NeuroImage 89 (2014) 143–151

Contents lists available at ScienceDirect

NeuroImage

j ourna l homepage: www.e lsev ie r .com/ locate /yn img

Multi-channel atomic magnetometer for magnetoencephalography:A configuration study

Kiwoong Kim a, Samo Begus b, Hui Xia c, Seung-Kyun Lee c, Vojko Jazbinsek d,Zvonko Trontelj d, Michael V. Romalis c,⁎a Korea Research Institute of Standards and Science, South Koreab Faculty of Electrical Engineering, Ljubljana, Sloveniac Physics Department, Princeton University, NJ, USAd Institute of Mathematics, Physics and Mechanics, Ljubljana, Slovenia

⁎ Corresponding author.E-mail address: romalis@princeton.edu (M.V. RomalisURL: http://physics.princeton.edu/atomic/romalis/ (M

1053-8119/$ – see front matter © 2013 Elsevier Inc. All rihttp://dx.doi.org/10.1016/j.neuroimage.2013.10.040

a b s t r a c t

a r t i c l e i n f o

Article history:Accepted 22 October 2013Available online 1 November 2013

Keywords:Atomic magnetometerAuditory evoked fieldBiomedical signal processingMagnetoencephalography

Atomic magnetometers are emerging as an alternative to SQUIDmagnetometers for detection of biological mag-netic fields. They have been used to measure both the magnetocardiography (MCG) andmagnetoencephalogra-phy (MEG) signals. One of the virtues of the atomic magnetometers is their ability to operate as a multi-channeldetectorwhile usingmany common elements. Herewe study two configurations of such amulti-channel atomicmagnetometer optimized for MEG detection. We describe measurements of auditory evoked fields (AEF) from ahuman brain as well as localization of dipolar phantoms and auditory evoked fields. A clear N100m peak in AEFwas observed with a signal-to-noise ratio of higher than 10 after averaging of 250 stimuli. Currently the intrinsicmagnetic noise level is 4 fTHz−1/2 at 10 Hz. We compare the performance of the two systems in regards to cur-rent source localization and discuss future development of atomic MEG systems.

© 2013 Elsevier Inc. All rights reserved.

Introduction

One of the most successful applications of the superconductingquantum interference device (SQUID) is in the field of biomagnetism.Especially, high sensitivity of a low-Tc SQUID magnetometer enabledthe measurement of neuromagnetic fields from a human brain andopened the field of magnetoencephalography (MEG) (Cohen, 1972).Since the SQUID MEG system was developed, numerous successfulinvestigations in various fields have been conducted with MEG. Atpresent, MEG is one of the most useful modalities for studies of brainfunctions together with the functional magnetic resonance imaging(fMRI) and positron emission tomography (PET). Specifically, due toits high temporal resolution, MEG has a unique position in the field ofthe cognitive science (Kwon et al., 2005; Tarkiainen et al., 2003). Al-though MEG is a powerful tool for functional brain imaging in bothbasic brain research (Darvas et al., 2003; Kakigi et al., 1995) and clinicaldiagnoses (Cheyne et al., 2007; Colon et al., 2009), existing SQUID sys-tems have a number of technical limitations that hinder more wide-spread use of this technique. Among them are a high rate of liquidhelium consumption, fixed sensor configuration, frequent requirementfor cryogenic maintenance, high cost and the need for large shieldedrooms.

)..V. Romalis).

ghts reserved.

Atomic magnetometer (AM) which optically detects polarizationchange of alkali metal vapor under external magnetic field is emergingas a promising alternative to existing MEG sensors. The sensitivity of aspin-exchange relaxation free (SERF) magnetometer is sufficient formeasuring an MEG signal (Kominis et al., 2003). In addition to theabsence of cryogenics, atomic magnetometers can operate in a multi-channel configuration at a relatively low cost by utilizingmany commondetector elements. They can also operatewith a smallermagnetic shieldsince there is no need to place a bulky liquid helium dewar inside theshielded room. However, the geometrical constraints and other techni-cal features of the AM system are substantially different from SQUIDsystems and require careful consideration for maximum utilizationof the available sensitivity and for being practically useful in routineMEG studies.

The aim of this research is to show the ability of potassium AM as adetector of weak evoked brain signals. In this study, we compare specif-ic features of two multi-channel AMMEG configurations and introduceAM specific data analysis methods such as localization of effective sen-sor positions by application of gradients, noise reduction technique formode hopping elimination (MHE), and response time correction. Thefirst system uses a transmitted probe beam, similar to the arrangementreported in Xia et al. (2006). In the second system we use a retro-reflected probe beam and several new technical features. We showmeasurements of auditory evoked fields (AEF) with both AM systemsand demonstrate the source localization with dipolar phantoms andAEF signals.

144 K. Kim et al. / NeuroImage 89 (2014) 143–151

In addition to our approach using a single alkali-metal cell for multi-channel recording, it is also possible to use an array of individual fiber-coupled atomic sensors. Fiber-coupled sensors with small alkali-metalcells have reached sensitivity below 10fT/Hz1/2 (Shah and Romalis,2009) and have been used for recording of magnetoencephalographysignals (Johnson et al., 2010, 2013; Sander et al., 2012).

Material and methods

Transmitted probe atomic MEG system

GeneralsThe detailed description of this setup was reported previously (Xia

et al., 2006). By absorbing circularly polarized pumping photons, potassi-um atoms in a glass cell are spin-polarized, creating local magnetization.The magnetization is rotated in the presence of an external magneticfield. The tipped component of magnetization rotates the polarizationof linearly polarized probe light. We optically measure the distributionof By field components in the y–z plane by using a 16 × 16-channel pho-todetector array, with geometry shown in Fig. 1(a).

In comparison to SQUID systems, where the sensor positions arefixed at design state, the AM system has more flexibility. In our casethe head of a subject is placed on top of the cell, as shown in Fig. 2.This feature is advantageous, for example, for baby MEG detection

Fig. 1.Multichannelmagneticfieldmappingwith an atomicmagnetometer; the expanded probchannel detector array. (b) Spatial distribution of pumping rate in the vapor cell, whichmakes asignal measured with the magnetometer.

(Okada et al., 2006). An inherent restriction in the optical signal detec-tion is the fact that we lose the position information along the probebeam direction; i.e. what we measure is the integral of all the magneticfields along the probe beam path. However, to localize a neuroelectricsource, we need to also get the spatial field distribution along the xaxis. This can be achieved by slicing the pumping beam in multiple sec-tions and illuminating only part of the cell at any given time (Gusarovet al., 2009). Such time division measurement is possible for repetitivemeasurements, such as evoked fields. In the detection of the spontane-ous brainwave or epileptic spikes, one can also usemagnetic fieldmod-ulation techniques to measure different components of the field (Li,2006; Seltzer and Romalis, 2004).

In our configuration, the sensor arrangement also provides a depthprofile of the magnetic field (Fig. 1(c)), which can help with magneticsource multipole analysis. The depth profile measures the decrease ofthemagnetic field as a function of distance from the source to a sensinglocation. The decay of the magnetic fields depends on the order ofthe current multipole, hence one can analyze the source current struc-ture, particularly for the case of multiple sources. For a SQUID sensorsystem, such layered configuration of pickup coils is undesirable dueto the superconductive screening currents on the lower coils. Evenwith adoption of an external feedback SQUID scheme, distortion of thefields in the gradiometric pick-up coil configuration is unavoidable.In contrast, high buffer gas pressure in the vapor cell of the atomic

e laser beamdetects the spatial distribution of By components on the y–z planewith a 256-different sensor characteristic for each sensor. (c)Map of themagneticfieldmap of theAEF

Fig. 2. Auditory evoked field measurement with the atomic magnetometer system. Thepotassiumcell and a human subject are placed in a three-layer cylindricalMu-metal shieldwith inner diameter of 0.9 m and outer diameter of 1 m. To block heat flow from the ovencontaining the cell, chilled water is circulating through a thin water bag between the headand the potassium cell's heating system. Tone stimuli are applied to one ear through anon-magnetic pneumatic earphone.

Fig. 3. Spatial probe beam profile corresponding to the mode hopping of the laser diode.

145K. Kim et al. / NeuroImage 89 (2014) 143–151

magnetometer reduces the cross-talk between the sensing channelsand we can measure nearly continuous spatial field distributions. Thecross-talk free distance lD is determined by the distance that Rb atomscan diffuse during their spin coherence time. This distance is propor-tional to the square root of the ratio of the diffusion constant D to thespin relaxation rate Γ:

lD ¼ffiffiffiffiDΓ

r: ð1Þ

For our conditions the diffusion constant of Rb atoms in 1 atm. of Hebuffer gas at 180 °C is 0.45 cm2/s. The total spin relaxation rate of Rbatoms Γ is equal to 2π × 30 Hz, which is the bandwidth of themagneticfield response. The diffusion distance is then equal to 0.5 mm, muchsmaller than the actual separation of the magnetometer channels.Therefore, each channel can be considered to be independent.

Multichannel characteristicsIn optical polarimetry, the signal is a product of light intensity and

the rotation of light polarization. As each channel is illuminated withslightly different probe light intensity, we need to calibrate each chan-nel with a known magnetic field. Moreover, due to the finite opticaldepth, each sensor has a different pumping rate along the pumpingbeam direction (Fig. 1(b)). The sensitivity to By depends on thepumping rate, and the magnetometer bandwidth also depends on theoptical pumping rate as well as the polarization of atoms (Kominiset al., 2003). Therefore, spatially inhomogeneous pumping profile pro-duces different characteristics of each channel both in sensitivity andin detection bandwidth. If the frequency range of a signal of interest isoutside of the sensor bandwidth, one can get a significant phase delayof the signal component. Event related potential usually forms a peak;the delay will change the latency of the peak appearance, necessitatingappropriate correction, such as a digital high-pass filter to compensate

for low-pass response of the magnetometer. We note that even thoughthe response of atomic magnetometers drops at higher frequencies dueto limited bandwidth, the noise also often drops, so the sensitivity of themagnetometer can be flat over a frequency range substantially largerthan its bandwidth (Shah et al., 2010).

Mode hopping eliminationMode hopping in the semiconductor laser generates big stepwise

jumps in the signal baseline. Even averaging hundreds of the responseepochs is not enough to suppress this noise. By doing singular value de-composition on the covariance matrix of measured fields without stim-ulus, we can separate the large variance noise eigen-components andtheir corresponding spatialfield patterns (Fig. 3). Elimination of the pro-jection components of these spatial patterns effectively reduces notonly the mode hopping noise, but also other ambient magnetic fieldnoises and possibly spontaneous brain magnetic fields.

Retro-reflected probe atomic MEG system

As discussed in the previous sections, the source localization capa-bility is one of the most important design factors of an atomic magne-tometer system for the biomagnetism applications. The problemsmentioned above were dealt with in our new system design (Fig. 4).

First, we adopted a retro-reflected probe beam for circumventingthe blind direction. At the mirror placed on the measurement surface,the probe beam reflects back to the detector after passing through thepotassium cell twice. The rotation angle of the probe beam polarizationdue to paramagnetic Faraday rotation is added after two passes. Thereflected beam is recorded by 16 × 16 channel photodiode array andwe canmeasure the spatial distributions of By and Bz components by al-ternating the pumping beam direction with the beam switch as shownin Fig. 4(b). The By and Bz components are corresponding to the orthog-onal tangential field components of the MEG. Compared to the radialcomponent (Bx) measurement, the tangential components measure-ment provides deeper and wider localizable space for current dipolesources when we have the same detection coverage areas (Kim et al.,2004). In the tangential component measurement, the field maximumpatternwill be located right over the current dipole sourcewhile the ra-dial component measurement shows a diverse dipolar pattern. Besidesthe sensor coverage area, the By and Bz components guarantee the infor-mation orthogonality even for spatially correlated fields at nearby sen-sors; this configuration carries more information on the sources.Second, we use two separate pumping lasers which are detuned in theopposite direction from the D1 line in K. This reduces their absorptioncross-section and allows the light to propagate further into the cell toobtain a more uniform spatial pumping profile. The detuning of the

Fig. 4. Schematic diagram for the retro-reflected probe atomic MEG system. (a) The widely-expanded probe beam reflects back to the photodiode array. In the wide K vapor cell(12 × 12 × 4 cm), the beam path is doubled. It measures the tangential field distribution from the brain. (b) By alternately switching the beam switch, we can change the direction ofthe pumping beam tomeasure two orthogonal tangential field components, By and Bz. The balanced detuning of the pump beam compensates the light shift of each beam and the detunedbeam provides more uniform pumping profile.

146 K. Kim et al. / NeuroImage 89 (2014) 143–151

two lasers from resonance is equal and opposite in order to cancel thevirtualmagneticfield, so-called light shift, created by laser light detunedfrom the optical resonance in K atoms. The light is generated by twoDFBlasers which are combined and amplified by a 1-W tapered amplifier tosupply high enough power for pumping the 12-cm deep vapor cell(Souza, 2008). Third, the heating system has been changed from hotair flow to AC electric heating. A pair of non-magnetic heating wireswas twisted to prevent from generating magnetic fields in the cell.The heating power is generated by an audio amplifier at 20 kHz and isfar away from the detection bandwidth. The operation temperaturewas about 200 °C to get high enough potassium density.

Results

Transmitted probe system

N100m peak is a typical primary brain response that appears100 ms after hearing a single tone stimulus. Human subject recordingsfollowed a protocol approved by Princeton University Institutional Re-view Board. Since themain focus of this research is on the experimentalapparatus for acquiring MEG data, we used just a few subjects as repre-sentative examples of typical MEG signals. We applied 500-Hz tonestimuli to the subject's right ear using a pneumatic earphone and

measured the response of the contra-lateral cortex. A clear N100mpeak in AEF was observed with a signal-to-noise ratio of higherthan 10 after averaging of 250 stimuli (Fig. 5(a)). For comparison,Fig. 5(b) shows 100-times-averaged AEF measured by a hemispherical37-channel double relaxation oscillation SQUID (DROS) magnetometersystem having a 3 fTHz−1/2 white noise level (Lee et al., 2003). In termsof sensitivity, the twomodalities are comparable to each other. The sin-gle polarity peaks in the AM recording imply that the current AM sys-tem does not have enough detection area (~3 cm in a length) to coverthe dipolar pattern of the magnetic fields, while the hemisphericalSQUID system covers an area of 14 cm in diameter.

As a rule of thumb, tominimize the localization error of a current di-pole source, we require at least a

ffiffiffi2

ptimes wider measuring diameter

than the distance between the sensor array and the source when thesource is placed at the center of the array. Oncewedefine themeasuringdiameter as L,

ffiffiffi2

ptimes the depth of the target source current dipole, all

the lengths can be normalized by L. We calculated the required signal tonoise ratio for the source localization as functions of the coveragediameter and the deviation of the center of the detection area fromthe position of a target current dipole source, respectively (Fig. 6). Thecalculation result shows the minimum required SNR with which 95%of current dipoles could be localized in 1 cm3 error volume. As shownin Fig. 6, for a MEG source on the brain cortex, the distance between

Fig. 5. (a) Auditory evoked field traces for all atomicmagnetometer channels. The typical N100m peak appears 100 ms after the sound stimulus. (b) AEF traces measured by a 37-channelSQUID MEG system.

147K. Kim et al. / NeuroImage 89 (2014) 143–151

the sensor array and the source is about 60 mm, and L ~ 85 mm. The re-quired diameter of the detection area, 1.4L, is about 12 cm. Since theexact source location is not known, the cell size should be even largerto ensure optimal localization.

Another issue with the AEF localization using an AM is the directionof the source dipole. In our configuration, the AEF current dipole hap-pened to be oriented nearly perpendicular to the probe beam direction.As mentioned in the previous section, the probe beam integrates all thefield distribution along the beam path. If the current dipole was locatedin the middle of the sensing area, the positive and the negative fieldsfrom the dipole would cancel out each other and the signal amplitudewould be reduced. This problem can be solvedwith slicing the pumpingbeam. The optimal number of sliceswas found byusing a computer sim-ulation minimizing the source localization error. Interestingly, slicingthe pump beam in two gave the lowest localization error (Fig. 7). Inour configuration, the higher number of slices gives the lower signal in-tensity, which results in a larger localization error. We can realizethe pumping slice selection by using alternative half-blocking of thepumping beam. To investigate this issue further, we placed a current di-pole around the middle of the sensing area by using a spherical headconductor phantom consisting of a current dipole in saline solution.We applied currents simulating the AEF and alternatively measuredeach pumping area; we alternatively blocked each half of the pumpingbeam cross-section. Fig. 8 shows the measurement results. Figs. 8(a)and (c) are the measured waveforms of all the channels while the firsthalf of the pump beam is blocked and the spatial field distribution atthe instant of the maximum peak, respectively. Figs. 8(b) and (d) are

Fig. 6. Required signal-to-noise ratio for the source localization as functions of source po-sition deviation from the center of the cell anddetection area, respectively. The L is definedto be a measuring diameter which is

ffiffiffi2

ptimes the depth of the target source current di-

pole. The detection area over 1.4L shows a flat SNR profile less depending on the positionof the source.

the measured waveforms of all the channels while the other half ofthe pump beam is blocked and the spatial field distribution at the in-stant of the maximum peak, respectively. With these two sets of thespatial magnetic field distribution, we could make source dipole locali-zation by solving an iterative nonlinear optimization problemwith var-iables for three orthogonal components of the position and the twocomponents of the current dipole strength; the spherical conductormodel eliminates the other source component of the dipole (Sarvas,1987). The estimated position of the current dipole was deviated fromthe original position by Δx ~ 15 mm, Δy ~ 1 mm, Δz ~ 1 mm. Ofcourse, the error in the direction for the probe beam was largest.

Retro-reflected probe atomic MEG system

Prior to measurement of the audio evoked brain signals with AM acompensation of the residualmagneticfields in the volumeof the potas-sium cell is necessary. We used the compensation coils inside the innermagnetic shield. This way the SERF operation of the AM was assured.The AM channels sensitivity was calibrated by applying low frequencymagnetic field (50 pT, 10 Hz) across the AM cell. Noise spectrum ofone magnetometer's channel and of the two adjacent channels in thegradiometer configuration is shown in Fig. 9. The intrinsic magnetome-ter noise determined by dividing the noise in the difference betweentwo adjacent channels by a factor of

ffiffiffi2

pis 4 fTHz−1/2 above 10 Hz. At

higher frequencies, vibration and magnetic interference peaks are sig-nificant, but they are partly canceled in the gradiometer measurements.

By applying low frequency constant magnetic field gradient(10 pT/cm, 10 Hz) using the gradient compensation coils the spatialmapping information from the AM cell to each photodiode positionwas obtained. The two-dimensional field profile data were smoothed

1 2 4 8 160

2

4

6

8

10

Loc

aliz

atio

n er

ror

volu

me,

cm

3

Number of slices (pumping selections)

100 nAm dipole || z−axis100 nAm dipole || y−axis

Fig. 7. Simulated source localization error as a function of the number of the pumpingslices.

Fig. 8. Results of the sliced pumping experiment for the current dipole localization with a spherical brain phantom. One half of the pump beam is blocked alternatively. (a) Themeasuredwaveforms of all the channels while the first half of the pump beam is blocked. (b) Themeasured waveforms of all the channels while the other half of the pump beam is blocked (c) and(d) are the spatial field distributions at the instant of the maximum peaks of (a) and (b), respectively.

148 K. Kim et al. / NeuroImage 89 (2014) 143–151

with an average value of 5 by 5 elements (Reeves, 2009) before the spa-tial information has been calculated (Fig. 10).

Effective calculated positions of the magnetometer channels usingthe magnetic field gradient mapping are shown in Fig. 11. Channelswith good signal to noise ratios, used in data analysis are shown withlarge marks.

10 20 30 40 5010

0

101

102

103

104

B, f

TH

z−1/

2

f, Hz

Ch133Ch133 − Ch188

Fig. 9. Noise spectrum of single channel magnetometer and gradiometer configuration ofthe two adjacent channels.

Each auditory applied stimulus consists of 20 square pulses witha period of 1 ms and duty cycle of 50%. The interval between eachpulse train is varied randomly between 1.3 s and 2.0 s to avoid subject'sadaptation.

Prior to data analysis the signalswerefiltered by a band-pass 2–20 Hzfilter. After rejecting the subject's heartbeat signals, signals originatingfrom eye movements and disturbances due to mechanical vibrations,the N100m could be seen in several channels of the 256 channel AMafter averaging 710 stimulus epochs. This was achieved by combiningBy magnetometer channels into a gradiometer configuration: one mag-netometerwas selected as a reference channel andwas subtracted fromthe other channels. Fig. 12 shows gradiometric channels with the bestsignal to noise ratio.

Analysis of MEG data followed common biomagnetism techniques.We have chosen a current dipole source model in a conducting sphere(Sarvas, 1987). Using signals at 100 ms after the stimulus from 10 se-lected By gradiometric channels with good signal to noise ratio wefound the position (rp), the dipole moment p = (px2 + py

2)1/2 and theangle ϕ = arctan(px/py) for the best-fitting equivalent current dipole.The third dipole component pz was neglected in the analysis becausein our setup it was almost radial and therefore did not contribute signif-icantly to the magnetic field outside of the conducting sphere (Sarvas,1987). Relative error (RE) defined as a root mean square (RMS) differ-ence between measured and calculated data divided by the RMS ofmeasured data was 0.1 and corresponding correlation coefficient (CC)0.99. The obtained location rp = (−6.36, 2.15, 0.34) cm relative to thesphere center is in the expected region of the cortex (Pantev et al.,

0 5 10 150

2

4

6

8

10

12

14

16

−45

−40

−40

−35

−35

−35

−30

−3

0

−30−30

−25

−25

−25

−20

−20

−20

−15

−15

−15

−10

−10

−10

−5

−5

−5

0

0

0

5

5

5 10

10

10

15

15

Column nr.

Row

nr.

−40

−30

−20

−10

0

10

B, p

T

(a)

0

2

4

6

8

10

12

14

16

−20

−15

−15

−10

−10 −10

−5

−5 −5

0

00 0

5

5

5 5 5

10

10 10 10

15

15

15

15 15 15

20

2020

20

25

2525

25

30

303035

35

35

40

40

Row

nr.

−20

−10

0

10

20

30

40

B, p

T

(b)

0 5 10 15

Column nr.

Fig. 10. Smoothed measured magnetic field at different photodiode locations with applied constant magnetic field gradient over the potassium cell. Left: gradient in y-direction. Right:gradient in z-direction.

149K. Kim et al. / NeuroImage 89 (2014) 143–151

1987), distance from the center (|rp| = 6.7 cm), the dipole momentp = 74.1 nAm and the angle ϕ = 34.2°. Fig. 13 shows magnetic fieldmaps for the tangential and radial components calculated from theequivalent current dipole.

The alternative way to solve the biomagnetic inverse problem is theminimumnormestimation (MNE)method (Parker, 1977; Sarvas, 1987)where only the source space is constrained. Among all possible currentdistributions in that space, the onewith theminimum norm is selected.Fig. 14 shows reconstructed current distribution constrained to a spher-ical cap surface inside a conducting sphere. We have got the same RE(0.1) and CC (0.99) as in the case of the equivalent current dipolesource. Fig. 15 shows tangential and radial components of magneticfield calculated from the obtained MNE current distribution in Fig. 14.Again, these magnetic field maps are very similar to those obtained bya single equivalent dipole current source.

These results were confirmed by the following test: We dividedmagnetic field maps in Fig. 15 in regular 11 times 11 square grid ofpoints. From field values in grid points we found the best fitting dipolefor both field components. Results displayed in Fig. 16 show that we ob-tained similar equivalent current dipoles in both cases. Like in the caseof single equivalent dipole fit in Fig. 13, they are approximately 6.7 cmfrom the sphere center (2.3 cm from the sphere surface) and they arepositioned in the same region within ~1 cm and oriented in a similardirection.

−5 −4 −3 −2 −1 0 1 2−2

−1

0

1

2

3

4

Y, cm

Z, c

m

Fig. 11. Spatial correction of the distorted image on the photo detector array based on gra-dient imaging. Channels with good signal to noise ratios which are used in data analysisare shown with large marks. The non-uniformity of the effective location of the channelsis due to optical distortions near the edge of the cell.

Discussion

In this paper we described two multichannel systems for detectingMEG signals with atomic magnetometry. A unique aspect of these sys-tems is that they use many common elements, so a multichannel mag-netic fieldmapping can be realizedwith a relatively low complexity andcost. The geometry of atomic magnetometers also offers new possibili-ties. In one of the geometries we measured the radial magnetic field asa function of distance away from the source, which allows one to deter-mine the magnetic multipole order or disentangle multiple dipolesources. In another geometry we measured the two tangential compo-nents of themagnetic field using the same sensor. Unlike SQUIDs, atom-ic magnetometers do not suffer from cross-talk and can be used to mapvector magnetic fields in 3-D with millimeter resolution.

Atomic magnetometers also present unique challenges for MEG de-tection. Since the effective sensor positions are defined by laser beams,they need to be determined in-situ. We apply known magnetic fieldgradients to localize each of the channels based on detected magneticfields. The sensitivity and bandwidth of each channel is determinedfor each data recording session by applying calibrated magnetic fieldsat varying frequencies. The sensitivity of the sensors can reach 4 fT/Hz1/2 above 10 Hz and magnetic field noise cancelation has been dem-onstrated between multiple channels.

We have successfully detected evoked auditory magnetic fields anddemonstrated source localization using the homogeneous conducting

−0.2 0 0.2 0.4 0.6−120

−100

−80

−60

−40

−20

0

20

40

60

80

t, s

ΔB, f

T

Fig. 12. Auditory evoked magnetic gradients in By obtained with AM after averaging. Theauditory stimulus was at t = 0 s.

(a) (b)

Fig. 13.Magnetic field maps for the tangential, By (left panel), and radial, –Bx (right panel), components calculated from the best fitting equivalent current dipole. The current dipole hasbeen obtained by fitting measured signals at time t = 0.102 s after the stimuli from ten selected By gradiometer channels, defined as difference between the selected channels ◊ and thereference (x).Δ is step betweenmap lines, m andM denotemaximum andminimum field values in fT, red, green and blue lines represent positive, zero and negative field values, respec-tively. Dotted line denotes the head model with radius 9 cm, as well as nasal and neck regions.

(a) (b)

Fig. 14. The estimated MNE current distribution in two different views. The head model is a sphere with 9 cm radius, shown with orange open triangles. The reconstruction area of thecurrent distribution is shown with black triangles shaded with grey (spherical cap surface with height hc = 5 cm and radius rc = 8 cm). The estimated current distribution is shownwith cyan arrows.

150 K. Kim et al. / NeuroImage 89 (2014) 143–151

sphere model. The localization of the equivalent current dipole corre-spondswellwith the knownposition of the auditory evoked brain activ-ity. While such model is relatively simplistic, it demonstrates thatatomic magnetometers have sufficient signal-to-noise ratio and spatialmapping capabilities to be used for MEG studies. Our retro-reflected

(a) (

Fig. 15.Magnetic field maps calculated from the estimated current dist

probe beam scheme is expected to be very useful not only for theMEG measurement but also for all other biomagnetic applications likeMCG since it can measure the tangential component of the magneticfield for which the measured field is the maximum just above thesource, this is advantageous in source localization.

b)

ribution shown in Fig. 14 for By (left panel) and –Bx (right panel).

(a) (b)

Fig. 16. Calculatedmagnetic field maps for the current dipoles obtained by fitting magnetic field maps shown in Fig. 15. By (left panel): RE = 0.21, CC = 0.982, rp = (−6.5,1.55,−1.17)cm, |rp| = 6.8 cm, p = 55.2 nAm and ϕ = 25°, and –Bx (right panel): RE = 0.2, CC = 0.98, rp = (−6.25,2.19,−0.79) cm, |rp| = 6.7 cm, p = 54.3 nAm and ϕ = 29.7°.

151K. Kim et al. / NeuroImage 89 (2014) 143–151

Conclusion

The successful measurement and localization of human brain activi-tywith amultichannel atomicmagnetometer systemdemonstrates thatthis non-cryogenic magnetometer is a possible alternative to SQUIDsensors. The current system is still relatively immature compared withcommercial SQUID systems but represents the first step toward the de-velopment of an economical and flexible multi-channel atomic MEGsystem. We discussed specific procedures and technical challengesthat can guide future development of such systems.

Acknowledgment

The authors would like to acknowledge the financial support of NIHand Packard Foundation (US), Korea Research Foundation and KRISS(Korea), and Slovene Research Agency and MORS (Slovenia).

References

Cheyne, D., Bostan, A.C., Gaetz,W., Pang, E.W., 2007. Event-related beamforming: a robustmethod for presurgical functional mapping using MEG. Clin. Neurophysiol. 118 (8),1691–1704.

Cohen, D., 1972. Magnetoencephalography— detection of brains electrical activity with asuperconducting magnetometer. Science 175, 664–666.

Colon, A.J., Ossenblok, P., Nieuwenhuis, L., Stam, K.J., Boon, P., 2009. Üse of routine MEG inthe primary diagnostic process of epilepsy. J. Clin. Neurophysiol. 26 (5), 326–332.

Darvas, F., Pantazis, D., Kucukaltun-Yildirim, E., Leahy, R.M., 2003. Mapping human brainfunction with MEG and EEG: methods and validation. NeuroImage 23 (Suppl. 1:S),289–299.

Gusarov, A., Levron, D., Paperno, E., Shuker, R., Baranga, A.B., 2009. IEEE Trans. Magn. 45,4478–4481.

Johnson, C., Schwindt, P.D.D., Weisend, M., 2010. Magnetoencephalography with atwo-color pump-probe, fiber-coupled atomic magnetometer. Appl. Phys. Lett.97, 243703.

Johnson, C.N., Schwindt, P.D.D., Weisend, M., 2013. Multi-sensor magnetoencephalogra-phy with atomic magnetometers. Phys. Med. Biol. 58, 6065–6077.

Kakigi, R., Koyama, S., Hoshiyama, M., Kitamura, Y., Shimojo, M., Watanabe, S., 1995. Pain-related magnetic fields following painful CO2 laser stimulation inman. Neurosci. Lett.192 (1), 45–48.

Kim,K., Lee, Y.H., Kwon,H., Kim, J.M., Kim, I.S., Park, Y.K., 2004. Optimal sensor distribution formeasuring the tangential field compoments inMCG. Neurol. Clin. Neurophysiol. 60, 1–4.

Kominis, I.K., Kornack, T.W., Allred, J.C., Romalis, M.V., 2003. A subfemtotesla multichan-nel atomic magnetometer. Nature 422, 596–599.

Kwon, H., Kuriki, S., Kim, J.M., Lee, Y.H., Kim, K., Nam, K., 2005. MEG study on neural ac-tivities associatedwith syntactic and semantic violations in spoken Korean sentences.Neurosci. Res. 51 (4), 349–357.

Lee, Y.H., Kwon, H., Kim, J.M., Kang, C.S., Kim, K., Kim, I.S., Park, Y.K., Lee, S.G., 2003. A mul-tichannel SQUID magnetometer system based on double relaxation oscillationSQUIDs. IEEE Trans. Appl. Supercond. 13 (2), 755–758.

Li, Z., 2006. Development of a Parametrically Modulated Serf Magnetometer. Thesis(Ph. D.) University of Wisconsin-Madison.

Okada, Y., Pratt, K., Atwood, C., Mascarenas, A., Reineman, R., Nurminen, J., Paulson, D.,2006. BabySQUID: a mobile, high-resolution multichannel magnetoencephalographysystem for neonatal brain. Rev. Sci. Instrum. 77, 024301.

Pantev, C., Hoke, M., Lütkenhöner, B., Lehnertz, K., 1987. Influence of stimulus intensity onthe location of the equivalent current dipole in the human auditory cortex. In:Atsumi, K., Kotani, M., Ueno, S., Katila, T., Williamson, S.J. (Eds.), Biomagnetism, 87.Tokyo Denki University Press, pp. 146–149.

Parker, R.L., 1977. Understanding inverse theory. Ann. Rev. Earth Planet. Sci. 5, 35–64.Reeves, G., 2009. Smoothing Function for 2D Data ßmooth2a. Division of Biology Caltech

(March, http://www.mathworks.com/matlabcentral/fileexchange/23287-smooth2a).Sander, T.H., Preusser, J., Mhaskar, R., Kitching, J., Trahms, L., Knappe, S., 2012. Magnetoen-

cephalography with a chip-scale atomic magnetometer. Biomed. Opt. Express 3,981–990.

Sarvas, J., 1987. Basic mathematical and electromagnetic concepts of the biomagnetic in-verse problem. Phys. Med. Biol. 32, 11–22.

Seltzer, S.J., Romalis, M.V., 2004. Ünshielded three-axis vector operation of a spin-exchange-relaxation-free atomic magnetometer. Appl. Phys. Lett. 85, 4804.

Shah, V., Romalis, M.V., 2009. Spin-exchange relaxation-free magnetometry using ellipti-cally polarized light. Phys. Rev. A 80, 013416.

Shah, V., Vasilakis, G., Romalis, M.V., 2010. High bandwidth atomic magnetometery withcontinuous quantum nondemolition measurements. Phys. Rev. Lett. 104, 013601.

Souza, M.J., 2008. A method for making large rectangular optical cells for SERF magne-tometers. Fusion 55, 17–21.

Tarkiainen, A., Helenius, P., Salmelin, R., 2003. Category-specific occipitotemporal activa-tion during face perception in dyslexic individuals: an MEG study. NeuroImage 19(3), 1194–1204.

Xia, H., Baranga, B.A., Hoffman, D., Romalis, M.V., 2006. Magnetoencephalographywith anatomic magnetometer. Appl. Phys. Lett. 89, 211104.