Available online at www.sciencedirect.com

Magnetic Resonance Imaging 31 (2013) 86–95

On the utility of spectroscopic imaging as a tool for generatinggeometrically accurate MR images and parameter maps in the presence of

field inhomogeneities and chemical shift effectsChris J.G. Bakkera, b,⁎, Hendrik de Leeuwb, Gerrit H. van de Maatb, Jetse S. van Gorpb,

Job G. Bouwmanb, Peter R. SeevinckbaDepartment of Radiology, University Medical Center Utrecht, Heidelberglaan 100, 3584CX Utrecht, the NetherlandsbImage Sciences Institute, University Medical Center Utrecht, Heidelberglaan 100, 3584CX Utrecht, the Netherlands

Received 15 February 2012; revised 18 May 2012; accepted 21 June 2012

Abstract

Lack of spatial accuracy is a recognized problem in magnetic resonance imaging (MRI) which severely detracts from its value as astand-alone modality for applications that put high demands on geometric fidelity, such as radiotherapy treatment planning andstereotactic neurosurgery. In this paper, we illustrate the potential and discuss the limitations of spectroscopic imaging as a tool forgenerating purely phase-encoded MR images and parameter maps that preserve the geometry of an object and allow localization ofobject features in world coordinates.

Experiments were done on a clinical system with standard facilities for imaging and spectroscopy. Images were acquired with a regularspin echo sequence and a corresponding spectroscopic imaging sequence. In the latter, successive samples of the acquired echo were used forthe reconstruction of a series of evenly spaced images in the time and frequency domain. Experiments were done with a spatial linearityphantom and a series of test objects representing a wide range of susceptibility- and chemical-shift-induced off-resonance conditions.

In contrast to regular spin echo imaging, spectroscopic imaging was shown to be immune to off-resonance effects, such as those causedby field inhomogeneity, susceptibility, chemical shift, f0 offset and field drift, and to yield geometrically accurate images and parameter mapsthat allowed object structures to be localized in world coordinates.

From these illustrative examples and a discussion of the limitations of purely phase-encoded imaging techniques, it is concluded thatspectroscopic imaging offers a fundamental solution to the geometric deficiencies of MRI which may evolve toward a practical solutionwhen full advantage will be taken of current developments with regard to scan time reduction. This perspective is backed up by ademonstration of the significant scan time reduction that may be achieved by the use of compressed sensing for a simple phantom.© 2013 Elsevier Inc. All rights reserved.

Keywords: Spectroscopic imaging; Field inhomogeneity; Geometric distortion; Magnetic susceptibility; Chemical shift; Interventional MRI; Quality assurance

1. Introduction

Lack of spatial accuracy is a recognized problem inmagnetic resonance imaging (MRI). In standard MR scans,the problem displays itself in two ways: first, the images donot faithfully represent the geometry of the object underinvestigation, especially in the presence of implants, internaland external devices, surgical tools, localization frames, etc.;second, the images do not exactly reveal the position of the

⁎ Corresponding author.E-mail address: c.j.g.bakker@umcutrecht.nl (C.J.G. Bakker).

0730-725X/$ – see front matter © 2013 Elsevier Inc. All rights reserved.http://dx.doi.org/10.1016/j.mri.2012.06.023

object with respect to the world outside (‘the laboratory’),i.e., uncertainty arises when trying to convert imagecoordinates to laboratory coordinates. This uncertaintydiminishes the accuracy by which a certain volume ofinterest within the image (a ‘target’) can be localized inphysical space, for example, with respect to the tabletop ofthe scanner. Both geometric deficiencies have a detrimentaleffect on the value of MRI as a stand-alone modality forapplications that put high demands on geometric fidelity,such as radiotherapy treatment planning, stereotactic neuro-surgery, stereotactic laser and radiosurgery, and image-guided interventions, e.g., needle biopsies.

87C.J.G. Bakker et al. / Magnetic Resonance Imaging 31 (2013) 86–95

According to the vast literature on the subject (see,e.g., Ref. [1] and references therein), the lack ofgeometric accuracy in MRI is mainly caused by thenonlinearity of the imaging gradients and by off-resonance effects. The first factor, nonlinearity of thegradients, is largely a matter of system and sequencedesign and virtually independent of the object. It isoptionally corrected for on most clinical MR systems and,therefore, not considered an issue here [2]. With regard tothe second factor, off-resonance effects, several sourcescan be identified, including system-related field inhomo-geneities and instabilities, susceptibility-induced fieldinhomogeneities, chemical-shift-related frequency offsetsand an overall frequency offset associated with the choiceof a particular reference frequency, f0. Usually, this f0 isinferred from the frequency distribution in the part of theobject to be imaged. The resultant f0 is used to define theorigin of image space along the slice selection andreadout axes but does not refer to a specific or exactlyknown location in physical space. As a consequence, theorigin of image space will exhibit a usually small butunknown shift with respect to the origin of the laboratoryframe in the slice selection and readout directions.Additional shifts may occur when the static magneticfield exhibits temporal variations, for example, due toresistive heating of magnet structures [3]. Such shifts willbe particularly harmful when scanning in the dynamicmode, as is typically done during diagnostic andtherapeutic interventions.

An obvious way to deal with the geometric deficienciesof MRI is the application of very strong slice selection andreadout gradients. Since distortions are generally inverselyproportional to the strengths of the gradients [4], thisapproach is effective in many cases. Yet, strong gradientsdo not provide the final solution for several reasons. First,strong slice selection and readout gradients go togetherwith large excitation and readout bandwidths and, hence,large radio frequency (RF) power depositions and imageswith a low signal-to-noise ratio (SNR). Second, rapidlyswitching strong gradients is noisy and may causeperipheral nerve stimulation [5]. Besides, it induces short-term and long-term eddy currents, which act as sources offield inhomogeneities themselves and are often hard tocompensate for. Third, gradient capabilities on clinicalsystems are usually not sufficient to overcome off-resonance effects at higher field strengths or invoked bylarge inhomogeneities such as those associated withmetallic devices, implants, needles, etc. At 7 T, forinstance, a readout gradient of 30 mT/m would alreadybe required to keep distortions associated with air cavities(Δχ≈10 ppm) within 1 mm.

A more conclusive and fundamental solution to theindicated geometric problems would be to refrain fromfrequency encoding altogether and rely on phase encodinginstead. In effect, this is exactly what is done in spectro-scopic imaging, albeit with a different purpose in mind [6,7].

In standard proton spectroscopic imaging (SI), for instance,the signals from water and fat are suppressed in order toenable visualization and analysis of low-concentration tissuemetabolites like N-Acetyl Aspartate (NAA), creatine andlactate. The signal from these metabolites is several orders ofmagnitude smaller than that from water, so to go from tissuechemistry to structural imaging is a matter of leaving out thesuppression techniques and an adaptation of the processingand display facilities. With suppression switched off, theacquired, purely phase-encoded images will be completelydominated by the properties of the bulk water and fat protonsand by line broadening and T2

* effects from microscopic andmacroscopic field inhomogeneities. Since only one phaseencoding step is applied after each excitation and prior tocollection of a free induction decay (FID) or spin echo (SE)signal, each acquired data point in k-space will exhibit thesame influence of dipolar interactions, chemical shifts, fieldinhomogeneities, diffusion and other mechanisms. Theresultant images, one T2

*-weighted complex image for eachsample point, will, therefore, not be distorted by off-resonance effects, irrespective of the strength of the appliedgradients. Evidently, this will also be the case for thefrequency-resolved images and the parameter maps, e.g.,T2* maps and ΔB0 maps, that may be derived from the

acquired data.To date, this nonstandard use of spectroscopic imaging

— i.e., for obtaining geometrically accurate images ofwater and fat protons rather than biochemical information— has not aroused much interest in the medical communitysince the lengthy examination times associated with purelyphase-encoded imaging techniques have always beenconsidered prohibitive. Recent developments with regardto scan time reduction such as parallel imaging andcompressed sensing, however, make us believe that thisbarrier may not be as impregnable as might have seemed inthe past. We therefore thought it a timely exercise to startand promote interest in this nonstandard use of spectro-scopic imaging by demonstrating its utility in generatingtime- and frequency-domain-resolved images and parame-ter maps that (a) correctly depict the geometry of an objectand (b) allow localization of object features in laboratory/world coordinates. By way of experiments with a spatiallinearity phantom and various configurations of test tubes,spectroscopic imaging will be shown to yield images andparameter maps that are geometrically immune to off-resonance effects, such as those caused by field inhomo-geneity, susceptibility, chemical shift, f0 offset and fielddrift. Once we have demonstrated the potential ofspectroscopic imaging in providing a fundamental solutionto the geometric deficiencies of MRI, we will discuss howthis solution may evolve towards a practical solution whenfull advantage will be taken of current developments withregard to scan time reduction. To back up this perspective,we will demonstrate the significant scan time reduction thatmay be achieved by the use of compressed sensing for asimple phantom.

�

�

88 C.J.G. Bakker et al. / Magnetic Resonance Imaging 31 (2013) 86–95

2. Theory

In a one-dimensional (1D) spin echo experiment withfrequency encoding along the x-axis (Fig. 1, top), the p-thsample of the acquired signal can be written as:

s pΔkxð Þ∝∫ρ xð Þexp −TE + pτT2 xð Þ

� �exp −i γ ΔB xð Þ pτ½ �

exp −i 2π pΔkx x +ΔB xð ÞGr

� �� �dx; ð1Þ

with −Nx/2≤pbNx/2, Nx the number of samples, ρ(x) theeffective 1D spin density, incorporating T1 and many otherfactors, Gr the amplitude of the applied readout gradient,Δkx= γGrτ the step-size in k-space and ΔB(x)=B(x)−B0 thespatially dependent deviations from B0 such as those causedby field inhomogeneities, susceptibility deviations, chemicalshifts and f0 shifts. The first and second exponentials inEq. (1) describe the signal variations, as a function of p, dueto T2⁎ decay during data sampling. Although often ignored,such variations cause blurring and impose a limit to theresolution that can be achieved in an experiment [4]. Thethird exponential in Eq. (1) describes the distortion that iseffected by field inhomogeneities. Utilizing the Fourier shifttheorem, i.e., s(k)exp(−i2πka)=FT[ρ(x-a)], it readily followsfrom Eq. (1) that the reconstructed spin density distributionwill be geometrically distorted according to:

ρ x0ð Þ = ρ x−ΔB xð ÞGr

� �: ð2Þ

With regard to the static field B0 and, hence, thedeviations ΔB, it should be realized that, in the context of

Fig. 1. Pulse sequence diagram of, respectively, a 1D SE experiment with a frequencwith a stepped phase encoding gradient Gp (bottom). In the spectroscopic imagingimage for each sample point t=TE+mδt of the spin echo, with Ns/2≤mbNs/2 and

a particular imaging experiment, B0 is not a measuredquantity but is defined in terms of the resonance frequencyvia B0≡f0/γ. The resonance frequency f0 is usually derivedfrom the frequency distribution within the part of the objectto be imaged. This f0 serves as a reference in image space butis not related to a specific position in object space. As aconsequence, an f0-related, small-but-unknown shift mayoccur between the origins of the coordinate systems in imagespace and physical space (the laboratory) in the frequencyencoding direction(s).

In a 1D spectroscopic imaging experiment with spin echodetection, the p-th sample of the acquired k-space data foreach point t=TE+mδt of the spin echo (Fig. 1, bottom) canbe written as:

s pΔkxð Þ∝∫ρ xð Þexp −TE + mδtT2 xð Þ

� �exp −i γ ΔB xð Þmδt½ �

exp −i 2π pΔkxx½ �dx: ð3Þwith −Nx/2≤pbNx/2, Nx the number of samples, ΔGp thestep-size of the phase encoding gradient, tp the duration ofthe phase encoding gradient and Δkx= γ ΔGptp the step-size in k-space. Eq. (3) shows that, by applying only onephase encoding step per excitation, the s(pΔkx) for eachpoint t=TE+mδt of the spin echo will be equally affectedby off-resonance effects. An undistorted image can thus bereconstructed for each individual sample point t=TE+mδtof the spin echo, and blurring will be negligible when thedwell time δt is chosen sufficiently short. For eachindividual image, the influence of dipolar interactions,chemical shifts, field inhomogeneities, diffusion and othermechanisms is constant, but the influence may vary among

y encoding gradient Gr (top) and a 1D SE spectroscopic imaging experimenexperiment, a series of Ns images is obtained by reconstructing a separate

Ns the number of samples.

t

89C.J.G. Bakker et al. / Magnetic Resonance Imaging 31 (2013) 86–95

the images. Off-resonance effects may thus be exploited asa contrast mechanism, without introducing geometricdistortion and blurring.

3. Materials and methods

3.1. Phantoms

To demonstrate the utility and illustrate the capabilities ofthe purely phase-encoded spectroscopic imaging sequences,we used a spatial linearity phantom and various configura-tions of cylinders and test tubes. The spatial linearityphantom consisted of a regular grid of 23×23 square holes of9 by 9 mm with a depth of 20 mm and a spacing of 1 mmdrilled in an acrylic plate with a thickness of 2.5 cm (PromisElectro Optics BV, Eindhoven, the Netherlands). In thereported experiments, only the central 5×5 elements of thegrid were filled with an MR visible fluid, viz., manganese-doped water with a T1 of about 750 ms and a T2 of about 80ms at 1.5 T. The second phantom setup featured variousconfigurations of cylinders and NMR test tubes with theirlongitudinal axes perpendicular to the tabletop and B0. Thefirst configuration (Fig. 3, columns 1 and 2) consisted of asingle cylinder with a diameter of 2.25 cm and a height of 10cm and was filled with manganese-doped water. In thesecond configuration (Fig. 3, column 3), a 1-cm-diametertest tube with sunflower oil was placed inside and parallel tothe outer cylinder, so as to create a chemical shift difference.In the third configuration (Fig. 3, column 4), oil was replacedby air to create a susceptibility difference of about 9 ppm. Inthe fourth configuration (Fig. 5, column 1), fourteen 5-mm-diameter test tubes with air were placed inside the cylinder tocreate multiple inhomogeneities at the scale of a voxel. In thefifth configuration (Fig. 5, column 2), a 1-cm-diameter testtube with doped water was placed outside the cylinder tocreate an external source of field disturbances. In the sixthconfiguration (Fig. 6), a 1.8-mm titanium needle was placedin a 2.25-cm-diameter test tube with doped water. The needlewas oriented perpendicular to B0 so as to create local fielddisturbances between −90 and +90 ppm.

3.2. Scan techniques

All imaging was done on a 1.5-T clinical whole-bodysystem (Achieva 1.5 T, Philips Healthcare, the Netherlands)using an 11-cm-diameter, two-element flexible phased-arraycoil for signal reception. Prior to the experiments, shimsettings were determined for the homogeneous cylinderphantom (Fig. 3, column 1) and maintained during furtherexperimentation. Single-slice coronal images were obtainedwith a standard spin echo sequence (Fig. 1, top) and acorresponding non-water-suppressed spectroscopic imagingsequence (Fig. 1, bottom). In the spectroscopic imagingexperiment, the spin echo signal was sampled at 128 equallyspaced intervals of 125 μs, corresponding to a bandwidth of8000 Hz. Note that a dwell time of 125 μs was considered to

be adequate for the objects in our study but that the dwelltime can readily be shortened, without technical impedimentbut at the cost of increased image noise, when T2⁎ lossesduring the dwell time would become a problem.

Examination parameters included an excitation band-width of 4 kHz, an echo time (TE) of 30 ms, a repetition timeof 150 ms, a flip angle of 90°, a scan matrix of 64×64, a fieldof view of 32×32 mm2 or 64×64 mm2, a reconstructionmatrix of 64×64 or 128×128, a slice thickness of 10 mm, twosignal averages in SE and one signal average in spectro-scopic imaging, resulting in a scan duration of 20.4 s for thestandard SE sequence and a scan duration of 10 min 14 s forthe SI sequence. In the standard SE sequence, readout wasalong the left–right axis with a readout bandwidth of 50 Hzper pixel. The same gradient power was used for phase andfrequency encoding in SE and phase encoding in SI (Gr=Gp

in Fig. 1). In the needle experiments, SE images were alsoacquired with an increased bandwidth of 500 Hz per pixeland 64 signal averages, yielding a scan time of 10 min 14 s.Prior to further analysis, images were corrected for non-linearities of the imaging gradients by the scanner software.The phantom setup, slice thickness and excitation bandwidthwere such that the effects of an f0 shift on the selected slicecould be ignored.

3.3. Processing and display

Following on the spectroscopic imaging sequence, acomplex image was reconstructed for each sample point t=TE+mδt of the spin echo, with −64≤mb64 and δt=0.125ms, yielding a time-resolved series of 128 complex images,which were converted to magnitude and phase maps forfurther analysis. Field maps were calculated by averagingacross 10 consecutive phase images acquired between t=TE (m=0) and t=TE+1.25 ms (m=10). A frequency-resolved series of images was obtained by a 1D Fouriertransform of the time-resolved images along the time axis.This yielded 128 complex images with f=mδf, −64≤mb64and δf=1/(128δt)=62.5 Hz, covering the total spectroscopicbandwidth of 1/δt=8000 Hz. Magnitude averaging wasapplied when a lower temporal or spectral resolution wasdesired than the resolution at which the data were acquired,e.g., for display purposes.

3.4. Positioning of the object

To position the object, first the tabletop was moved out ofthe magnet and locked into the parking position. Next, thelight visor system was used to project laser cross-hairs ontothe object indicating, respectively, a sagittal reference planethrough the isocenter of the magnet and a transversereference plane at a distance of 80.5 cm from the isocenter.The coronal reference plane through the isocenter was notindicated by the laser system but was verified to be located ata height of 12.0 cm above the tabletop. The object waspositioned with respect to these reference planes and securedto the tabletop using the appropriate immobilization tools.

90 C.J.G. Bakker et al. / Magnetic Resonance Imaging 31 (2013) 86–95

Finally, the travel-to-isocenter function was activated tomove the transverse reference plane to the isocenter of themagnet and to complete positioning of the object withrespect to the laboratory frame. The tabletop accuracy, asspecified by the manufacturer, was 0.5 mm.

3.5. Determination of the resonance frequency

In preparation of each imaging experiment, the resonancefrequency f0 was determined with the object in place andusing the f0 of the previous scan as the reference for selectionand readout. This was done by acquiring a volume selectivespectrum of a 100×100×50-mm box centered at the slice tobe imaged using a point-resolved spectroscopy (PRESS)with a TE of 100 ms. The f0 sequence was always precededby a short TI inversion recovery (STIR) sequence to null thefat signal. The center of gravity of the resultant powerspectrum was taken as the frequency offset for updating f0.In case of shimming, the f0 determination was repeated aftershimming. In some experiments, the f0 sequence wasskipped, and f0 was left unaltered or specified by the user.

4. Results

To demonstrate the different geometric responses of SEand SI to field inhomogeneity, in this case invoked by theshape, orientation and composition of the object, the spatiallinearity phantom was subjected to an SE and an SIacquisition with equal gradient power for phase andfrequency encoding. Prior to being imaged, the central 5×5elements of the grid were brought to the isocenter andaligned with the principal axes of the gradient system usingthe laser positioning device. Slice selection effects wereavoided by exciting a nominal slice of 10 cm, well beyondthe 2.5-cm thickness of the phantom. The results of the SEand SI acquisitions are displayed in Fig. 2 and give rise to thefollowing observations. First, standard SE images show

Fig. 2. Coronal images of a grid phantom with its center positioned in the origin of th5×5 elements was filled with doped water. The left panel displays the regular spin ecthe right panel displays the corresponding SI image (m=0, t=TE). White lines rep

correct alignment in the phase encoding direction and exhibitdistortions in the read direction (Fig. 2, left). Second, thealignment is perfect in both phase encoding directions in theSI images (Fig. 2, right). Third, the standard SE imageexhibits a small f0 shift along the read axis, which isassociated with the determination of f0 and which is absent inthe SI image.

To demonstrate the utility of SI in eliminating geometricdistortion due to off-resonance effects, various arrangementsof test tubes were subjected to a standard SE acquisition anda corresponding spectroscopic imaging acquisition, the latteryielding 128 images with different t=TE+mδt. In allexperiments, the center of the larger cylinder was alwayspositioned in the origin of laboratory frame. The results aredisplayed in Fig. 3 and lead to the following observations.First, both SE and SI yield geometrically reliable images inthe absence of field inhomogeneities (Fig. 3, column 1).Second, a shift of the resonance frequency causes atranslation of the SE image along the readout axis, butleaves the SI images unaffected (Fig. 3, column 2). In thiscase, f0 was manually increased by 800 Hz. Third, chemicalshift leads to misregistration along the readout axis in SE butis absent in SI images (Fig. 3, column 3). Fourth,susceptibility deviations give rise to geometric distortionsalong the readout axis in SE and are absent in SI images(Fig. 3, column 4). Overall, we observed correct alignment ofSI images in all cases; correct alignment of SE images in thephase encoding direction; and susceptibility distortions,chemical shift and f0 shift along the read direction in SE.

Next, the coaxial cylinder phantom (Fig. 3, column 4) wasused to demonstrate the utility of non-water-suppressedproton spectroscopic imaging in providing detailed infor-mation with regard to the distribution of microscopic andmacroscopic field inhomogeneities within an object. For thispurpose, the acquired time-resolved images of the phantom,some of which are displayed in the top row of Fig. 4, were1D Fourier transformed to a set of frequency-resolved

e laboratory frame and aligned with the magnet gradient system. An array ofho image with readout along the left–right axis (bandwidth 50 Hz per pixel)resent the left–right and feet–head axes of the laboratory frame.

;

Fig. 3. Coronal spin echo images (top row) and corresponding averaged spectroscopic images (bottom row) of a 2.25-cm-diameter cylinder with doped wate(columns 1 and 2) and the same cylinder with, respectively, a 10-mm-diameter test tube with oil (column 3) and a 10-mm test tube with air (column 4). In the spinecho images (top row), readout was always along the left–right axis. Black crosses in the left four images represent the origin of the laboratory frame.

91C.J.G. Bakker et al. / Magnetic Resonance Imaging 31 (2013) 86–95

images, some of which are displayed in the bottom row ofFig. 4. The time- and frequency-resolved images, respec-tively, reflect the evolution of time signal and the frequencycontent of each voxel, as is illustrated for a few voxels by thediagrams in Fig. 4D and H. The sinc modulation of theechoes in Fig. 4D is indicative of the presence ofmacroscopic field gradients and points to the invalidity ofa uniexponential model in analyzing such data.

Time- and frequency-resolved images constitute an idealstarting point for detailed studies of field inhomogeneityeffects and off-resonance phenomena and for the derivationof various kinds of undistorted parameter maps andcomposite images. This was demonstrated by experiments

Fig. 4. Coronal time-resolved (top row) and frequency-resolved (bottom row) spectroscopic images of a 2.25-cm-diameter cylinder with doped water and a 10-mm coaxial test tube with air. Top row: t=TE+mδt with m=−64 (A), m=0 (B) and m=+63 (C). Bottom row: f=mδf with m=−1 (E), m=0 (F) and m=+1 (G)Herein δt=0.125 ms, Ν=128 and δf =1/(Nδt)=62.5 Hz. Graphs (D) and (H), respectively, represent the time- and frequency-resolved signals for voxels at thelocations indicated by the symbols in graph (B).

r

with a cylinder with multiple susceptibility deviations(Fig. 5, left column) and a cylinder in the presence of anexternal field disturbance (Fig. 5, right column). In bothcases, spectroscopic imaging yielded 128 SI images withdifferent t=TE+mδt, resulting in a set of undistortedmagnitude images with systematically varying T2

* weightingand a set of undistorted phase maps describing the temporalevolution of the phase on either side of the spin echo top. Asillustrated by Fig. 5, such data sets can, for instance, be usedfor generating undistorted field maps.

Finally, the needle phantom was used for a comparativeevaluation of conventional SE and spectroscopic imaging inthe presence of larger field disturbances than commonly

.

Fig. 5. Coronal SE images (top row), SI images (middle row) and SI-derived field maps (bottom row) of a 2.25-cm-diameter cylinder withdoped water and fourteen 5-mm-diameter test tubes with air (left column)and an external 10-mm-diameter test tube with doped water (righcolumn). In the spin echo images (top row), readout was along the left–right axis. The field of view was doubled for examining the configurationwith the external tube (right column). The field maps (bottom row) werederived from 10 consecutive SI phase images that were acquired betweenδt=0.0 ms (m=0) and δt=1.25 ms (m=10).

Fig. 6. Coronal spin echo images (A, B, C) and corresponding averaged specdiameter titanium needle. (A) SE with BW=50 Hz/pixel and two signal avertime 20.4 s; (C) SE with BW=500 Hz/pixel and 64 signal averages, scan timvertical axis; readout along the horizontal axis in SE.

92 C.J.G. Bakker et al. / Magnetic Resonance Imaging 31 (2013) 86–95

t

encountered in the human body. Fig. 6 shows the severelydistorted conventional spin echo image that results when asmall readout bandwith is used (Fig. 6A). The distortion isstrongly reduced when using a much larger bandwidth, albeitat the expense of SNR, which is inversely proportional to thebandwidth (Fig. 6B). When the number of signal averages isincreased, so as to make the scan time of the conventional

troscopages, sce 10 min

acquisition equal to the scan time of the spectroscopicacquisition, the SNRs of the former and the latter are aboutequal (Fig. 6C versus D).

5. Discussion

In this work, proton spectroscopic imaging was deployedas a tool for investigating object structure and object-inducedfield distortions rather than tissue chemistry. This wasachieved by leaving out the water and fat suppressiontechniques — so as to make image contrast governed by theproperties of the bulk water and fat protons — and byadapting the processing and display capabilities of anotherwise standard spectroscopic imaging package. By wayof phantom experiments, non-water-suppressed protonspectroscopic imaging was shown (a) to yield geometricallyaccurate time- and frequency-resolved images, (b) to enablelocalization of object structures in laboratory coordinates and(c) to effectuate geometrical immunity to off-resonanceeffects, such as those caused by field inhomogeneity,susceptibility, chemical shift, f0 offset and field drift.Spectroscopic imaging was further shown (d) to allowdetailed analyses of microscopic and macroscopic fieldinhomogeneities and (e) to provide a powerful base for thederivation of various kinds of undistorted parameter mapsand composite images, e.g., undistorted field maps andpositive contrast images.

In order to achieve geometric fidelity and localization inlaboratory coordinates, several conditions have to besatisfied. First, the point where all three gradient coilsproduce zero magnetic fields, i.e., the isocenter of thegradient system (which typically is the same point as theisocenter of the magnet), has to exhibit a known relation withthe origin of the laboratory frame as, for instance, indicatedby the laser positioning system. Second, the laser positioningsystem has to be exactly aligned with the principal axes ofthe gradient system. Third, images have to be corrected fornonlinearities of the imaging gradients. Once these condi-tions have been satisfied, non-water-suppressed spectro-scopic imaging emerges as an attractive research tool forapplications that put high demands on geometric fidelity,

ic image (D) of a 2.25-cm-diameter cylinder with doped water and a 1.8-mman time 20.4 s; (B) SE with BW=500 Hz/pixel and two signal averages, scan14 s; (D) averaged spectroscopic image, scan time 10 min 14 s. B0 along the

-

93C.J.G. Bakker et al. / Magnetic Resonance Imaging 31 (2013) 86–95

such as radiotherapy treatment planning, stereotactic neuro-surgery, stereotactic laser and radiosurgery, and image-guided interventions, e.g., needle biopsies, and adds to thevalue of MRI as a stand-alone modality for such applica-tions. In addition, geometric fidelity and localization inlaboratory coordinates may prove beneficial in manyoccasions, for instance, when integrating MR images fordifferent nuclei, e.g., 1H and 19F, or with other modalities,e.g., computed tomography, or when there is a need to defineor reproduce the geometry of a scan with respect to externallandmarks, etc.

As compared to conventional imaging, spectroscopicimaging displays some nice features. Apart from providingadditional information (see above), spectroscopic imaging isattractive in that encoding and acquisition are separated, as aresult of which geometric perfection can be achieved withmodest gradient strengths. This is different in conventionalimaging, where the detrimental effects of field disturbancescan be alleviated by using stronger readout gradients, butwhere the maximum gradient capability sets a limit to thefield disturbances that can be accommodated. In addition tothis, it should be realized that the use of very stronggradients may be undesirable for several reasons: (a) theSNR is decreased by increasing the readout bandwidth; (b)rapidly switching strong gradients induces large eddycurrents which are a source of geometric distortionthemselves; (c) rapidly switching strong gradients producesacoustic noise and may invoke physiological effects such asperipheral nerve stimulation.

With regard to applications, it is important to make adistinction between two complementary aspects of geometricfidelity, viz., absence of geometric distortion and localizationin laboratory coordinates. The first property refers to theability to accurately depict the object, even in the presence ofimplants, devices, surgical tools, etc. Obviously, this is adesirable property for procedures in which targets aredefined with respect to external frames or fiducial markers,as is the case in, e.g., breast biopsy and stereotactic surgery,and for procedures in which MRI data are to be matched withdata from other modalities, as in radiotherapy treatmentplanning [8]. The second property, localization in laboratorycoordinates, is a feature that— to our knowledge — has notyet received any attention in the MRI literature. It impliesthat image coordinates can be directly converted intolaboratory coordinates and vice versa. So, what is measuredin the images will also be valid in the laboratory, andconversely, what is planned and measured in the laboratorywill also be found in the images. This also implies that scanplanes can be directly defined or reproduced with respect tothe light visor system, without having to rely on geometri-cally suspect scout views. It also means that the target of anintervention can be directly specified in laboratory co-ordinates. For interventions that do not or only minimallydisturb the geometry of an object, e.g., stereotacticradiosurgery and MR-guided thermal therapies, this featurecould open up new ways to MR-guided treatment.

From the foregoing discussion, MR spectroscopic imagingseems to come out as an ideal tool for detailed time andfrequency domain analyses of object-inducedmicroscopic andmacroscopicfield inhomogeneities and as the gold standard forapplications that put high demands on geometricfidelity. Apartfrom effects associated with a limited coverage of k-space, thisis indeed the case, at least for research purposes. Unfortunately,however, the favorable geometric properties of spectroscopicimaging and its wealth of information go together with a lackof efficiency. Since only one point in k-space is acquired aftereach excitation, as compared to, e.g., 256 points when areadout gradient is applied, spectroscopic imaging tends to beabout two orders of magnitude slower than the correspondingstandard imaging sequence, which imposes a limit to itsclinical utility. In the present study, a two-dimensional setupand small field of views and acquisition matrices were used toachieve adequate spatial resolution, while avoiding excessivescan times and still being able to demonstrate essentialfeatures. For a purely phase-encoded three-dimensional scan,scan times may easily become prohibitive, even more sobecause, in general, an oversized scan volume has to be chosento ensure that all spins in the volume of interest have beenexcited. The situation may even get worse when, in thepresence of extremely large field inhomogeneities, forexample, near steel implants, the maximum available orallowable RF bandwidth becomes too small to cover the fullrange of off-resonance frequencies. Multiband excitationtechniques, like those used in dedicated techniques for imagingnear implants such as slice encoding for metal artifactcorrection (SEMAC) [9] and multi-acquisition variable-resonance image combination (MAVRIC) [10], may beinvoked to overcome this problem at the cost of efficiency.Further, it should be noted that conducting structures, such asmetallic implants, also give rise to RF artifacts, including flipangle errors and phase errors, which may favor FID techniquesover the SE techniques used in our study [11].

To put the efficiency issue in proper perspective, it shouldbe emphasized that our study was performed on a clinicalMR system with standard spectroscopic imaging capabilitiesand that no attempts were made to optimize the sequencewith regard to scan duration, SNR, etc. In the past decades,however, several techniques have been proposed to speed upspectroscopic imaging, e.g., echo-planar techniques [12],multiple spin echo techniques [13], spiral imaging tech-niques [14], parallel imaging techniques [15] and sparsesampling [16]. To further improve efficiency, it can beconsidered to refrain from spectroscopic and temporalinformation altogether and only use the initial amplitude ofthe FID for the reconstruction of a single image. This is infact what is done in so-called single-point imaging (SPI), atechnique proposed by Emid and Creyghton in 1985 [17] andmainly employed in solid-state imaging and MR microscopyever since. To date, applications of SPI in medical researchhave been largely restricted to microimaging of solid andquasi-solid tissue specimens, e.g., bone, cartilage, tendonand teeth [18], and prosthetic materials [19].

Fig. 7. Coronal averaged spectroscopic images of a coaxial cylinder phantom with oil in the 10-mm-diameter inner compartment and doped water in the 2.25-cmdiameter outer compartment; images were retrospectively undersampled and reconstructed with, respectively, 100% (A), 25% (B), 10% (C) and 5% (D) of thecomplex k-space data.

94 C.J.G. Bakker et al. / Magnetic Resonance Imaging 31 (2013) 86–95

As with spectroscopic imaging, many techniques havebeen proposed to speed up the basic SPI experiment,including turbo spin-echo SPI [20–25]; single-point rampedimaging with T1 enhancement, a technique in which one ofthe phase-encode gradients is ramped in equally spaceddiscrete steps [26]; multipoint k-space mapping techniques[27,28]; SPI with variable phase encoding interval [29];and, last but not least, sparse k-space sampling [30]. Thelatter technique, also termed compressed sensing (CS), hasalready been shown to be very effective when imagingobjects which possess a sparse representation in a certaindomain. This has, for instance, been shown to be the casefor active and passive catheters [31,32]. An illustration ofthe potential of CS for one of our phantoms is provided byFig. 7. The figure shows retrospectively reconstructedspectroscopic images of the water–oil phantom withundersampling factors 4, 10 and 20. A single spectroscopicinput image was created by taking the average of 15spectroscopic images around the echo top. Undersamplingand reconstruction were done as described in Ref. [33].

Taking into account rapid developments in sparsesampling, parallel imaging and multichannel transmittechniques, spectroscopic imaging for investigating structurerather than chemistry may be expected to make its way toclinical MRI in the near future. As matters stand at present,spectroscopic imaging techniques may already be consideredfor low-resolution and volume-selective studies, when thegoal is to explore what MRI can achieve with regard togeometric accuracy, or when a gold standard or referencemethod is required against which to measure the effects offield perturbations in conventional MRI. Further, it should berealized that purely phase-encoded imaging is a naturalalternative to conventional imaging when low SNR demandsexcessive signal averaging in conventional imaging. In thatsituation, which may easily arise in, for instance, fluorineand sodium imaging, phase encoding may take the place ofsignal averaging.

6. Conclusion

Non-water-suppressed spectroscopic imaging offers afundamental solution to the geometric deficiencies of MRI.

-

This solution may evolve towards a practical solution —and lend force to MRI as a stand-alone modality forapplications that require geometric accuracy — once fulladvantage will have been taken of recent developments withregard to scan time reduction.

Acknowledgments

Grant sponsor/funding agency: Dutch Technology Foun-dation STW; grant number: 10712.

References

[1] Schenck JF. The role of magnetic susceptibility in magnetic resonanceimaging: MRI magnetic compatibility of the first and second kinds.Med Phys 1996;23:815–50.

[2] Langlois S, Desvignes M, Constans JM, Revenu M. MRI geometricdistortion: a simple approach to correcting the effects of non-lineargradient fields. J Magn Reson Imaging 1999;9:821–31.

[3] El-Sharkaway AM, Schaer M, Bottomley PA, Atalar E. Monitoringand correcting spatio-temporal variations of the MR scanner's staticmagnetic field. MAGMA 2006;19:223–6.

[4] Haacke EM, Brown RW, Thompson MR, Venkatesan R. Magneticresonance imaging: physical principles and sequence design. NewYork: Wiley; 1999.

[5] Glover PM. Topical review: interaction of MRI field gradients with thehuman body. Phys Med Biol 2009;54:R99–115.

[6] Brown TR, Kincaid BM, Ugurbil K. NMR chemical shift imaging inthree dimensions. Proc Natl Acad Sci U S A 1982;79:3523–6.

[7] Brown TR. Practical applications of chemical shift imaging. NMRBiomed 1992;5:238–43.

[8] Hajnal JV, Hill DLG, Hawkes DJ. Medical image registration. NewYork: CRC Press; 2001.

[9] Lu W, Pauly KB, Gold GE, Pauly JM, Hargreaves BA. SEMAC: sliceencoding for metal artifact correction in MRI. Magn Reson Med2009;62:66–76.

[10] Koch KM, Lorbiecki JE, Hinks RS, King KF. A multispectral three-dimensional acquisition technique for imaging near metal implants.Magn Reson Med 2009;61:381–90.

[11] Camacho CR, Plewes DB, Henkelman RM. Nonsusceptibility artifactsdue to metallic objects in MR imaging. J Magn Reson Imaging 1995;5:75–88.

[12] Mansfield P. Spatial mapping of chemical shift in NMR. Magn ResonMed 1984;1:370–86.

[13] Duyn JH, Moonen CT. Fast proton spectroscopic imaging ofhuman brain using multiple spin-echoes. Magn Reson Med 1993;30:409–14.

95C.J.G. Bakker et al. / Magnetic Resonance Imaging 31 (2013) 86–95

[14] Adalsteinsson E, Irarrazabal P, Topp S, Meyer C, Macovski A,Spielman DM. Volumetric spectroscopic imaging with spiral-based k-space trajectories. Magn Reson Med 1998;39:889–98.

[15] Dydak A, Pruessman KP, Weiger M, Tsao J, Meier D, Boesiger P.Parallel spectroscopic imaging with spin-echo trains. Magn ResonMed2003;50:196–200.

[16] Larson PE, Hu S, Lustig M, Kerr AB, Nelson SJ, Kurhanewicz J, et al.Fast dynamic 3D MR spectroscopic imaging with compressed sensingand multiband excitation pulses for hyperpolarized 13C studies. MagnReson Med 2011;65:610–9.

[17] Emid S, Creyghton JHN. High resolution NMR imaging in solids.Phys B 1985;128:81–3.

[18] Gruwel MLH, Latta P, Tanasiewicz M, Volotovskyy V, Sramek M,Tomanek B. MR imaging of teeth using a silent single point imagingtechnique. Appl Phys A 2007;88:763–7.

[19] Ramos-Cabrer P, van Duynhoven JPM, van der Toorn A, Nicolay K.MRI of hip prostheses using single-point methods: in vitro studiestowards the artifact-free imaging of individuals with metal implants.Magn Reson Imaging 2004;22:1097–103.

[20] Bendel P. Echo projection imaging— a method to obtain NMR imagesundistorted by magnetic field inhomogeneities. IEEE Trans MedImaging 1985;4:114–9.

[21] Miller JB, Garroway AN. Removal of static field inhomogeneity andchemical shift effects in NMR imaging. J Magn Reson 1986;67:575–9.

[22] Wong TS, Rosenfeld D. A technique for NMR imaging undermagnetic fields with high field inhomogeneities. IEEE Trans MedImaging 1987;6:148–56.

[23] Sharp JC, Bowtell RW, Mansfield P. Elimination of susceptibilitydistortions and reduction of diffusion attenuation in NMR microscopyby line-narrowed 2DFT. Magn Reson Med 1993;29:407–11.

[24] McIntyre DJO, Hennel F, Morris PG. SPARE: a robust method formagnetic resonance imaging in inhomogeneous fields. J Magn Reson1998;130:58–62.

[25] Beya SD, Balcom BJ, Mastikhin IV, Bremner TW, Armstrong RL,Grattan-Bellew PE. Imaging of heterogeneous materials with a turbospin echo single-point imaging technique. J Magn Reson 2000;144:255–65.

[26] Balcom BJ, MacGregor RP, Beya SD, Green DP, Armstrong RL,Bremner TW. Single-point ramped imaging with T1 enhancement(SPRITE). J Magn Reson A 1996;123:131–4.

[27] Cho ZH, Ro YM. Multipoint k-space point mapping (KPM) techniquefor NMR microscopy. Magn Reson Med 1994;32:258–62.

[28] Fernandez-Seara MA, Wehrli SL, Wehrli FW. Multipoint mapping forimaging of semi-solid materials. J Magn Reson 2003;160:144–50.

[29] Latta P, Gruwel MLH, Volotovskyy V, Weber MH, Tomanek B.Single-point imaging with a variable phase encoding interval. MagnReson Imaging 2008;26:109–16.

[30] Parasoglou P, Malioutov D, Sederman AJ, Rasburn J, Powell H,Gladden LF, et al. Quantitative single point imaging with compressedsensing. J Magn Reson 2009;201:72–80.

[31] Schirra CO, Weiss S, Krueger S, Pedersen SF, Razavi R, Schaeffter T,et al. Toward true 3D visualization of active catheters usingcompressed sensing. Magn Reson Med 2009;62:341–7.

[32] Yerly J, Lauzon ML, Chen HS, Frayne R. A simulation-based analysisof the potential of compressed sensing for accelerating passive cathetervisualization in endovascular therapy. Magn Reson Med 2010;63:473–83.

[33] Lustig M, Donoho D, Pauly JM. Sparse MRI: the application ofcompressed sensing for rapid MR imaging. Magn ResonMed 2007;58:1182–95.