www.centauro.it Rivista di Neuroradiologia 16: 995-1000, 2003

995

64 EPI volumes (3 x 3 x 6 mm,DF 0.20, TE 45 / TR 108 ms, FA40°) were acquired at constantinterscan intervals RT of 0.324seconds avoiding delays betweenconsecutive slices (3 interleavedcoronals; Gy H � F) and vol-umes. Thereupon, scanning washeld for 30 seconds to let pre-sumed haemodynamic BOLDresponses disappear in silence.Repetition of 30-50 runs sepa-rated by these silent intervalswas to ensure an in-between fad-ing of BOLD effects evoked bypGGN audition (pGGNA). Dis-carding the initial 12 volumes toapproximate to a steady state,the data were explored model-free by a principal componentanalysis (PCA) using pca_imageconfined to the remaining 52time points in each run 5. Howe-ver, the first component still pos-sessed extreme temporal valuesfor the initial volumes and it wasincluded as a confound into fur-ther model-driven analysis. Thus,statistical parametric mapping(SPM99) was performed to spe-cifically demonstrate the tem-poral delay of haemodynamicBOLD activity evoked by pG-GNA in an event-related designof finite impulse responses (FIR).In the FIR model, the 52 scanskept of each run were structured

Introduction

Echo Planar Imaging (EPI) ge-nerates considerable acoustic noi-se by steep gradient pulses at un-restricted bandwidths. In func-tional neuroimaging, unshieldedEPI sounds inevitably activate theauditory system impairing the sen-sitivity to detect concomitant andsuccessive auditory stimulations.Instead of vanishing the scanner’snoise transmission 1, our goal wasto utilize it for the functional neu-roimaging of audition. In contrastto previous approaches 2-4, whichimplemented additional gradientswitches for sound generation, wefocussed on EPI’s genuine gradientnoise (GGN) of clinical 1.5 T MRscanners to detect auditorily e-voked BOLD effects. By taking u-nique advantage of the delayedrise of haemodynamic responsefunctions, EPI finally ought to de-tect itself within T2*-weighted sig-nals related to onsets of persistent(A) or intermittently accelerated(B) but otherwise unmodifiedGGN of the scanning per se.

Material and MethodsMethod A (figure 1)

Auditory autodetection of EPI’spersistent genuine gradient noise(pGGN):

into 26 mini-epochs contrastingthe final 20 (1/20 x 20 = +1) a-gainst the initial 6 (-1/6 x 6 = -1;t-test, see below).

Method B (figure 2)

Auditory autodetection ofEPI’s intermittent genuine gradi-ent noise (IGGN):

Delays between each of a lim-ited number EPI-repetitions we-re specified separately: 17 repeti-tions of axially (parallel to theSylvian fissure) acquired EPIvolumes (5 slices, 3 x 3 x 4 mm,FA 70°) were recorded at an RTof 2.000 seconds, except that the5th and 6th scan were repeated af-ter 1.000 seconds already. Con-sidering its GGN setting in inter-mittently, the 5th scanning wasdiscarded from the statisticalanalyses of the time series butdefined as the auditory stimula-tion of interest. Repeating thisprotocol at least 12-15 times,data were analysed by StatisticalParametric Mapping (SPM99) inan event-related design. The de-sign matrix contained the HRFand its temporal derivative as ba-sis functions as well as the ave-raged signal intensity of each ima-ge to account for fluctuations ofthe global signal due to the addi-tional scan.

Detection of the Scanner’s Genuine GradientNoise by Functional Echo Planar Imaging

A.J. BARTSCH, K. SPECHT*Division of Neuroradiology, Bavarian Julius Maximilians-University of Würzburg, Würzburg* fMRI-Section, Division of Neuroradiology, Medical Center Bonn, Bonn; Germany

Key words: MRI, EPI, GN, noise

XVII Symposium Neuroradiologicum 2002: Detection of the Scanner’s Genuine Gradient Noise ... A.J. Bartsch

996

Figure 1 Regressor (left) and design matrix (right) used in the SPM analysis of method A.

Figure 2 Design matrix (left) and extracted signal (right), which is used as additional covariate in the design matrix of method B.

www.centauro.it Rivista di Neuroradiologia 16: 995-1000, 2003

997

Method A & B Equipment andPulse Sequences

Measurements were perfor-med on clinical 1.5 T SiemensMR scanners (A: Magnetom Vi-sion / B: Symphony), running un-der Numaris 4 and 3, respec-tively. EPI’s genuine gradientnoise (GGN) amounted up to96dB at the level of the auriclesunshielded but external foamand muffle pads as well asearplugs were used for fixationpurposes and to minimize thesubjects motion predominantlydue to initial startling. FollowingTRUFI-localizers, the time seriesof BOLD-sensitive Mosaic EPIswere acquired. Two methods (A,left, and B, right boxes) were es-tablished for auditory EPI au-todetection after extensive expe-rimental pretestings. For anato-mical inference, a T1-weightedMP-RAGE sequence (≤ 1 x 1 x 1mm_, TE 4 / TR 12 ms, FA 12°)of the entire brain was recorded.

Subjects and DataPreprocessing

Method A

Five right handed, healthy malevolunteers, age 33.2 ± 4.3 years.

Method B

Seven right handed, healthyvolunteers (five male, two fe-male), age 30.3 ± 7.7 years.

Method A & B

The functional EPI volumeswere spatially realigned by themost accurate rigid body trans-formations (SPM99) and reslicedthrough a modified Fourier spa-ce interpolation. Motion correc-tion parameters were included asconfounds into parametric ana-lyses (not shown in the figures).Slice timing was renounced toavoid band-limitation, aliasing,and loss of experimental power

rise. The signal change amountedto about 2.5% above the globalmean of the scaled BOLD-sensi-tive images. Spatially, this effectwas traced by small volume cor-rections for multiple compari-sons performed on anatomicallypredefined spheres of 10 mm ra-dius to clustered activations cov-ering the primary auditory cor-tex around the gyri temporalestransversi of Heschl and to bra-instem centres of audition (nu-clei cochleares et corporis trape-zoidei, oliva superior, colliculi in-feriores et brachium, corpus ge-niculatum mediale) by the near-est cluster approach (FDR-cor-rected p < 0.01). These resultswere visualized by AMIRA ren-dering the activated clusters ontoa likelihood voltex of grey matterand an isosurface representationof white matter.

Results for Method B (figure 4)

The method was sensitive e-nough to detect BOLD responsesin the auditory cortex as evokedby scans of the pulse sequenceitself intermittently inserted atshortened RTs. Modelling thetemporally delayed BOLD effectevoked by GGN, it revealed theincrease of the T2*-weighted sig-nal superimposed by a singular s-timulus scan. The inclusion of theoverall mean intensities capturesmost of the confounding signalfluctuations. However, the criti-cal point here is that the signalhas to be stable at least twoscans after the “stimulationscan”. Otherwise, the signal fluc-tuation could be on the sametime scale as the expected tem-porally delayed BOLD response.We observed further that thesefluctuations vary profoundly ondifferent scanners (Vision, Sym-phony, Sonata). Therefore, ap-propriate adjustments of the flipangle used are mandatory. Re-moving this global effect andtesting for the main effect, i.e. for

at Nyquist’s frequency of RT.Considering an interstimulus in-terval of interest at or below 2 xRT this was particularly war-ranted by method A. Smoothingwas performed by a Gaussian k-ernel of twice the voxel size. Fi-nally, the structural 3D MP-RA-GE image was coregistered tothe mean functional by a mutualinformation algorithm.

Results

Results for Method A (figure 3)

In all subjects, the first compo-nent extracted by PCA regularlydetected a decay of the T2*-weighted signal finally approach-ing its global steady state. Due tothe longer T1-values of the cere-brospinal fluid, it centred spa-tially on the inner and outerliquor-filled spaces. Temporally,reaching a steady state can befastened by increasing the flipangle (FA) closer to 90°. At la-tencies of a few seconds, the sec-ond or third component caughtby PCA commonly exhibited asubsequent rise of the signal am-plitude attributable to a BOLDhaemodynamic response func-tion (HRF) evoked by pGGNA.In two of the fife subjects mea-sured, however, there wereprominent initial startling move-ments extending over the dis-carded volumes in most sessionsand preventing a meaningfulanalysis due to the correlation ofconfounding motions with theeffect searched for. In the otherthree subjects, the componentpresumably related to pGGNAfocussed bilaterally on the sub-cortical and cortical centres ofaudition, peaked at around 10seconds, and then usually attenu-ated to some extend over time.In SPM, the corrected responsefit of the filtered and adjusteddata for t-contrasting the initial 6against the remaining 20 FIRsalso detected a delayed signal

XVII Symposium Neuroradiologicum 2002: Detection of the Scanner’s Genuine Gradient Noise ... A.J. Bartsch

998

the basis function of the HRF,we found significant activationswithin the primary auditory sys-tem. The activated areas were re-stricted to the gyri temporalestransversi (Heschl’s gyri). Re-

sure, no brainstem centres of au-dition were within the field ofview. However, in addition to theresults above, we were able todemonstrate a lateralized pro-cessing of iGGN. In all cases, the

gions known to be involved inhigher order processing of an au-ditory signal were not detected.Due to the small number ofslices (n = 5) and their acquisi-tion parallel to the Sylvian fis-

Figure 3 Method A. Left: PCA analysis; right: Fitted response and rendering of the SPM results.

Figure 4 Method B. Overlay of the SPM results onto anatomical sections (left) and the fitted response (right).

www.centauro.it Rivista di Neuroradiologia 16: 995-1000, 2003

999

most (or only) significant BOLDsignal was detected within theright auditory cortex. The figurebelow depicts in the left panelthe result of a single subjectSPM analysis. It shows a signifi-cant peak within the auditorycortex, displayed onto the indi-vidual anatomy (data not nor-malized).

In this case, only the BOLDsignal of the right auditory cor-tex reached the significance level(p < 0.001). The fitted responsefrom that area, which is dis-played in the right panel of thatfigure, demonstrates very nicelythe temporal delay of the BOLDresponse of 5 - 6 seconds.

Discussion and Conclusions

Both methods are sensitive todetect audition of the pulse se-quence itself through BOLD ac-tivations evoked by EPI’s persis-tent and intermittent genuinegradient noise (p/iGGN), respec-tively. Lateralized processing ob-served for iGGN is in good a-greement with current models ofauditory perceptions of non-ver-bal auditory stimuli.Essentially,the data- and model-drivenanalyses correspond well to eachother and are consistent with thesampling of a HRF evoked byGGN audition. Obviously, EPI isfast enough to catch the initialbut naturally delayed BOLD in-crease in its own T2*-weightedsignal induced by the onset ofpersistent and intermittently ad-vanced scanning noise. Accordingto the PCA, the mean signalwithin the initial six seconds ofcontinuous scanning is indeed be-low the mean of the rest even ifthe BOLD signal may finally at-tenuate to some extent. The par-ticular timing of the FIR modelemphasizes the temporal featuresof the component presumably e-voked by pGGNA. A peaking ofthe corresponding signal laterthan six but not much beyond ten

re may be good reasons for sup-posing non-Gaussian features inour data, especially of method A,the results of the PCA performedwere meaningful interpretable forour purpose. Electrophysiologicalstudies habitually refer to the al-leged disadvantage of BOLDimaging to measure epiphenom-ena but no neural events in realtime. We present the single andmost outstanding case for whichthe temporal delay of the BOLDeffect is not only of advantage buta prerequisite for the demonstra-tion of the presumed neural re-sponse.

This is equivalent to “scanningfor the scanner” through auditoryautodetection of EPI by sound lis-tening to the temporal character-istics of the BOLD responsewhich otherwise limit temporalresolution and functional sensitiv-ity. Notably, coronal slices aresuited to catch both cortical andsubcortical auditory activations.Our data show that audition canbe investigated by just listening tothe scanner. Limiting the band-width will increase the signal-to-noise ratio but lower the GGNlevel. Efficient unilateral earplug-ging can facilitate lateralizedGGN processing. In the future,functional autodetection of EPImay gain clinical value to investi-gate audition with neither particu-lar soft- nor hardware such as sti-mulation devices, trigger signals,or designer EPIs. We are workingon a customized protocol of para-metric iGGNA to increase thesensitivity of the design and to al-low for an objective, clinically use-ful functional assessment of audi-tion within about 6 minutes byjust “scanning for the scanner”.

Acknowledgements

It is our pleasure to thank Dr.Stefan Thesen and Matthias Nit-tka from Siemens Medical Solu-tions for their support of the on-going project.

seconds after the onset of persi-stent scanning clearly supportsthe notion of single superposedBOLD events. To demonstrate li-near superposition of ultrafast sti-muli, the HRF (and its temporalderivative) presumably inducedby audition of rapid, successive s-timulations corresponding to theslice selection gradients (Gz) canbe extracted from SPM’s designmatrix (Sess{s}.bf{i}) set up for asingle run. Then, these vectorsshould be modified to accountfor the discarded scans, expandedto match the number of runs, andentered as user-specified regres-sors only into F-tests. Furthermo-re, nonlinear superposition, i.e.stimulus non-additivity, could beaddressed by expansions of thehaemodynamic convolution th-rough higher-order Volterra ker-nels. This may, however, eventu-ally overspecify the effects sear-ched for on cost of considerableloss of sensitivity. In addition,current SPM may not be most ef-ficient for the analysis of our datamainly because it models tempo-ral autocorrelations within thetime series by a single parameterof the first order autoregressionon the precoloured, i.e. smootheddata which essentially corre-sponds to a smoothing procedureby a kernel of an infinite FWHM.Local (voxel-by-voxel) pre-white-ning with an AR model whosecoefficients are smoothed spa-tially (such as in fmristat) orwithin tissue types (such as in fsl)is probably more sensible. Model-ling by SPM results, however, invery robust estimators. With re-gards to the data-driven analysis,the PCA presented is to catch thedirection of the maximal variancein the data whereas alternativemethods of independent compo-nent analyses (ICA) representhigher order statistics and canfind multiple directions of maxi-mum independence even in non-Gaussian data (see fsl-webpagefor further details). Although the-

XVII Symposium Neuroradiologicum 2002: Detection of the Scanner’s Genuine Gradient Noise ... A.J. Bartsch

1000

Andreas J. Bartsch, M.D.Division of NeuroradiologyBavarian Julius-Maximilians-Universityof WuerzburgJosef-Schneider-Str. 11D-97080 Wuerzburg, GermanyE-mail: bartsch@neuroradiologie.

uni-wuerzburg.de

References1 Ravicz ME, Melcher JR: Isolating the

auditory system from acoustic noiseduring functional magnetic resonanceimaging: Examination of noise con-duction through the ear canal, head,and body. J Acoust Soc Am 109: 216-231, 2001.

2 Bilecen D et Al: The MR tomographas a sound generator: fMRI tool for theinvestigation of the auditory cortex.Magn Reson Med 40: 934-937, 1998.

3 Bandettini PA et Al: Functional MRIof brain activation induced by scanneracoustic noise. Magn Reson Med 39:410-416, 1998.

4 Talavage TM et Al: Measurement ofsignal chnages induced by fMRI Imag-ing Noise. NeuroImage 7: S360, 1998.

5 Worsley KJ et Al: Characterizing theresponse of PET and fMRI data usingMultivariate Linear Models (MLM).NeuroImage 6: 305-319, 1998.