Inability to Directly Detect Magnetic Field ChangesAssociated With Neuronal Activity

Laura M. Parkes,* Floris P. de Lange, Pascal Fries, Ivan Toni, and David G. Norris

The ability to directly detect neuronal magnetic fields by MRIwould help investigators achieve the “holy grail” of neuroimag-ing, namely both high spatial and temporal resolution. Bothpositive and negative findings have been reported in the litera-ture, with no clear consensus as to the feasibility of directdetection. The aim of this study was to replicate one of the mostpromising published in vivo results. A second aim was to inves-tigate the use of steady-state visual evoked potentials (ssVEPs),which give a large evoked response and offer a well-controlledapproach because the frequency of the neuronal response canbe dictated by the experimenter. For both studies we used ageneral linear model (GLM) that included regressors for boththe expected blood oxygen level-dependent (BOLD) signal andthe magnetic source (MS) signal. The results showed no activitythat could be attributed to the neuromagnetic signals in eitherstudy, and no frequency component corresponding to the fre-quency of the ssVEPs. This study demonstrates that for theparticular stimuli and hardware used, the sensitivity of the mag-nitude MRI signal to detect evoked neuronal currents is too lowto be of practical use. Magn Reson Med 57:411–416, 2007.© 2007 Wiley-Liss, Inc.

Key words: direct detection; neuronal activity; visual evokedpotentials; fMRI; ssVEP

There is continuing interest in the possibility of using MRIto directly detect the magnetic fields produced by neurons(1–4). This would allow direct measurement of neuronalactivity at high temporal resolution with unambiguouslocalization. This would be an improvement over the cur-rent leading neuroimaging techniques of functional MRI(fMRI), which has low temporal resolution, and magne-toencphalography (MEG), which offers no unique localiza-tion of activity. Earlier theoretical work showed that themagnetic fields produced by some of the brain’s strongestelectrical activity, such as the alpha rhythm, should belarge enough to be detected by MRI (5,6). While in vitrowork has shown promise (7,8), no in vivo study has yetbeen reliably reproduced.

In theory, the magnetic fields generated by neuronscould affect both the magnitude and phase of the MRsignal. Following synaptic activity, post-synaptic ioniccurrents flow along the dendrites, creating the currentdipoles seen with EEG and MEG. Ionic currents also flowalong axons in the form of action potentials. These cur-rents induce transient magnetic fields that can locally alterthe precession rate of water molecules, and hence alter the

phase and magnitude of the MR signal. Given the millionsof dendrites or axons that are present in a single MR voxel,any overall phase variation at the source of activity wouldbe expected to average out to zero, whereas loss of signalmagnitude caused by intravoxel dephasing should remaindue to microscopic field inhomogeneities. At a distancefrom the source, however, the phase variations add up togive a resultant phase change, as demonstrated in twophantom studies using an electric current dipole (5,7).Disappointingly, in vivo results obtained with phase sig-nals are variable (1–4). The most promising in vivo studieshave found changes in the signal magnitude (4,9,10), andtherefore we considered only the magnitude signal in thisstudy.

The aim of this work was twofold. The first goal was toreplicate one of the most positive in vivo results from aprevious study (9) that used a well-established visuomotorparadigm. That study showed clustered magnetic source(MS) activity with locations and latencies consistent withthe electrophysiology and fMRI literature. In a secondexperiment, we investigated the use of steady-state visualevoked potentials (ssVEPs), a well-controlled approachthat has also shown promise (10). ssVEPs are an attractivechoice of stimulation because they give a strong electro-physiological response, and the frequency of the responsecan be completely controlled. If the magnetic fields fromthis activity can be detected, and the frequency of themeasured signal follows the frequency of the driving stim-ulus, this would be strong evidence for the ability of MRIto directly detect neural activity.

MATERIALS AND METHODS

Experiment 1: Visuomotor Response Task

For experiment 1 we closely followed the experimentaldesign of Ref. 9. Six subjects (three females and threemales, 21–35 years old) took part.

Stimulus

The stimulus consisted of a 50-ms checker wedge pre-sented in the lower right visual field, to which the subjecthad to respond as quickly as possible with a left index-finger button press. The stimulus was presented at randomintervals of 0.6–1.5 s (mean interval � 1 s, step size �0.1 s) in six blocks of 120 s interspersed with 30 s of rest,for a total time of approximately 15 min. This designdiffered from Xiong et al.’s (9) original study in two ways.First, Xiong et al. (9) used a longer mean interstimulustime of 2 s, and second, they achieved an effective jitter of1.8–2.2 s (step size � 0.1 s) in the original study bychanging the time of the scanning acquisition relative tothe stimulus in five separate blocks, such that there was no

F.C. Donders Center for Cognitive Neuroimaging, Radboud University, Nijme-gen, The Netherlands.*Correspondence to: Dr. Laura M. Parkes, Magnetic Resonance and ImageAnalysis Research Centre, University of Liverpool, Liverpool L69 3GE, UK.E-mail: laupar@liverpool.ac.ukReceived 18 May 2006; revised 11 September 2006; accepted 3 October2006.DOI 10.1002/mrm.21129Published online in Wiley InterScience (www.interscience.wiley.com).

Magnetic Resonance in Medicine 57:411–416 (2007)

© 2007 Wiley-Liss, Inc. 411

jitter within each block. The shorter mean interval be-tween stimuli used in the present study minimizes anypotential signal change due to the BOLD effect, which is apossible confound in Xiong et al.’s study, as will be dis-cussed below. The jitter within each block prevents thesubject from predicting the appearance of the checkerwedge and thus helps to maintain attention.

Scanning Parameters

Scanning was performed on a 1.5T Siemens system usinga multiecho EPI sequence with prospective motion correc-tion (11) and three echo times (TEs) of 30, 60, and 90 ms.This allowed the optimum BOLD signal to be obtained at aTE of 60 ms (on the basis that this matches the T2*), andthe MS signal to be obtained at the longer TE of 90 ms,which has been argued to be close to the optimum TE (9).A longer TE is optimum because it allows a longer periodfor intravoxel dephasing due to the neuronal magneticfields (9). Five slices with an in-plane resolution of 3.5 mmwere carefully positioned to cover both the primary motorand visual cortices. A repetition time (TR) of 1 s gave ainterslice time of 200 ms. The stimulus presentation wassynchronized with the scan timing such that the flashesoccurred 50 ms or 150 ms preceding slice collection.

Analysis

We considered only the magnitude of the MR signal,which gave positive results in the original study (9). Thedata were analyzed using BrainVoyager (Brain Innova-tions, Maastricht, The Netherlands) on an individual basis.Preprocessing consisted of linear trend removal and a4-mm 2D spatial filter only. A general linear model (GLM)was produced for each of the five imaging slices, whichconsisted of one BOLD regressor and four MS regressors.The four MS regressors comprised two motor and two

visual regressors: one for the interval between 100–200 mspreceding slice collection, and one for the interval be-tween 0–100 ms preceding slice collection, as shown inFig. 1. For each slice from each subject, regions of “acti-vation” (i.e., regions in which a significant amount ofvariance in the MR signal time course is accounted for bythe regressor) associated with each regressor were care-fully assessed by eye. A low statistical threshold of P �0.01 uncorrected was chosen in order to fully consider anyweakly activated regions. The analysis was carried out onthe image sets collected at TEs of 60 ms and 90 ms.

Experiment 2: ssVEPs

Three subjects (two females and one male; 24, 35, and 24years old, respectively) took part in experiment 2.

Stimulus

The stimulus consisted of a 6-Hz flashing checkerboardpresented in 12 blocks of 120 s interspersed with 120 s ofrest, for a total time of 48 min. This is similar to themethod used in Ref. 10, except that in the previous studya 10-Hz flashing checkerboard was used. Throughout theexperiment, a central fixation cross was displayed thatrandomly changed color (red, blue, and green). The subjectwas instructed to fixate on the cross and respond to thecolor changes by pressing one of two buttons.

Scanning Parameters

Scanning was performed on a 3T Siemens system using anEPI sequence with prospective motion correction (11), aTE of 50 ms, and a TR of 1050 ms. Six slices of 3.5-mm3

resolution were positioned over the primary visual cortex.Following the design in Ref. 10, a dead time of 50 ms wasintroduced at the end of each TR, such that each slice was

FIG. 1. Time course and regressors for ex-periment 1. The figure shows stimulus tim-ing on the top row and slice timing on thesecond row, with slices corresponding topositions as shown in the picture. For ex-ample, if a visual stimulus occurs 0–100 msbefore slice collection, as is the case for thefirst visual stimulus for slice 1, the visualregressor 1 for that slice is set to one (row3). At other times, this regressor is set tozero.

412 Parkes et al.

collected at the same time with respect to the 6-Hz flash,but each volume was sampled at a different time withrespect to the flashes. This produces the effect that themeasurement slowly subsamples the 6-Hz signal, asshown in Fig. 2. Sampling theory states that if a periodicsignal with frequency f1 is subsampled at frequency f2, themeasured signal will have frequency components of �f1 �n * f2, where n is an integer. In our case, f1� 6 Hz and f2 �0.9524 Hz, giving an expected frequency component of0.286 Hz within the measurable range of 0–0.5 Hz.

Analysis

As for experiment 1, only the magnitude of the MR signalwas considered. The data were analyzed using BrainVoy-ager (Brain Innovations, Maastricht, The Netherlands) onan individual basis. Preprocessing consisted of lineartrend removal and a 4-mm 2D spatial filter only. Theelegant design of this study means that a single GLM isappropriate for all slices comprising the envelope BOLDresponse plus two MS regressors, since all slices have thesame timing with respect to the flashes. The MS regressorswere constructed using sine and cosine waves (wavelengthcorresponding to 6 Hz), subsampled according to the tim-ing of the scan triggers, as shown in Fig. 2. Although theneuronal bursts in response to the light flashes are brief,the dendritic currents purported to cause the MS signalextend over several tens of milliseconds, producing thecharacteristic oscillatory signal seen in EEG (12). At stim-ulus frequencies below 10 Hz, EEG power is largely con-tained in the fundamental harmonic (13), which suggeststhat a sinusoidal model is accurate. Both sine and cosinewaves were included because a linear combination ofthese regressors could allow for any phase in the MS signalwith respect to the neural activity. Hence, detection ofactivation is insensitive to the exact delay time betweenthe flash of light and the neural response. It is important toset the average magnitude of the MS regressors during theflash period equal to the resting baseline value. Otherwise,the MS regressors will simply “pick up” the signal differ-ence between the rest and active periods that were notcompletely modeled by the BOLD regressor. For example,the signal at the onset and offset of activation is unlikely to

be perfectly modeled by the BOLD regressor due to inac-curacies in the assumed delay and dispersion.

In each subject, we carefully assessed regions of “acti-vation” associated with each regressor by eye. Frequencyanalysis was also performed on the time-course signalfrom regions of interest (ROIs) in the visual cortex. Tworegions were chosen: one containing approximately 100voxels where the strongest BOLD activation was seen, andone containing approximately 20 voxels in the occipitalcortex where the strongest MS activation was seen. Sinceno significant MS activation was found, these 20 voxelswere selected with the use of a nonsignificant activationthreshold. The time-course data were segmented into 24blocks of 90 s, with 12 blocks corresponding to periodswhen the visual stimulus was on, and 12 rest blocks. Thecentral 90 s was chosen out of the available 120 s to avoidtransition effects. A Fourier transform was performed oneach block, and the resulting spectra were averaged toreveal any differences in frequency components in therespective “active” and “rest” periods. For comparison, wealso performed an identical Fourier analysis on the MSregressors to identify the subsampled frequency.

RESULTS

Experiment 1: Visuomotor Response Task

All of the subjects showed BOLD activation in regions ofthe motor and visual cortices. In all subjects the MS re-gressors also accounted for some variance in the MR data(P � 0.01, uncorrected), but the activation was scatteredevenly throughout the brain with few clusters. For eachsubject there were less than 100 voxels activated for eachMS regressor at P � 0.01 uncorrected, and on average only10 voxels at P � 0.001 and one voxel at P � 0.0001. Giventhat the images contained approximately 10000 voxels, itappears likely that the “activated” voxels show an effect bychance. In line with this, the data did not survive any formof correction for multiple comparisons. The number of“activated” voxels was not significantly different for theimage sets collected at TEs of 60 ms and 90 ms.

Figure 3 shows the results of the subject with the bestexample of clustered MS activity, but, as can be seen, it

FIG. 2. Time course and regressors for ex-periment 2. The figure shows the timecourse of the flashing visual stimulus on thetop row and the constructed MS regressorson the second row. The subsampling of the6-Hz stimulus for a duration TE at each TRresults in a much slower frequency of oscil-lation of the MS regressors, as shown on thesecond row. The time course shown on thetop row is constructed using either a singlecosine wave or a single sine wave at theonset of each checkerboard flash, resultingin the two MS regressors shown on thesecond row. Note that the regressors will beidentical for each slice, since the stimulusand slice sampling rate are both 6 Hz.

Inability to Detect Neuronal Magnetic Fields 413

does not appear to be within the activated regions asdefined by the BOLD signal. Furthermore, the activatedpixels show a roughly even distribution between positiveand negative correlations, whereas any MS activity shoulddephase the MR signal, giving only negative correlations.

Experiment 2: ssVEPs

All three subjects showed strong BOLD activity in thevisual cortex. In contrast, in all three subjects the MSregressors accounted for no significant variance in thedata, even at the very low statistical threshold of P � 0.05uncorrected. Figure 4 shows the BOLD and MS responsesfrom one typical subject. The MS results show only a fewrandomly scattered activated voxels at the P � 0.05 level,which cannot be considered significant given the largenumber of comparisons that were performed.

Figure 5 shows the results of the frequency analysis. The“BOLD region” included clustered groups of voxels fromall six brain slices. The “MS region” consisted of scatteredvoxels within the occipital cortex on a single slice. Thespectrum of the cosine MS regressor (Fig. 5g) shows a

single component at 0.29 Hz, as predicted from samplingtheory. None of the ROIs in any of the three subjects showa response at this frequency; however, there is a responseat approximately 0.3–0.35 Hz in both the BOLD and MSROIs. This frequency component is present in the signalregardless of whether the visual stimulation is on (red line)or off (blue line).

DISCUSSION

Even though we followed a design very similar to that usedin the original study by Xiong et al. (9), we were unable toreproduce their results (Fig. 3). One explanation for thisdiscrepancy could be the difference in timings betweenthe two experiments. In the present study we used a meaninterstimulus time of 1 s, whereas Xiong et al. used alonger time of 2 s. Two seconds could be long enough toallow small modulations in the BOLD signal to occur, andcertainly long enough to see modulations in the initial dipof the BOLD signal (14). The model used by Xiong et al.did not include any possible BOLD effect, whereas we

FIG. 3. Regions of significant activation from ex-periment 1 in one individual subject showing themost clustered MS activation. Activation associ-ated with the MS regressors is shown at thresholdP � 0.01 uncorrected, and for the BOLD regressorat P � 0.00001 uncorrected. The figure shows MSdata with a TE of 90 ms and BOLD data with a TEof 60 ms. Column 5 shows clear BOLD activation inthe visual and motor cortices. There is scatteredMS activation, but it does not appear to be in thesame location as the BOLD activation.

FIG. 4. Regions of significant activation from experiment 2 in one typical subject. Activated voxels associated with (a) the BOLD regressorat threshold P � 0.00001 uncorrected, and (b) the MS regressors at P � 0.05 uncorrected.

414 Parkes et al.

included both MS regressors and BOLD regressors in ourGLM. Our use of a 1-s interval, however, effectively re-moves the possibility of detecting BOLD modulations, andcould therefore explain the apparent contradiction be-tween the two studies. Other minor differences between

the two studies, such as field strength (present study: 1.5T;original study: 1.9T) and voxel size (present study: 3.5 mmisotropic; original study: 3 � 3 � 6 mm), are unlikely tohave a major effect on the results. Chu et al. (3) offered afurther explanation for the discrepancies, i.e., that thedesign of the Xiong et al. (9) study, which involved mul-tiple time-staggered runs, is prone to artifacts related tovariations across the runs, such as fatigue and attention.We reduced this potential confound by instead jitteringthe stimulus presentation within a single run.

The results of experiment 2 also fail to show clusteredactivity corresponding to the MS regressors (Fig. 4b), de-spite a very strong BOLD response (Fig. 4a). The frequencyanalysis identified a frequency component in the MR sig-nal time course of 0.3–0.35 Hz (Fig. 5), which is close tothe predicted MS frequency of 0.29 Hz. However, thecomponent has equal amplitude during both visual stim-ulation and rest, and therefore cannot be due to neuronalmagnetic fields. The component is most likely due to on-going, spontaneous physiological fluctuations, such asmovement due to respiration, as was observed in a previ-ous work (15).

An alternative to our strategy of subsampling the 6-Hzneural signal would be to directly sample the activityusing a single-slice acquisition with a TR of �80 ms. Thiswould have the added benefit of directly sampling thephysiologic noise, which would allow these componentsto be easily identified. One drawback of a low TR is thatthe high degree of partial saturation makes the signal sen-sitive to spin history effects, and hence more sensitive tomovement.

Recent in vivo studies have shown some success indirectly detecting neuronal fields with MRI (1,2,4). Onestudy found increased power in the alpha frequency range(8–12 Hz) of the MR phase signal when the subjects closedtheir eyes (1), which suggests that alpha waves may bedirectly detectable. However, the magnetic field variationsproduced by alpha waves are an order of magnitude largerthan evoked fields, and indeed in the same study the signalfrom evoked fields could not be detected. Interesting workby Chow et al. (4) shows the possibility of detecting mag-netic field changes produced by action potentials along theoptic nerve. This is a new approach compared to previouswork, which focused on the direct detection of dendriticfields in the cortex. Given the lack of a sound consensus inthis emerging field of research, we believe it is importantfor all positive results to be replicated and all negativeresults to be reported.

CONCLUSIONS

We were unable to replicate the positive findings of Xionget al. (9) or to detect a MS signal from ssVEPs. Thus, thisstudy adds support to previous work (1,3) and demon-strates that the sensitivity of the magnitude MRI signal fordetecting evoked neuronal currents is too low to be ofpractical use for the particular stimuli and hardware usedin this work.

REFERENCES1. Konn D, Leach S, Gowland P, Bowtell R. Initial attempts at directly

detecting alpha wave activity in the brain using MRI. Magn ResonImaging 2004;22:1413–1427.

FIG. 5. Frequency spectra from ROIs in experiment 2. Spectra areshown for MR time-course data during periods of visual stimulation(red) and rest (blue). Two ROIs were considered: the 100 voxelsshowing the most significant BOLD activation (a, c, and e), and the20 voxels showing the most significant MS activation (b, d, and f;white boxes indicate the ROIs). g: The frequency spectrum of thecosine MS regressor shows a strong component at approximately0.29 Hz during activation (red), and no fluctuations during rest (blue).The strong frequency component at �0.35 Hz shown in b, d, and fis present during both activation (red) and rest (blue), and thus islikely due to ongoing, spontaneous physiological fluctuations, suchas respiration.

Inability to Detect Neuronal Magnetic Fields 415

2. Bianciardi M, Di Russo F, Aprile T, Maraviglia B, Hagberg GE. Combi-nation of BOLD-fMRI and VEP recordings for spin-echo MRI detectionof primary magnetic effects caused by neuronal currents. Magn ResonImaging 2004;22:1429–1440.

3. Chu RN, de Zwart JA, van Gelderen P, Fukunaga M, Kellman P, HolroydT, Duyn JH. Hunting for neuronal currents: absence of rapid MRI signalchanges during visual-evoked response. Neuroimage 2004;23:1059–1067.

4. Chow LS, Cook GC, Whitby E, Paley MNJ. Investigating direct detectionof axon firing in the adult human optic nerve using MRI. Neuroimage2006;30:385–346.

5. Konn D, Gowland P, Bowtell R. MRI detection of weak magnetic fieldsdue to an extended current dipole in a conducting sphere: a model fordirect detection of neuronal currents in the brain. Magn Reson Med2003;50:40–49.

6. Bodurka J, Jesmanowicz A, Hyde JS, Xu H, Estkowski L, Li SJ. Current-induced magnetic resonance phase imaging. J Magn Reson 1999;137:265–271.

7. Bodurka J, Bandettini PA. Toward direct mapping of neuronal activity:MRI detection of ultraweak, transient magnetic fields changes. MagnReson Med 2002;47:1052–1058.

8. Park TS, Lee SY, Park JH. Effect of nerve cell currents on MRI images insnail ganglia. Neuroreport 2004;15:2783–2786.

9. Xiong JH, Fox PT, Gao JH. Directly mapping magnetic field effects ofneuronal activity by magnetic resonance imaging. Hum Brain Mapp2003;20:41–49.

10. Kilner JM, Stephan KE, Josephs O, Friston KJ. Comparison of phase andmagnitude of the MR signal in measuring neuronal activity. In: Pro-ceedings of the 10th Annual Meeting of the Organization for HumanBrain Mapping, Organization for Human Brain Mapping, Budapest,Hungary, 2004. p 299.

11. Thesen S, Heid O, Mueller E, Schad LR. Prospective acquisition cor-rection for head motion with image-based tracking for real-time fMRI.Magn Reson Med 2000;44:457–463.

12. Herrmann CS. Human EEG responses to 1–100 Hz flicker: resonancephenomena in visual cortex and their potential correlation to cognitivephenomena. Exp Brain Res 2001;137:346–353.

13. Rager G, Singer W. The response of cat visual cortex to flicker stimuliof variable frequency. Eur J Neurosci 1998;10:1856–1877.

14. Yacoub E, Hu X. Detection of the early decrease in fMRI signal in themotor area. Magn Reson Med 2001;45:184–190.

15. Kiviniemi V, Ruohonen J, Tervonen O. Separation of physiological verylow frequency fluctuation from aliasing by switched sampling intervalfMRI scans. Magn Reson Imaging 2005;23:41–46.

416 Parkes et al.