Fast Three-Dimensional 1H MR Spectroscopic Imaging at7 Tesla Using “Spectroscopic Missing Pulse – SSFP”

Christian Schuster,1,2 Wolfgang Dreher,1,2 Jorg Stadler,3 Johannes Bernarding,4 andDieter Leibfritz1,2*

The use of spectroscopic Missing Pulse - SSFP (spMP-SSFP)for fast three-dimensional (3D) proton MR spectroscopic im-aging (MRSI) at 7 Tesla (T) is demonstrated. Sequence mod-ifications were required regarding the limits of the specificabsorption rate as well as hardware limitations with respectto maximum B1 field strength and B0 gradient slew rate, ascompared to previous studies performed at 3T. The combi-nation of two spatially selective radiofrequency (RF) pulses(with orthogonal slice orientation) and a dual-band chemicalshift selective RF pulse for simultaneous water and lipidsuppression proved to enable fast 3D MRSI measurements ofthe brain of healthy volunteers. Using a total measurementtime of approximately 8.5 minutes and a nominal and realvoxel size of 0.62 cm3 and 2.6 cm3, respectively, signals ofN-acetyl aspartate, total creatine, choline containing com-pounds, myo-inositol, and glutamate�glutamine could bedetected. Magn Reson Med 60:1243–1249, 2008. © 2008Wiley-Liss, Inc.

Key words: spectroscopic imaging; human brain; steady state;missing pulse

Fast three-dimensional (3D) proton MR spectroscopic im-aging (1H MRSI) (1,2) of the human brain using steady statefree precession (SSFP) based sequences like spectroscopicCE-FAST (3), spectroscopic FAST (4) and spectroscopicMissing Pulse - SSFP (5) (spMP-SSFP) provides rapidlyacquired maps of the spatial distribution of brain metabo-lites. These spectroscopic methods have been derivedfrom their corresponding MR imaging sequences (6,7)which exhibit a short minimum measurement time and ahigh signal-to-noise ratio per unit measurement time(SNRt). While the speed of these sequences is a conse-quence of the inherently short repetition times (TR) usedto obtain a strong steady-state signal, the spectral resolu-tion of the spectroscopic variants is limited due to theconstraints on the available acquisition time. Therefore,SSFP based methods for 1H MRSI have been introduced forhigh static magnetic field strengths of 3 Tesla (T) and 4.7T(8) for which an appropriate spectral resolution isachieved. Moving toward even higher B0 magnetic fields,

SSFP based spectroscopic sequences would specificallybenefit because the spectral resolution could be furtherincreased and/or the repetition time could be shortenedfor a given spectral resolution (in ppm) resulting in ashorter minimum measurement time and an additionalSNRt gain. This work explores the feasibility of using fastSSFP based 3D 1H MRSI of the human brain on a 7T wholebody system to make use of the general increase in theSNRt and the spectral resolution at higher B0 fieldstrengths (9–11).

In vivo 3D 1H MRSI of the human brain benefits fromefficient and reproducible water and lipid suppressionmethods by which artifact signals and/or an SNRt decreasemay be avoided. The SSFP based sequences as introducedin (3–5) use spectral-spatial RF pulses for a simultaneousslice and chemical shift selective excitation/refocusing.These spectral-spatial RF pulses are derived from a com-posite pulse proposed in (12). The constant time interval �between the individual six subpulses of the compositepulse determines the periodicity of the frequency responseexhibiting minima at frequencies of � N/� (N �0,1,2,3,. . .), thus suppressing signals at these frequencies.At 3T (MAGNETOM Allegra, Siemens Healthcare Sector,Erlangen, Germany), a 2 ms interpulse time was used toachieve suppression bands at 0 and 500 Hz for a robustwater and lipid suppression (5). However, the interpulsetime has to be shorter at 7T due to the increased dispersionof the chemical shift (in Hz). Thus the interpulse timeneeds to be as short as 857 �s corresponding to minima ofthe suppression bands at 0 and 1167 Hz. When consider-ing the pulse duration of 600 �s of one individual sub-pulse of the composite pulse, the inevitably short inter-pulse time causes gradient switching problems with com-mercial gradient systems. For this reason, the spectral-spatial RF pulses were replaced at 7T. However, using asimple modification of the spMP-SSFP (5) pulse sequence,both slice selection and chemical shift selection can bepreserved. Therefore, this SSFP based MRSI method hasbeen considered first to be implemented on a 7T wholebody system and applied to phantoms and healthy humanbrains.

MATERIALS AND METHODS

Pulse Sequence

The modified spMP-SSFP sequence is shown schemati-cally in Figure 1. The spectral-spatial composite pulsesused for the RF excitation and refocusing at 3T were re-placed by slice selective Hamming filtered sinc-pulses of2.8-ms duration. The second RF pulse excites an orthogo-nal slice with respect to the slice of the first RF pulse sothat only spins from the intersection of both slices are

1University of Bremen, FB 2 (Chemistry), Bremen, Germany.2Center for Advanced Imaging (CAI), Bremen, Germany.3Leibniz Institute for Neurobiology, Magdeburg, Germany.4Institute for Biometry and Medical Informatics, University Hospital of Mag-deburg, Magdeburg, Germany.Dr. Schuster’s present address is Siemens AG Healthcare Sector, Erlangen,Germany.*Correspondence to: Dieter Leibfritz, University of Bremen, FB 2 (Chemistry),Leobener Str. NW2/C, D-28359 Bremen, Germany.E-mail: dieter.leibfritz@uni-bremen.deReceived 14 March 2008; revised 8 July 2008; accepted 21 July 2008.DOI 10.1002/mrm.21787Published online 3 October 2008 in Wiley InterScience (www.interscience.wiley.com).

Magnetic Resonance in Medicine 60:1243–1249 (2008)

© 2008 Wiley-Liss, Inc. 1243

refocused and contribute to the spin echo signal, thusenabling 2D spatial preselection (13). A fixed 180° phaseshift between �1 and �2 is applied to achieve the maximumintensity for metabolite signals with uncoupled spins (5).The slice selective gradients are fully balanced to avoidany disturbance of the steady state.

The spatially selective outer volume suppression (OVS)RF pulses included within the first T interval in the se-quence for 3T (5) were substituted by one spectrally selec-tive dual-band RF pulse which serves for water as well asfor lipid signal suppression in the sequence for 7T. Thedual-band RF pulse was calculated by superimposing twosymmetrical Gaussian functions which decay to exp(-3). Aphase difference of 180° at the center of the pulse was usedto minimize the SAR. Using a flip angle of �s � 180°achieves an optimal suppression factor of 100%. An inclu-sion of additional OVS RF pulses in the first T interval wasnot possible because of SAR reasons.

The full spin echo signal comprises a coherent additionof the primary spin echo from �1 and �2 with an echo timeof TE � 2T as well as further higher order spin echoeswhose contributions depend on the relaxation times.Other coherences are destroyed by spoiler gradients afteror before the RF pulses. Phase encoding gradients in two orthree dimensions are reversed after each signal acquisitionto maintain the steady state.

Experimental Setup

For all measurements, an eight-channel head array coil(RAPID Biomedical, Rimpar, Germany) was used for bothRF transmission and signal reception on a 7T whole bodyMR scanner (Siemens Healthcare Sector, Erlangen, Ger-many).

Experiments were performed with 2D spatial resolutionon a spherical phantom with a diameter of 18 cm. Thephantom contained an aqueous solution of 12.5 mM N-acetyl-L-aspartic acid (NAA), 10 mM creatine (Cr), 3 mMcholine chloride (Cho), 7.5 mM myo-inositol (m-Ins),12.5 mM L-glutamic acid (Glu), and 5 mM lactic acid (Lac).

Parameters of the 2D 1H SI phantom measurement wereas follows: field of view (FOV) (x,y): 200 � 200 mm2,volume of interest (VOI) (x,y): 200 � 100 mm2, slice thick-ness: 25 mm, phase encoding steps (x,y): 16 � 16, TR �183 ms, TE � 122 ms (echo time of primary spin echo),�1 � 40°, �2 � �40°, acquisition bandwidth � 5000 Hz,512 complex spectral data points, total measurement time:0:59 min including 64 dummy scans. For the spectrallyselective dual-band RF pulse a duration of 10 ms and a flipangle of only �s � 120° was applied to use the same valueas for the in vivo measurements where the flip angle had tobe limited for SAR reasons.

In vivo 3D 1H SI measurements were recorded from thebrain of two healthy volunteers. All measurements havebeen conducted in compliance with the regulations of thelocal authorities. After three-plane scout imaging and theacquisition of T1-weighted images with an MP-RAGE (14)sequence, the FOV of the 3D spMP-SSFP MRSI measure-ment with a size of 190 � 190 � 100 mm3 was positionedto cover the whole brain. One spatially selective RF pulseexcited a slice of 50 mm thickness (foot to head direction),while the other selected an orthogonal slice of 120 mmthickness (anterior–posterior) to exclude extracranial lipidsignals originating from frontal and occipital head regions.The flip angle of the water and lipid suppression dual-band RF pulse had to be limited to �s � 120° because ofSAR reasons.

Further measurement parameters were as follows: nom-inal phase encoding steps (x,y,z): 22 � 22 � 12 (zero-filledto 32 � 32 � 16), TR � 183 ms, TE � 122 ms (echo time ofprimary spin echo), �1 � 50°, �2 � �50°, acquisitionbandwidth � 5000 Hz, 512 complex spectral data points,64 dummy scans, total measurement time: 8:35 min usingan acquisition weighted k-space sampling (15,16). Thiselliptical k-space sampling scheme was performed by ac-quiring the k-space center with 3 averages while the aver-ages of the residual k-space points were reduced comply-ing with a symmetrical Hamming function around thek-space center based on the nominal phase encoding steps.Therefore, the nominal voxel size of 0.62 cm3 was in-creased to a real voxel size of 2.60 cm3 by the acquisitionweighting with subsequent 3D spatial Hamming filtering.The flip angles �1 and �2 were chosen based on simula-tions for uncoupled spins and typical T1 and T2 values (5).

Signal combination of time domain data from individualcoil elements was done using the residual water signal forweighting and normalizing voxel by voxel (17). Furtherdata processing with the interactive data language IDL(Research Systems, Boulder, CO) included apodization ink� with a sine-bell function, zerofilling from 512 to 4096spectral data points and Fourier transformation. Spectrawere displayed in magnitude mode without decreasing thespectral resolution because full spin echoes are acquiredwith the spMP-SSFP sequence.

RESULTS

Figure 2 displays all spectra of the 2D 1H SI phantommeasurement acquired with the modified spMP-SSFP se-quence as well as a typical spectrum from one voxel whoseposition is indicated in the reference image. The solidlines within the image indicate the FOV, while the dashed

FIG. 1. Diagram of the modified spectroscopic Missing-Pulse SSFPsequence for 7T. A 2D spatial selection is realized by two sliceselective RF pulses exciting two orthogonal slices. The spectrallyselective dual-band RF pulse introduced in the first T interval servesfor water and lipid suppression. A full spin echo is acquired, whichcomprises the primary spin echo (TE � 2T) and several higher orderechoes dependent on the relaxation times. Unwanted coherencesare eliminated by spoiler gradients.

1244 Schuster et al.

lines show the slice of 100 mm thickness excited by the RFpulse �2.

The excellent spatial selection is proven by the noisyspectra outside the VOI because spins excited by only oneRF pulse (either �1 or �2) do not contribute to the spin echosignal. The typical ring-shaped B1 inhomogeneities occur-ring at high static magnetic fields are visible in the refer-ence image and cause a nonuniform distribution of metab-olite signals within the spectral map. Furthermore, somevoxels exhibit significantly lower signal intensities thantheir neighboring voxels which is attributed to the nonop-timal combination of data from the elements of the eight-channel array coil (for details see the Discussion section).

Resonances of both uncoupled and J-coupled spins aredetected and labeled in the enlarged magnitude spectrum.The peak of Lac at 1.3 ppm is partially suppressed becausethe dual-band RF pulse is applied to that spectral regionfor in vivo lipid suppression.

Results of an in vivo 3D 1H SI measurement are shown inFigures 3 and 4. The 2D spectral maps whose positions areindicated with a dotted square in the corresponding refer-ence images represent the distribution of metabolites fromtwo different slices through the brain, one through theventricles (Fig. 3) and one closely above the ventricles(Fig. 4). Again, the FOV (solid line) and the VOI in ante-rior–posterior direction (dashed lines) are shown in thereference image.

The enlarged magnitude spectra from the voxelsmarked in the images show metabolite signals of uncou-pled and J-coupled spins which are assigned to NAA,total creatine (tCr), choline containing compounds(tCho), m-Ins, and glutamateglutamine (Glx). Residuallipid signals originating from voxels in the scalp indi-cate that lipid suppression was not perfect. They areparticularly strong at the edges of the illustrated part ofthe spectral maps and contaminate some voxels inside

FIG. 2. Results of the 2D 1H SImeasurement on a sphericalphantom containing NAA, Cr,Cho, m-Ins, Glu, and Lac. The 2Dspectral map shows all magni-tude spectra from the FOV whichis indicated by white solid lines inthe reference image. The spatialselection by �2 is marked withdashed lines. The position of thevoxel whose magnitude spectrumis displayed in the lower right isindicated within the reference im-age.

Fast 3D 1H MRSI at 7T Using spMP-SSFP 1245

the brain, but do not affect the NAA peak at 2.0 ppm inthe majority of the voxels.

DISCUSSION

The modified spMP-SSFP sequence without spectral-spa-tial pulses enables fast 3D MRSI of the human brain at 7Tusing simple slice selective RF excitation/refocusingpulses and a spectrally selective dual-band RF pulse forboth water and lipid suppression. This spectrally selectivepulse performs well despite the typical B1 inhomogene-ities present at 7T. Even the small flip angle of �s � 120°,which had to be limited because of SAR reasons, is suffi-cient to suppress the lipid signals in such a way that

voxels inside the brain are contaminated only minimally.Disregarding the constraints given by the SAR limitation,for a perfect lipid suppression, the flip angle of choice is�s � 180°, because then, the dephasing of lipid spinswould be 100% (18). However, note that the chosen flipangle values are based on a spatially global RF poweradjustment and that the large B1 inhomogeneities willcause severe deviations from these values over the sample.Thus, the use of an adiabatic RF pulse for a B1-insensitivelipid suppression might be even more appropriate for 7Talbeit not possible with the given measurement parametersdue to the larger SAR requirements of an adiabatic RFpulse. Because the SAR was already 99% of the maximumallowed value with the parameters given above (�1 � 50°,

FIG. 3. Results of the 3D 1H SImeasurement on the brain of ahealthy volunteer. A spectral mapfrom one partition through theventricles whose position ismarked with a dotted square inthe reference image shows mag-nitude spectra within a range of4.3–0.0 ppm. The FOV is indi-cated by solid lines while the sliceselected by �2 in anterior–poste-rior direction is marked by thedashed lines. The enlarged mag-nitude spectrum originates fromthe voxel outlined in the referenceimage.

1246 Schuster et al.

�2 � �50°, �s � 120°, TR � 183 ms), an optimization of RFpulses with regard to a minimal power deposition is cur-rently under investigation. Using such pulses, an alterna-tive approach for lipid suppression could be achieved byadditional OVS slices within the first T interval as hasbeen done in the sequence for 3T (5).

Large B1 inhomogeneities do not only affect the efficacyof water and lipid suppression but cause spatial differ-ences in the created transverse magnetization and the ob-served NMR signal, too. Considering the principle of rec-iprocity, the sensitivity with which a signal is observed isproportional to the local B1 field produced by a unit cur-rent in the RF coil. Additionally, for a missing pulse SSFPsequence one may expect very large differences in the

observed transverse magnetization because the echo-likesignal is created by two or more RF pulses and thus theeffects of B1 inhomogeneities and the corresponding dif-ferences in flip angles may multiply. However, as shownin Figure 2 of Schuster et al. (5) the dependence of theobserved transverse magnetization on the flip angle �1 ���2 is rather weak, particularly for flip angles larger than30–40°. Therefore, the spMP-SSFP sequence is not moreor even less susceptible to spatial B1 inhomogeneities thanother SI sequences.

The slice selective RF pulses for excitation/refocusinghave been optimized with respect to minimum chemicalshift artifacts in the in vivo measurements. With a mini-mum duration of 2.8 ms, the sinc-shaped pulses had a

FIG. 4. Further results of the 3D1H SI measurement on the brainof a healthy volunteer. A spectralmap (position marked with a dot-ted square in the reference image)from another partition slightlyabove the ventricles shows mag-nitude spectra within a range of4.3–0.0 ppm. The FOV is indi-cated by solid lines while the se-lected slice in anterior–posteriordirection is marked by dashedlines. The enlarged magnitudespectrum originates from thevoxel outlined in the reference im-age.

Fast 3D 1H MRSI at 7T Using spMP-SSFP 1247

bandwidth of 3.125 kHz resulting in a relative chemicalshift displacement of approximately 9.5% for 1 ppm. Be-cause the excitation frequency was set to 3 ppm, the max-imum displacement for the metabolites of interest between4 and 2 ppm was � 11.4 mm in-plane (anterior–posterior)and � 4.75 mm in the foot-head direction.

As stated before in the results section, individual voxelsin the phantom measurement exhibit significantly lowersignal intensities than their neighboring voxels. This is aconsequence of the automatic combination algorithm thatsums up the time-domain signals from each coil element ofthe eight-channel Tx/Rx coil. The combination algorithmuses the first point in the time-domain data of each indi-vidual voxel to determine zero-order phasing and coilweighting reflecting the local coil reception sensitivity.This is certainly an appropriate procedure for FID-like SIdata provided that the NMR signal is strong enough for thealgorithm to work properly. Because full spin echoes areacquired in the spMP-SSFP sequence, the first point in thetime-domain data might not be large enough for the deter-mination of a correct weighting factor. Thus signals wereinherently suppressed in voxels with low signal intensi-ties and almost complete water suppression. For an appro-priate combination of the spMP-SSFP data the echo max-imum (i.e., the central time-domain signal) should beused, which has not been possible up to now. However,the suppression effect of single voxels due to impropersignal combination was more significant in phantom mea-surements in which a better water suppression wasachieved compared with the in vivo measurements. In thein vivo spectra the residual water signal was sufficientlylarge for the combination algorithm to work properly, eventhough the first point of the time-domain data was used.An alternative way of determining the weights of the coilelements would be to use additional prescans using MRI orSI (without water suppression) sequences. The latterwould increase the total measurement time, but improvethe accuracy of the coil weights determined and avoid anydependence on the quality of water suppression.

Because the same sequence timing as for 3T (5) wasused, an increase of the spectral resolution in comparisonto 3T could be achieved. However, an accurate quantita-tive comparison of spectral resolutions is hard to deter-mine due to the different measurement and processingconditions. A rough estimation shows an increase in spec-tral resolution by a factor of approximately 1.6 � 0.2 forthe NAA signal at 7T compared with 3T taking into ac-count the inferior shim and the stronger spectral apodiza-tion function needed at 7T to avoid ringing artifacts be-cause of the stronger nonrelaxed water and lipid signals. Itis noteworthy that the SNR could be increased by using ashorter interpulse delay T than in the current implemen-tation, however, at the cost of a reduced spectral resolutionand an increased SAR.

At 7T additional metabolite signals, that is, Glx, aredetected which have not been visible with the spMP-SSFPsequence at 3T. The reason might be the higher SNR, thebetter spectral resolution and/or the different field depen-dent behavior of J-coupled spins which also changes, e.g.,the appearance of the m-Ins peaks at 3.5/3.6 ppm. As

discussed in Schuster et al. (5), the quantification of spMP-SSFP spectra is mainly possible by using peak ratios be-cause the signal intensities depend on the T1 and T2 relax-ation times and on the coupling pattern of J-coupled spinsmaking absolute quantification difficult or even impossi-ble.

Until now, the sequence parameters have been adjustedempirically to maximize the SNRt of specific metabolitesof interest. While the signal intensity of uncoupled spinscan be calculated easily (19), simulation programs for theSSFP signal of J-coupled spins have not been available.However, following the ideas of Anand et al. (20) mayallow optimizing the sequence parameters for the detec-tion of both uncoupled and J-coupled spins in the future.

CONCLUSION

A modification of the spMP-SSFP sequence is presented,which allows for fast 3D 1H spectroscopic imaging of thehuman brain at 7T. By using slice selective excitation/refocusing RF pulses, the 2D spatial selection of the se-quence originally proposed for 3T can be preserved whilean additional spectrally selective dual-band RF pulse sup-presses both water and lipid signals. Future work regard-ing optimized RF pulses for both spectral and spatial se-lection will be done to further increase the SNRt and thespectral quality, despite the constraints given by the SARlimit and problems arising from B1 inhomogeneities.

ACKNOWLEDGMENT

The concepts and information presented in this study arebased on research and are not commercially available.

REFERENCES

1. Brown TR, Kincaid BM, Ugurbil K. NMR chemical shift imaging inthree dimensions. Proc Natl Acad Sci U S A 1982;79:3523–3526.

2. Maudsley AA, Hilal SK, Perman WH, Simon HE. Spatially resolvedhigh resolution spectroscopy by “four-dimensional” NMR. J Magn Re-son 1983;51:147–152.

3. Geppert C, Dreher W, Schuster C, Leibfritz D. Very fast 1H spectroscopicimaging of the human brain using spectral-spatial RF pulses and spCE-FAST. In: Proceedings of the 22nd Annual Meeting of ESMRMB, Basel,Switzerland, 2005. p. 224.

4. Schuster C, Dreher W, Leibfritz D. Very fast 2D proton spectroscopicimaging of the in vivo human brain at 3 Tesla with high spatial reso-lution using the SSFP based sequence spectroscopic FAST. In: Proceed-ings of the 14th Annual Meeting of ISMRM and 24th Annual Meeting ofISMRM Berlin, Germany. 2007. p. 774.

5. Schuster C, Dreher W, Geppert C, Leibfritz D. Fast 3D 1H spectroscopicimaging at 3 Tesla using spectroscopic Missing-Pulse SSFP with 3Dspatial preselection. Magn Reson Med 2007;57:82–89.

6. Gyngell ML. The application of steady-state free precession in rapid2DFT NMR imaging: FAST and CE-FAST sequences. Magn Reson Im-aging 1988;6:415–419.

7. Patz S, Wong STS, Roos MS. Missing pulse steady-state free precession.Magn Reson Med 1989;10:194–209.

8. Dreher W, Geppert C, Althaus M, Leibfritz D. Fast proton spectroscopicimaging using steady-state free precession methods. Magn Reson Med2003;50:453–460.

9. Gruetter R, Weisdorf SA, Rajanayagan V, Terpstra M, Merkle H, TruwitCL, Garwood M, Nyberg SL, Ugurbil K. Resolution improvement in invivo 1H NMR spectra with increased magnetic field strength. J MagnReson 1998;135:260–264.

1248 Schuster et al.

10. Tkac I, Andersen P, Adriany G, Merkle H, Ugurbil K, Gruetter R. In vivo1H NMR spectroscopy of the human brain at 7 T. Magn Reson Med2001;46:451–456.

11. Otazo R, Mueller B, Ugurbil K, Wald L, Posse S. Signal-to-noise ratioand spectral linewidth improvements between 1.5 and 7 Tesla inproton echo-planar spectroscopic imaging. Magn Reson Med 2006;56:1200–1210.

12. Starcuk Z, Sklenar V. New hard pulse sequences for solvent signal sup-pression in Fourier transform NMR. II. J Magn Reson 1986;66:391–397.

13. Hwang KP, Flask C, Lewin JS, Duerk JL. Selective missing pulse steadystate free precession (MP-SSFP): inner volume and chemical shift se-lective imaging in a steady state sequence. J Magn Reson Imaging2004;19:124–132.

14. Mugler JP, Brookeman JR. Three-dimensional magnetization-preparedrapid gradient-echo imaging (3D MP RAGE). Magn Reson Med 1990;15:152–157.

15. Mareci TH, Brooker HR. High-resolution magnetic resonance spectrafrom a sensitive region defined with pulsed field gradients. J MagnReson 1984;57:157–163.

16. Hugg JW, Maudsley AA, Weiner MW, Matson GB. Comparison ofk-space sampling schemes for multidimensional MR spectroscopic im-aging. Magn Reson Med 1996;36:469–473.

17. Brown MA. Time-domain combination of MR spectroscopy data ac-quired using phased-array coils. Magn Reson Med 2004;52:1207–1213.

18. Mescher M, Tannus A, O’Neil Johnson M, Garwood M. Solvent sup-pression using selective echo dephasing. J Magn Reson A 1996;123:226–229.

19. Scheffler K. A pictorial description of steady-states in rapid magneticresonance imaging. Concepts Magn Reson 1999;11:291–304.

20. Anand CK, Bain AD, Nie Z. Simulation of steady-state NMR of coupledsystems using Liouville space and computer algebra methods. J MagnReson 2007;189:200–208.

Fast 3D 1H MRSI at 7T Using spMP-SSFP 1249