Rapid communication

Functional magnetic resonance imaging with intermolecularmultiple-quantum coherences

Wolfgang Richtera,*, Marlene Richtera,b, Warren S. Warrenc, Hellmut Merkleb,Peter Andersenb, Gregor Adrianyb, Kamil Ugurbilb

aInstitut du biodiagnostic, Conseil national de recherches Canada, Winnipeg, Manitoba, CanadabCenter for Magnetic Resonance Research, University of Minnesota, Minneapolis, MN, USA

cDepartment of Chemistry, Princeton University, Princeton, NJ, USA

Received 28 December 1999; accepted 3 February 2000

Abstract

For the first time, we demonstrate here functional magnetic resonance imaging (fMRI) using intermolecular multiple-quantumcoherences (iMQCs). iMQCs are normally not observed in liquid-state NMR because dipolar interactions between spins average to zero.If the magnetic isotropy of the sample is broken through the use of magnetic field gradients, dipolar couplings can reappear, and henceiMQCs can be observed. Conventional (BOLD) fMRI measures susceptibility variations averaged over each voxel. In the experimentperformed here, the sensitivity of iMQCs to frequency variations over mesoscopic and well-defined distances is exploited. We show thatiMQC contrast is qualitatively and quantitatively different from BOLD contrast in a visual stimulation task. While the number of activatedpixels is smaller in iMQC contrast, the intensity change in some pixels exceeds that of BOLD contrast severalfold. © 2000 Elsevier ScienceInc. All rights reserved.

Keywords:fMRI; Intermolecular multiple quantum coherences; Brain function

1. Introduction

1.1. Intermolecular multiple-quantum coherences

The phenomenon of intermolecular multiple quantumcoherences (iMQCs) was first described a few years ago[1–4]. In NMR, we candirectly observe only single-quan-tum single-spin coherences (this corresponds to magnetiza-tion). Two-dimensional NMR methods [5] make it possibleto observe other coherences between states in a multi-spinsystem as well; these coherences are made to evolve ‘si-lently’ during a successively incremented time interval andare detected after transformation into magnetization. In or-der for this transformation to take place, a net couplingbetween the spins involved must exist. In the case of spins1/2 (such as the hydrogen nucleus), we can distinguish twotypes of couplings: scalar couplings, which act through

chemical bonds, and dipolar couplings, which are muchstronger than scalar couplings and act through space. Scalarcouplings are commonly used to effect the abovementionedtransformation of multi-spin coherences into magnetization;hence 2-D NMR method can provide information about theconnectivity between spins in a molecule. On the otherhand, dipolar couplings are normally not observable inliquids. The dipolar coupling strength between two spinsscales as

Dij } 3cos2u 2 1

whereu is the angle between the interspin vector and themain magnetic field. This coupling averages to zero whenintegrated over all directions (the surface of a sphere). In thecase of short-range dipolar interactions, the interspin vectorsamples all directions on an NMR time scale through mo-lecular diffusion.

This is not true for long-range dipolar interactions, whereu is almost constant in time. However, in that case thedistribution of spins is quasi-continuous, and the dipolarinteractions average to zero in space as long as the liquid ismagnetically isotropic. Magnetic field gradient pulses ap-

* Corresponding author. Institute for Biodiagnostics, National Re-search Council, 435 Ellice Ave., Winnipeg MB R3B 1Y6 Canada. Tel.:11-204-984-6564; fax:11-204-984-7036.

E-mail address:richter@ibd.nrc.ca (W. Richter).

Magnetic Resonance Imaging 18 (2000) 489–494

0730-725X/00/$ – see front matter © 2000 Elsevier Science Inc. All rights reserved.PII: S0730-725X(00)00133-8

plied during the experiment can break this isotropy; hencelong-range dipolar couplings can be reintroduced by theexperimenter in a controlled fashion. In this manner, thestructure of a sample can be probed on the distance scale onwhich the dipolar couplings act. Most importantly, thisdistance may be tuned through the choice of experimentalparameters. The gradient pulse causing the reappearance ofdipolar couplings (the so-called ‘correlation gradient’) canbe thought to wind up a helix of magnetization along itsaxis; the dipolar interactions that are reintroduced in thismanner act over a distance scale of approximately one halfpitch of that helix. Hence the larger the area under thegradient pulse is, the shorter is the correlation distance. Inpractice, the correlation distance ranges from tens to hun-dreds of micrometers, which is of course far above themicroscopic range that is provided by scalar couplings, yetbelow the size of an imaging voxel. The method describedhere is therefore among the few NMR methods that canprovide structural information on a mesoscopic distance scale.

1.2. Functional MRI (fMRI)

FMRI was first demonstrated a few years ago [6–8] andis today one of the most powerful neuroimaging techniques.FMRI can measure brain activity noninvasively, with aspatial resolution of millimeters and a temporal resolutionof seconds. Most fMRI techniques are based on the ‘bloodoxygen level dependent’ (BOLD) effect. The BOLD effectis thought to arise from localized changes in the concentra-tion of the strongly paramagnetic deoxyhemoglobin mole-cules in the brain, which are coupled to alterations in neu-ronal activity. Blood flow increases within seconds near thesite of activation and overcompensates for the increasedmetabolic demand, resulting in decreased deoxy- and in-creased (diamagnetic) oxy-hemoglobin contents. The ensu-ing changes in susceptibility gradients across capillaries andvenous blood vessels result in an increase of the apparenttransverse relaxation time (T2

*) of the spins [9]. Therefore,an image whose intensity is weighted by T2

* will showneuronal activation, through the secondary effect of bloodoxygenation, as an increase in signal intensity. The typicalsignal increase upon activation is on the order of a fewpercent; hence this method is relatively insensitive.

The application of iMQCs to fMRI is based on therationale that multi-spin coherences themselves do have adifferent, and possibly higher, sensitivity to susceptibilitygradients than single-spin coherences. They may also bemore specific to thesite of activation. For example, anintermolecular zero-quantum coherence (iZQC) evolves atthedifferenceof the single-quantum frequencies of the twospins involved, hence the zero-quantum signal intensity is afunction of thedistribution of susceptibility gradients. TheBOLD signal, on the other hand, is a function of theaveragestrength of those gradients within a voxel. Hence iZQCcontrast isfundamentallydifferent from T2

* contrast [this hasbeen shown previously by our group [3]], but still a function

of blood oxygenation. Furthermore,the choice of correla-tion distance provides an additional degree of freedom tooptimize contrast, which conventional methods do not pos-sess. Susceptibility gradients due to deoxyhemoglobin arisefrom blood vessels ranging from densely distributed capil-laries to low density large veins. The latter are clearlyundesirable in fMRI experiments since they will not haveaccurate correspondence with the actual sites of enhancedneuronal activity. iMQCs offer the possibility of altering thedistance scale over which the dipolar-couplings lead toobservable signal, and hence the sensitivity to blood vesselsof different sizes. Here, we provide initial evidence thatiMQCs (here actually intermolecular double quantum co-herences that are, in our pulse sequence, made to depend onthe frequencydifferencebetween the spins, like iZQCs) canindeed generate large contrast related to functional activa-tion of the brain.

2. Methods

2.1. Hardware

Experiments were carried out with a whole-body 7 Teslaimaging system (Varian/Magnex) with a head gradient in-sert and a double-loop surface coil.

2.2. Pulse sequences

The pulse sequence used for iMQC contrast is shown inFig. 1. It is a modification of previously used pulse se-quences, with the goal to minimize signal fluctuations. Thissequence was discussed in ref. [3] and selects for double-quantum coherences (iDQCs). Hence, after the first r.f.pulse a double quantum coherence between spins 1 and 2,such as (Ix1Ix2 2 Iy1Iy2) evolves at the sum of the individualresonance frequencies (Dv1 1 Dv2), for a duration oftdq.

Fig. 1. iDQC pulse sequence. The iDQC preparation period (including thepair of correlation gradients) is followed by a double spin echo and asegmented EPI readout sequence. Note that the timing is modified from thesimplified version that is described in the text, on account of the doublespin echo. The effective evolution time after the second gradient is 2tdq, asrequired.

490 W. Richter et al. / Magnetic Resonance Imaging 18 (2000) 489–494

The second r.f. pulse transforms this into two-spin, single-quantum coherences, such as (Ix1Iz2 1 Ix2Iz1); those coher-ences evolve at the single-quantum frequencies6 Dv1

and6 Dv2, respectively. After a time interval of 2*tdq, apartial coherence transfer echo forms, with a total phaseevolution of6 tdq*(Dv1 2 Dv2). In order to read out theimage, we then used a conventional (segmented) echo pla-nar imaging sequence. Note that the detected signal (includ-ing evolution over both time periods) is a function of theresonance frequencydifference of the two spins (eventhough the DQ coherence evolves at the sum of the reso-nance frequencies). This functional dependence is the sameas that of a zero-quantum coherence (iZQC). However, thesecond gradient, which is not present in the previouslypublished pure zero-quantum version of this pulse se-quence, helps cancel spurious magnetization.

Relevant pulse sequence parameters were the following:matrix size5 64 3 64; TR5 3.5 s; NEX5 2 (with phasecycling of the first pulse); 2 segments with linear k-spacesampling (hence the total imaging time was 14 s per vol-ume);tdq 5 0 (this means that the double quantum signalevolution takes place entirely during the correlation gradi-ent);t1 5 3.8 ms;t2 5 23.7 ms;t3 5 27.5 ms; duration offirst correlation gradient5 8.27 ms; G5 1.5 G/cm. Thecorrelation distance was, therefore, on the order of 100mm [1].

Conventional BOLD experiments were carried out with asegmented Gradient-Echo EPI pulse sequence: matrixsize5 64 3 64; TR 5 3.5 s; NEX5 1; 2 segments withcenter-out k-space sampling (for an imaging time of 7 s pervolume); an effective echo time (to the first echo) of 5 ms;and a nominal echo time (to the center of the echo train) of19 ms.

2.3. Human subjects

Seven normal subjects (six men and one woman) werestudied, in accordance with the guidelines of the Institu-tional Review Board of the University of Minnesota. In-formed consent was obtained from all subjects. Technicallyacceptable data were obtained from five subjects (four menand one woman); only those were considered further.

2.4. Visual stimulation task

Monocular or binocular visual stimulation (red LEDs,flashing at 8 Hz) was provided through commercial goggles(Grass Instruments, Quincy, MA, USA).

2.5. Experimental protocol

First, a series of anatomic FLASH images was obtained.From those images, a single slice (along one of the Carte-sian axes) containing part of the primary visual cortex wasselected. Next, a series of iDQC images with differentcorrelation gradients was prepared, in order to assess andevaluate the properties of the iDQC signal. Then, a conven-

tional (BOLD) gradient-echo EPI functional imaging exper-iment was carried out (50 volumes, three rest periods andtwo stimulation periods of 10 volumes each). After that, aniDQC functional imaging experiment was carried out (25volumes, three rest periods and two stimulation periods of 5volumes each). For several subjects, another pair of BOLDand iDQC experiments was carried out, with a differentslice orientation. In total, we obtained four coronal, threeaxial, and one sagittal data set from the five subjects.

2.6. Data analysis

Data were analyzed using the software package Stimu-late [10] and various software routines written in InteractiveData Language (Research Systems, Boulder, CO, USA).Functional maps were obtained by calculating Pearson’scorrelation coefficient between each pixel’s time course andthe stimulation (off/on) waveform. Hemodynamic lag wastaken into account by shifting the reference waveform by oneor two volumes; the highest correlation coefficient resultingfrom this multiple comparison was retained. Maps were thresh-olded to an uncorrected confidence interval of 0.01. Addition-ally, spatial filtering was applied so that only spatial clustersof at least two activated pixels were included in the maps.

3. Results

3.1. Authenticity of the iDQC signal

An important property of iMQCs is the overall scaling ofthe signal by a factor of (3 cos2u 21) (this is similar to thescaling factor for dipolar couplings), where in this caseu isthe angle between the main magnetic field and the correla-tion gradient direction. Accordingly, we expect that thesignal is twice as large for the longitudinal (z) gradientdirection as for the transverse (x or y) gradient direction,and that the phase difference between them isp. This isshown in Fig. 2. Shown on the top are two phase imageswith z and x correlation gradients, respectively. Note thatthe phase difference pixel-by-pixel is approximatelyp in-deed. On the bottom, the average magnitude of the signal inthe whole slice is shown. Note that the magnitude for the zgradient is indeed twice that of the transverse gradients.Furthermore, an analysis of signal-to-noise ratios (SNRs)shows that the SNR for the iDQC images is, on average,reduced by a factor of 6.9 from the GE-EPI images; this ison the order of magnitude of our expectations from theo-retical considerations [3]. Hence we are confident that thesignal comes essentially from iDQCs.

3.2. Anatomic and functional contrast

All data sets evaluated showed activation throughout theoccipital lobe in both the BOLD and the iDQC experiments.A representative data set is shown in Fig. 3. On the left, we

491W. Richter et al. / Magnetic Resonance Imaging 18 (2000) 489–494

show a conventional GE-EPI image in gray scale. Super-imposed on it, in color, is the corresponding BOLD activa-tion map. On the right, we show an iDQC image in grayscale, with the iDQC activation map in color superimposed.In both images and maps, Gaussian smoothing (interpola-tion) was performed for display purposes only. Note that the

anatomic contrast is fundamentally different between thetwo images, as we expect according to the mechanismpostulated. The two activation maps are qualitatively dif-ferent from one another as well; the iDQC map is consid-erably more sparse, and the foci of activation are onlypartially overlapping.

Fig. 2. Phase and magnitude of the iDQC signal for two different correlation gradient directions. Note that the the signals for longitudinal and transversegradients arep out of phase, and that their intensity is different by approximately a factor of two, in accordance with the theoretical model for iDQCs.

Fig. 3. BOLD and iDQC activation maps on a background of GE-EPI and iDQC contrast. Note that the anatomic contrast is fundamentally different, as arethe functional maps.

492 W. Richter et al. / Magnetic Resonance Imaging 18 (2000) 489–494

In order to quantify this result, we calculated four parame-ters for each map. These are shown in Table 1. Here weindicate the slice orientation (coronal, axial, sagittal) and thestimulation (left eye, right eye, both eyes). We also give thenumber of pixels in the region of interest (ROI), which was thepart of the occipital lobe intersected by the slice. We then givethe number of pixels activated in the iDQC and BOLD exper-iment, respectively, and the overlap of those two sets. Note thatin all subjects but one, some but not all pixels activated in theiDQC experiment are also activated in the BOLD experiment.

An important issue of interest is the intensity changeupon activation. Inspection of the data sets revealed that thebaseline fluctuations were considerably larger for the iDQCsignal than for the BOLD signal. This is not unexpected,given the presence of the two gradient pulses, introducing apotential source of instability that is absent in the BOLDexperiment. At the same time, the signal intensity changefor the activated pixels was larger (on the order of 10%increase from baseline) for the iDQC signal as well. Thus,the smaller “activation” volumes in the iDQC data mayinitially be thought to be an artifact of the method of analysis,arising from the larger overall signal fluctuations. We investi-gated this by comparing the pixels that werecommonlyacti-vated in both experiments (corresponding to the last row inTable 1). In Fig. 4, we graph the relative intensity change of

those pixels in both methods. Note that for almost all pixels theiDQC intensity change is larger than the BOLD intensitychange (these are the points below the x5 y line). Theanalysis reveals no significant correlation between these twoquantities (confidence interval. 0.2). However, the aver-age intensity change in these pixels was 10.0% for the iDQCsignal, as compared to 4.2% for the average BOLD signalchange. In Fig. 5, we show timecourses from commonlyactivated pixels in one subject. Horizontal bars denote theactivation periods. Again, the signal intensity change ismuch larger in the iDQC experiment than in the BOLDexperiment, while the baseline fluctuation is larger as well.Note that these time courses are not representative of pixelswith the highest correlation in either method.

4. Discussion

This experiment constitutes the first demonstration offunctional activation revealed by iMQCs. At this moment,we lack a physiological model to relate the observed acti-vation to blood flow, blood oxygenation, blood volume,vascular structure, oxygen consumption, and other poten-tially relevant parameters quantitatively; this is also largelytrue for the BOLD mechanism. We find it particularly note-worthy that the foci of activation are partially, but notcompletely, congruent in the two methods. This observationstrongly suggests that iDQC contrast is fundamentally dif-ferent from BOLD contrast, but is also sensitive to changessubsequent to neuronal activation. These results are in ac-cordance with our expectations.

The functional signal change in commonly activatedpixels was on average two to three times larger in the iDQCmethod compared to BOLD for the particular echo timeused in the BOLD experiments. This number has to beconsidered carefully, as a fair comparison between the twomethods is by no means straightforward. Signal intensitychanges depend on a variety of factors, many of which arerooted in the technical details of each experiment, and not ina physiological phenomenon. In this preliminary effort, asystematic comparison of “contrast-to-noise” was not per-formed either for the BOLD or for the iDQC studies. Thiswould also require that differential sensitivity to differentvessel sizes is considered, and that the comparison is per-formed when the two methods are displaying changes cou-

Table 1Activited pixels in the ROI with either method alone and in both methods simultaneously

Subject #

1 (sag)both eyes

1 (cor)both eyes

1 (ax)both eyes

2 (cor)both eyes

3 (ax)left eye

3 (ax)right eye

4 (cor)both eyes

5 (cor)both eyes

Total ROI 557 694 1091 962 927 927 806 723iDQC 18 37 12 5 11 6 28 2BOLD 241 276 320 83 119 115 288 67Overlap 15 26 7 1 4 4 12 0

Fig. 4. BOLD and iDQC signal intensity changes in commonly activatedpixels. Note that the iDQC intensity changes are generally larger, at leastfor this set of parameters.

493W. Richter et al. / Magnetic Resonance Imaging 18 (2000) 489–494

pled to neuronal activity for similar types of blood vessels.In this respect, it would be more appropriate to comparespin-echo BOLD studies with the current iDQC experiment.However, we can confidently state thatthere exist pixelsthat are functionally activated, in which the iDQC signalshows a larger intensity change than the BOLD signal.

The activation maps are much more sparse in the iDQCmethod. The reason for this may be purely that the signal ismuch less stable in the present implementation of the iDQCmethod, thereby obscuring true activation. This could berectified in future studies, for example with the use ofnavigator-echo type approaches. However, this observationmay also reflect true differences in the physiological param-eters that actually influence the signal. It should also bepointed out that the correlation distance, which is possiblythe most crucial parameter determining contrast, was notoptimized or investigated in this preliminary study. A care-ful investigation and optimization of the iDQC approach indetecting signal changes coupled to neuronal activity re-mains to be performed. However, we find it encouragingthat a large signal change was observed in the iDQCmethod.

5. Conclusion

We demonstrate here that the iMQC method yields sig-nal changes in the brain that are coupled to alterations in

neuronal activity. Future work will have to explore methodsto increase the stability of the signal toward physiologicaland instrumental fluctuations, and the influence of the cor-relation distance on the observed contrast. The iMQCmethod may become an invaluable tool for the elucidationof brain structure and function.

Acknowledgments

Supported by NIH National Resources grant RR08079(University of Minnesota), the Keck Foundation, and NIHGM 35253 and the McKnight Foundation (Princeton Uni-versity).

References

[1] Warren WS, Richter W, Hamilton Andreotti A, Farmer BT II. Gen-eration of impossible cross peaks in solution nuclear magnetic reso-nance. Science 1993;262:2005.

[2] Richter W, Lee S, Warren WS, He Q. Imaging with intermolecularmultiple quantum coherences. Science 1995;267:654–7 .

[3] Warren WS, Ahn S, Rizi R, Hopkins J, Leigh JS, Richter W, MescherM, Garwood M, Ugurbil K. MR imaging contrast enhancement basedon intermolecular zero quantum coherences. Science 1998;281:247–51.

[4] Lee S, Richter W, Vathyam S, Warren WS. Quantum treatment of theeffects of dipole-dipole interactions in liquid nuclear magnetic reso-nance J Chem Phys 1995;105:874–900.

[5] Jeener J. Ampere International Summer School: Baski Polje, Yugo-slavia, 1972.

[6] Bandettini PA, Wong EC, Hinks RS, Tikofsky RS, Hyde JS. Timecourse EPI of human brain function during task activation. MagnReson Med 1992;25:390–7.

[7] Kwong KK, Belliveau JW, Chesler DA, Goldberg IE, Weisskoff RM,Poncelet BP, Kennedy DN, Hoppel BE, Cohen MS, Turner R, et al.Dynamic magnetic resonance imaging of human brain activity duringprimary sensory stimulation. Proc Natl Acad Sci USA 1992;89:5675–9.

[8] Ogawa S, Tank DW, Menon R, Ellermann JM, Kim S-G, Merkle H,Ugurbil K. 4 Tesla gradient recalled echo characteristics of photicstimulation-induced signal changes in the human primary visual cor-tex. Proc Natl Acad Sci USA 1992;89:5951–5.

[9] Ugurbil K, Ogawa S, Kim S-G, Chen W, Zhu XH. Imaging brainactivity using nuclear spins. In: Maraviglia B, editor. Magnetic Res-onance and Brain Function: Approaches from Physics. Amsterdam:BIOS Press, 1999. p. 261–310.

[10] Strupp, JP. A GUI based fMRI analysis software package. Neuroim-age 1996;3:S603.

Fig. 5. Time courses from activated pixels for BOLD (bottom) and iDQC(top) methods. Note that the signal intensity change is larger in the iDQCtime course, as are the baseline fluctuations.

494 W. Richter et al. / Magnetic Resonance Imaging 18 (2000) 489–494