comparison of diffusion-weighted fmri and bold fmri …cogns.northwestern.edu/cbmg/dfmri and bold ni...

8
Comparison of diffusion-weighted fMRI and BOLD fMRI responses in a verbal working memory task Toshihiko Aso a, , Shin-ichi Urayama a , Hidenao Fukuyama a , Denis Le Bihan a, b a Human Brain Research Center, Kyoto University Graduate School of Medicine, Kyoto 606-8507, Japan b NeuroSpin, CEA-Saclay Center, Gif-sur-Yvette CEDEX 91191, France abstract article info Article history: Accepted 2 November 2012 Available online 10 November 2012 Keywords: Diffusion-weighted MRI Functional MRI BOLD contrast Diffusion-weighted functional MRI (DfMRI) has been reported to have a different response pattern in the visual cortex than that of BOLD-fMRI. Especially, the DfMRI signal shows a constantly faster response at both onset and offset of the stimulus, suggesting that the DfMRI signal might be more directly linked to neu- ronal events than the hemodynamic response. However, because the DfMRI response also contains a residual sensitivity to BOLD this hypothesis has been challenged. Using a verbal working memory task we show that the DfMRI time-course features are preserved outside visual cortices, but also less liable to between-subject/ between-regional variation than the BOLD response. The overall ndings not only support the feasibility of DfMRI as an approach for functional brain imaging, but also strengthen the uniqueness of the DfMRI signal origin. © 2012 Elsevier Inc. All rights reserved. Introduction It is known that there are acute changes in the microscopic structure of neural tissues accompanying neuronal activation in the vertebrate brain. Such phenomena were initially observed as a change in tissue optical properties (light transmittance) occurring during activation and interpreted as cell swelling (Andrew and MacVicar, 1994; Lipton, 1973; MacVicar and Hochman, 1991). This hypothesis has been later conrmed through the observation of a transient decrease of the extracellular volume by ion-sensing microelectrodes (Dietzel et al., 1980; Pannasch et al., 2011).On the other hand, a transient decrease in water diffusion in the activated visual cortex has been reported in the human and rat brain using heavily sensitized diffusion fMRI (DfMRI) (Le Bihan et al., 2006; Yacoub et al., 2008). The temporal precedence up to 4 s in time-to-peak of the DfMRI signal response to that of hemodynamic BOLD-fMRI suggests a non-vascular mechanism, superimposed to a residual tissue BOLD component (Aso et al., 2009). This temporal prece- dence of the DfMRI response to hemodynamic events was further extended to near-infrared spectroscopy (NIRS) signals which was close to BOLD response, but not with DfMRI (Kohno et al., 2009). Furthermore, the sharp onset and offset of the DfMRI signal at both the initiation and termination of the stimulus matches the expected neuronal response, while the temporal prole of the BOLD fMRI response lags several sec- onds, reecting the astrocyte response which is known to mediate the blood ow increase (Schummers et al., 2008). Similarly, the asymmetric shape and timing of the intrinsic DfMRI response also closely mimic the infrared transmittance optical signal features observed during the activation of brain slices ex vivo, to the opposite of the hemodynamic BOLD fMRI response. As more direct evidence, the activation-induced decrease in water diffusion has also been observed with diffusion MRI in such brain slices where no blood ow is present (Flint et al., 2009). Based on the exquisite sensitivity of diffusion MRI to variations of cell volume (Anderson et al., 1996; Sotak, 2004), those activation-induced changes in water diffusion have been tentatively ascribed to cell swelling in the neuropil (Le Bihan, 2007; Le Bihan et al., 2006). However, this pro- posal has been challenged by some groups who believe that DfMRI only reects vascular events, although they do not suggest mechanisms (Kim and Kim, 2006; Miller et al., 2007). In the present study our aim was to investigate whether the DfMRI response features initially reported in the visual cortex (Le Bihan et al., 2006) could also be observed in other brain cortical regions activated with more cognitive paradigms. The stability of this DfMRI response (obtained using a spin-echo sequence with b = 1800 s/mm 2 ) was also assessed across brain regions and across individuals, and com- pared with the spatial/individual variations of the BOLD fMRI signals acquired with both spin-echo (SE) MRI (b = 0) and gradient-echo (GRE) sequences, as the two methods are known to be sensitive to different vascular pools (Boxerman et al., 1995; van Zijl et al., 1998). The ndings altogether suggest that BOLD and diffusion fMRI rely on different mechanisms. Materials and methods Subjects Twenty-one healthy Japanese subjects participated in the study. Experiments were approved by a local ethics committee and all NeuroImage 67 (2013) 2532 Corresponding author. Fax: +81 75 751 3202. E-mail address: [email protected] (T. Aso). 1053-8119/$ see front matter © 2012 Elsevier Inc. All rights reserved. http://dx.doi.org/10.1016/j.neuroimage.2012.11.005 Contents lists available at SciVerse ScienceDirect NeuroImage journal homepage: www.elsevier.com/locate/ynimg

Upload: others

Post on 18-Aug-2020

6 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Comparison of diffusion-weighted fMRI and BOLD fMRI …cogns.northwestern.edu/cbmg/dfMRI and BOLD NI 2013.pdf · 2013. 1. 22. · Comparison ofdiffusion-weighted fMRI and BOLD fMRI

NeuroImage 67 (2013) 25–32

Contents lists available at SciVerse ScienceDirect

NeuroImage

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

Comparison of diffusion-weighted fMRI and BOLD fMRI responses in a verbal workingmemory task

Toshihiko Aso a,⁎, Shin-ichi Urayama a, Hidenao Fukuyama a, Denis Le Bihan a,b

a Human Brain Research Center, Kyoto University Graduate School of Medicine, Kyoto 606-8507, Japanb NeuroSpin, CEA-Saclay Center, Gif-sur-Yvette CEDEX 91191, France

⁎ Corresponding author. Fax: +81 75 751 3202.E-mail address: [email protected] (T. Aso).

1053-8119/$ – see front matter © 2012 Elsevier Inc. Allhttp://dx.doi.org/10.1016/j.neuroimage.2012.11.005

a b s t r a c t

a r t i c l e i n f o

Article history:Accepted 2 November 2012Available online 10 November 2012

Keywords:Diffusion-weighted MRIFunctional MRIBOLD contrast

Diffusion-weighted functional MRI (DfMRI) has been reported to have a different response pattern in thevisual cortex than that of BOLD-fMRI. Especially, the DfMRI signal shows a constantly faster response atboth onset and offset of the stimulus, suggesting that the DfMRI signal might be more directly linked to neu-ronal events than the hemodynamic response. However, because the DfMRI response also contains a residualsensitivity to BOLD this hypothesis has been challenged. Using a verbal working memory task we show thatthe DfMRI time-course features are preserved outside visual cortices, but also less liable to between-subject/between-regional variation than the BOLD response. The overall findings not only support the feasibility ofDfMRI as an approach for functional brain imaging, but also strengthen the uniqueness of the DfMRI signalorigin.

© 2012 Elsevier Inc. All rights reserved.

Introduction

It is known that there are acute changes in the microscopic structureof neural tissues accompanying neuronal activation in the vertebratebrain. Such phenomena were initially observed as a change in tissueoptical properties (light transmittance) occurring during activation andinterpreted as cell swelling (Andrew and MacVicar, 1994; Lipton, 1973;MacVicar andHochman, 1991). This hypothesis has been later confirmedthrough the observation of a transient decrease of the extracellularvolume by ion-sensing microelectrodes (Dietzel et al., 1980; Pannaschet al., 2011).On the other hand, a transient decrease in water diffusionin the activated visual cortex has been reported in the human and ratbrain using heavily sensitized diffusion fMRI (DfMRI) (Le Bihan et al.,2006; Yacoub et al., 2008). The temporal precedence up to 4 s intime-to-peak of the DfMRI signal response to that of hemodynamicBOLD-fMRI suggests a non-vascular mechanism, superimposed to aresidual tissue BOLD component (Aso et al., 2009). This temporal prece-dence of the DfMRI response to hemodynamic events was furtherextended to near-infrared spectroscopy (NIRS) signals which was closeto BOLD response, but notwith DfMRI (Kohno et al., 2009). Furthermore,the sharp onset and offset of the DfMRI signal at both the initiation andtermination of the stimulus matches the expected neuronal response,while the temporal profile of the BOLD fMRI response lags several sec-onds, reflecting the astrocyte response which is known to mediate theblood flow increase (Schummers et al., 2008). Similarly, the asymmetricshape and timing of the intrinsic DfMRI response also closely mimicthe infrared transmittance optical signal features observed during the

rights reserved.

activation of brain slices ex vivo, to the opposite of the hemodynamicBOLD fMRI response. As more direct evidence, the activation-induceddecrease in water diffusion has also been observed with diffusion MRIin such brain slices where no blood flow is present (Flint et al., 2009).Based on the exquisite sensitivity of diffusion MRI to variations of cellvolume (Anderson et al., 1996; Sotak, 2004), those activation-inducedchanges inwater diffusion have been tentatively ascribed to cell swellingin the neuropil (Le Bihan, 2007; Le Bihan et al., 2006). However, this pro-posal has been challenged by some groups who believe that DfMRI onlyreflects vascular events, although they do not suggest mechanisms (Kimand Kim, 2006; Miller et al., 2007).

In the present study our aimwas to investigate whether the DfMRIresponse features initially reported in the visual cortex (Le Bihan et al.,2006) could also be observed in other brain cortical regions activatedwith more cognitive paradigms. The stability of this DfMRI response(obtained using a spin-echo sequence with b=1800 s/mm2) wasalso assessed across brain regions and across individuals, and com-pared with the spatial/individual variations of the BOLD fMRI signalsacquired with both spin-echo (SE) MRI (b=0) and gradient-echo(GRE) sequences, as the two methods are known to be sensitive todifferent vascular pools (Boxerman et al., 1995; van Zijl et al., 1998).The findings altogether suggest that BOLD and diffusion fMRI rely ondifferent mechanisms.

Materials and methods

Subjects

Twenty-one healthy Japanese subjects participated in the study.Experiments were approved by a local ethics committee and all

Page 2: Comparison of diffusion-weighted fMRI and BOLD fMRI …cogns.northwestern.edu/cbmg/dfMRI and BOLD NI 2013.pdf · 2013. 1. 22. · Comparison ofdiffusion-weighted fMRI and BOLD fMRI

26 T. Aso et al. / NeuroImage 67 (2013) 25–32

volunteers gave informed consent. Subjects were excluded prior to theexperiment if they had anymedical, neurological, or psychiatric illnessor if they were taking any prescription medications. Furthermore, weselected subjects known for their good head immobility during MRIscanning (b5 mm among all the fMRI volumes of a whole session),based on their previous participation in various BOLD fMRI studiesin our laboratory, to ensure good quality for both BOLD and DfMRIdatasets.

Data acquisition

Visual stimulationThe 2-backworkingmemory paradigm consisted in the randompre-

sentation of single Kanji characters from a small database (20-charactersubset for each run from a pool of 200 Kanjis) projected at 1 Hz createdon a PC using Cogent toolbox (University College London, http://www.vislab.ucl.ac.uk/Cogent/) with MATLAB (MathWorks, Natick, US) anddelivered with a digital light processing projector (Kaga components,Tokyo, Japan) from outside the scanner room. Subjects were asked topress a button when the present Kanji was identical to the second pre-vious one (2 s memory retention period). Each task block lasted for 8 s(8 characters) separated by a 24 s resting period (Fig. 1).

MRI acquisitionBecause of the prohibitive acquisition length to perform DfMRI,

SE and GRE BOLD fMRI acquisitions, our subject population was splitin two groups, one group (13 subjects) with DfMRI/SE-BOLD and theother group (8 subjects) with DfMRI/GRE-BOLD. More subjects wererecruited for SE-BOLD group in order to obtain a clean time course com-pensating for relatively lower signal:noise ratio (SNR) than GRE-BOLD.

All MRI measurements were carried out with a 3 T whole-bodyscanner (Trio, Siemens, Erlangen, Germany) with an 8-channel phased-array head coil. The DfMRI and SE-BOLD acquisition were performedwith a twice-refocusing spin-echo DWI sequence implemented bySiemens for clinical DWI (Aso et al., 2009; Reese et al., 2003), and theGRE-BOLD acquisition was obtained using Siemens standard productsequence. The slice position was selected in semi-coronal orientationcovering the occipital and parietal lobes. Each subject went througheight DfMRI, four SE-BOLD or four GRE-BOLD fMRI runs each lasting312 s in an intermixed and balanced order. Other imaging parameterswere as follows: FOV 192×180 mm, image matrix 64 60 voxels, slicethickness 4 mm, slice gap 2 mm, the number of slices 8, phase-encodedirection R/L, the number of volumes (TRs) 312, partial Fourier factor3/4 (no acceleration), TR 1000 ms, TEs 85/85/30 ms for DfMRI/SE-BOLD/GRE-BOLD, b-values 1800/0 s/mm2 for DfMRI/SE-BOLD. After allfMRI scans, T1-weighted 3D acquisition with MPRAGE sequence wascarried out to obtain anatomical information with the same imagingparameters described in a previous paper (Aso et al., 2009). Adjustablefixation arms of the coil unit restricted the head motion bilaterally. Be-cause diffusion MRI is highly sensitive to body motion (Anderson andGore, 1994), careful instruction was given to the subjects not to movetheir heads during the whole session. Subjects were also asked notto communicate orally during the session if possible, using instead ahand-held rubber squeeze bulb, to avoid jaw joint and headmovement.

Button press

Stimulus

Fig. 1. Single letter, two-back working memory task lasted 8 s separated by 24 s rest.

Data processing

PreprocessingAll MRI data were processed using SPM8 (http://www.fil.ion.ucl.

ac.uk/spm/). Images were corrected for slice-timing of the interleavedacquisition and for head motion. The head motion parameters wereused later as confounding variables in the statistical analysis. All ofthe preprocessing steps were conducted separately for the DfMRI andthe BOLD-fMRI data sets. Automatic spatial registration of DfMRI vol-umes with anatomical images was prone to fail because diffusion MRIimages present asymmetric signal intensity in the white matter dueto diffusion anisotropy effects in fiber bundle. Instead, coregistrationbetween DfMRI and anatomical images was verified by careful visualinspection of brain contours and adjusted, if necessary, by manuallygiving the parameters for a rigid-body transformation using SPM8.

To improve co-registration of the sulcal structure spatial normali-zation to the MNI brain template was done using DARTEL (Ashburner,2007). Since our acquisition FOV did not include anterior part ofthe brain, we limited the bounding box of the final normalized vol-umes at Y=−15 mm plane (MNI atlas). Normalized volumes werere-sliced to have 3 mm isotropic voxels. Finally spatial smoothingwas applied on the volumes using an 8×8×6mm FWHM Gaussiankernel.

First level (individual) activation mapsBrain activation was estimated first for each individual for each of

the fMRI methods (DfMRI, SE-BOLD, GE-BOLD) using the canonicalHemodynamic Response Function (HRF) and its derivative. By includ-ing the first temporal derivative in the design matrix, possible tempo-ral shifts of the response due to inter-area/subject variation can becompensated (Friston et al., 1998). This was necessary because ourgoal was to evaluate response variation across subjects/brain areas.Head motion parameters were added to the individual design matrixin order for the first-order motion artifact to be regressed out. SPManalysis was conducted using the default settings except for the esti-mation of autocorrelation using AR(1) model. We cancelled this fea-ture implemented to model the noises in order to avoid possibledifferential effect on the three signal types because its effect on rela-tively noisy data has not been established. As described below, indi-vidual activation maps were then used to infer brain response ingeneral population based on the random-effect model. We created afigure to demonstrate the distribution of individual activation by thenumber of individuals.

Second level (group) activation mapsIn a previous report we used anatomically-defined ROI in the

calcarine sulci to avoid bias arising from voxel selection method(Aso et al., 2009). This time we conducted a group analysis to checkwhether DfMRI activation is detectable with specificity from broaderbrain regions, using the commonly used approach to process BOLDfMRI data. The group analysis was conducted separately for the threefMRI datasets (DfMRI, SE- and GRE-BOLD) as the second-level analysisbased on a random-effect model (Friston et al., 1999) to build or pop-ulation level inference from individual parametric maps. The resultingmaps were thresholded at pb0.001, uncorrected for multiple compar-isons. We chose this uncorrected threshold to include as many voxelsas possible to evaluate variation in response features across detectedvoxels at the cost of false-positives.

ROI analysesOur secondary aim in conducting group activation mapping was

to test how the temporal relationship between DfMRI and BOLD ispreserved against voxel selection criteria. Since the GRE-BOLD datasetwas from 7 subjects which is small sample for t-test, only the SE-BOLDactivation map was used to define the BOLD-based ROI for the follow-ing analyses.

Page 3: Comparison of diffusion-weighted fMRI and BOLD fMRI …cogns.northwestern.edu/cbmg/dfMRI and BOLD NI 2013.pdf · 2013. 1. 22. · Comparison ofdiffusion-weighted fMRI and BOLD fMRI

Task block Time-to-peak

Decay time (from offset)

Rise time

50% peak amplitude

Fig. 2. Definition of the temporal profile parameters. After filtering by a three-pointmoving average, the last TR of the resting period was used as the baseline from whichthe peak amplitude/latency is measured. The 50% peak height is then used to defineonset/offset slopes.

27T. Aso et al. / NeuroImage 67 (2013) 25–32

Group ROI analysis. The left V1, IPS and basal occipitotemporal junc-tion (basal OTJ, centered in the posterior part of the fusiform gyri)were defined using the Automated Anatomical Labeling (AAL) atlas(Tzourio-Mazoyer et al., 2002) and ROIs were drawn from thoseareas in the group activation maps obtained for DfMRI (n=20) orSE-BOLD (n=13). For the response profile analysis described later,all voxels that are detected by both DfMRI and SE-BOLD group analy-ses were used. For plotting the time course, we used only voxels with-in the 15 mm-radius spheres centered at the cluster peaks in order toeliminate marginal voxels.

The SE BOLD-ROIs were used to extract the signal time course ofthe DfMRI and BOLD fMRI (Fig. 5). Taking the DfMRI identified ROIs

DfMRI

BOLD-fMRI

20

0

Subject count

Fig. 3. Overlayed individual activationmaps fromall subjects for DfMRI (top) andSE-BOLD-fMRI (bottom) (thresholded at P=0.001, uncorrected). The color scale represents thenumber of subjects whose voxels were activated at each location.

as masks, which lied within the SE-BOLD ROIs, did not change theresults significantly.

To further investigate this point, we also looked at the time coursesof voxels detected as activated for SE-BOLD but not for DfMRI, usinga threshold of P=0.05 to create the exclusive mask to assure “noactivation” in DfMRI group SPM.

Amplitude correlation analysis.We conducted a correlation analysison response magnitude of DfMRI/BOLD to see how inter-individualvariation is correlated between DfMRI and BOLD.

Individual ROI analysis. To confirm the findings by group-based ROIresults, signals from individual DfMRI activation map (thresholdedat pb0.001, uncorrected) was extracted from the bilateral primaryvisual (V1, the calcarine sulci), the bilateral IPS and the left basalOTJ. ROIs were defined for each individual automatically by maskingthe activation map with the corresponding AAL region map (left fusi-form gyrus template was used for the basal OTJ).

Amplitude comparison. To compare the 3 fMRI methods in terms ofresponsemagnitude variation, time courseswere extracted separatelyfrom the corresponding activation map from DfMRI or BOLD (eitherSE- or GRE-). This is the only analysis in the present study where

-12

0

+12

+24

+36

+48

y = -64 x = -30

Fig. 4. Group activation maps from the 3 fMRI datasets. Maps are overlaid on the nor-malized gray matter template created by DARTEL from the individual brain volumes.DfMRI cluster is almost overlapped by the SE-BOLD clusters. GRE-BOLD cluster issmall because of a fewer subjects of the GRE-BOLD group.

Page 4: Comparison of diffusion-weighted fMRI and BOLD fMRI …cogns.northwestern.edu/cbmg/dfMRI and BOLD NI 2013.pdf · 2013. 1. 22. · Comparison ofdiffusion-weighted fMRI and BOLD fMRI

28 T. Aso et al. / NeuroImage 67 (2013) 25–32

different set of voxelswere used for the 3 fMRI signals in order to focuson the “best” voxels for the methods.

Response profile analysisThe spatial variation of the response time profiles was investigat-

ed using the voxels that were commonly detected by DfMRI andSE-BOLD group analyses. The signal response in each activated voxelwas extracted and pooled over all subjects before temporal profileswere calculated. In addition temporal smoothing was applied to theraw responses using three-point moving average before generatingspatial maps of the response time profile features. The following pa-rameters were investigated: Time-to-peak from the task onset, 50%rise time after stimulus onset and 50% decay time after stimulus off-set. Those rise and decay times were defined as the time when thesignal reaches 50% of the difference between response peak and thelast time point before the task onset and offset, respectively. We cre-ated maps showing these parameters as voxel values to visualize itsspatial distribution (Fig. 2). We also created histograms of the valuesfrom these maps using all voxels.

Results

Individual and group activation maps

Brain activation was successfully detected in the dorsal and ven-tral visual information streams (Mishkin and Ungerleider, 1982) inboth DfMRI and BOLD-fMRI maps. Although the voxel clusters weresmaller and more scattered in DfMRI than BOLD, they clearly distrib-uted to certain specificity in the primary visual (V1), IPS and basal OTJcortices (Fig. 3).

The group activation maps were similar across the three fMRImethods, demonstrating selective activations centered in the primaryto extrastriate visual cortices, as well as in the banks of IPS (Fig. 4).The overlapping voxels form a white cluster in the center. As can befound in the maps, both cluster size and the number of voxels in eachROI were smaller in DfMRI than in SE-BOLD: DfMRI/SE-BOLD in LeftV1=228/624, Left IPS 55/675, Left basal OTJ 40/183).

-0.5

0

0.5

1

0 5 10 15 20 25 30-0.5

0

0.5

1

Lt V1, 624 voxels Lt IPS,

-0.2

0

0.2

0.4

0 5 10-0.5

0

0.5

1

DfMRI (b=1800), n=20SE-BOLD, n=13GRE-BOLD, n=7

Time (s)

perc

ent s

igna

l cha

nge

norm

aliz

ed

Fig. 5. fMRI time courses extracted from group-activated SE-BOLD clusters. Using three-poismall crosses. The response profiles such as the time-to-peak and amplitude are essentiallchange plots (top) have error bars indicating standard error of mean over subjects.

Signal responses in the BOLD group activation map

Fig. 5 shows the signal responses extracted from the group-wiseactivated voxels defined by SE-BOLD fMRI. The normalized DfMRIresponse showed clear time precedence over both GE and SE BOLDresponses, confirming our previous report in the visual cortex (Asoet al., 2009; Le Bihan et al., 2006). Interestingly, while the time profileof the diffusion response is uniform across V1 and IPS, the initial slopeand the late undershoot differed between SE- andGRE-BOLD responses,and betweenbrain regions. Although the onsetwas later in SE thanGRE,the time-to-peak and the offset decay time were not so much differentbetween the two BOLD signals.

Fig. 6 shows signal response pooled over all 1846 voxels fromwhole brain that were detected with SE-BOLD, but not with DfMRIeven at the low threshold of p=0.05, uncorrected. A DfMRI responsecan nevertheless be seen with a marked temporal precedence overthe BOLD response (time-to-peak of DfMRI/SE/GRE-BOLD=7.5/9.5/9.5 s). Failure to detect those voxels as activated likely results fromusing the HRF model. Such a fast response would be better detectableusing the appropriate DhRF model (Aso et al., 2009).

Amplitude correlation analysisFig. 7A shows the inter-subject correlation between the DfMRI and

BOLD response amplitudes. Therewas no significant correlation betweenthe DfMRI and BOLD magnitude over subjects (r=0.18, P=0.45).

Individual ROI analysis

Fig. 7B shows the individual response magnitude from the 20subjects. The corresponding activation map for each method wasused in this analysis, unlike the previous one, in order to measurehighest response amplitude in each individual. There was smallerinter-subject variation in DfMRI magnitude than BOLD as shown bythe error bar (95% confidence interval), suggesting uncommonmech-anisms of signal change.

The raw signal time courses extracted from those regions are shownin Supplementary Fig. 1. The diffusion-weighted signal response started

Extracted from SE-BOLD group activation cluster

675 voxels

15 20 25 30

-0.5

0

0.5

1

0 5 10 15 20 25 30-0.5

0

0.5

1

Lt basal OTJ, 183 voxels

nt moving average, the curves are provided in addition to the raw values indicated byy the same with those from individual clusters (Supplementary Fig. 1). Percent signal

Page 5: Comparison of diffusion-weighted fMRI and BOLD fMRI …cogns.northwestern.edu/cbmg/dfMRI and BOLD NI 2013.pdf · 2013. 1. 22. · Comparison ofdiffusion-weighted fMRI and BOLD fMRI

-0.3

0.0

0.3

0.6

-5 0 5 10 15 20 25 30 35

Voxels detected only by BOLD fMRI

DfMRI

SE-BOLD

GRE-BOLD

scaled DfMRI

Time (s)

Per

cent

sig

nal c

hang

e

Fig. 6. Response time courses averaged over all the 1180 voxels that were detected by SE-BOLD group analysis (p=0.001, uncorrected, 2706 voxels) but not by DfMRI (p=0.05,uncorrected, 1149 voxels). There is small DfMRI response peaking at 7.5 s after onset,which is too fast for theHRFmodel resulting in false-negative. Error bars indicate standarderror of mean over subjects (n of D/SE/GRE=20/13/7).

Table 1Response properties compared over regions and fMRI methods (Mean±SD). All voxelsin the SE-BOLD group activation map masked with the AAL anatomical templates of thethree cortical regions were used.

DfMRI Group 1/2 V1(calcarine sulci)

Left parietal lobe Left fusiformgyrus

SE-BOLD

GRE-BOLD

Time-to-peak 7.0±1.4 6.6±2.0 7.9±3.510.6±1.2 10.4±0.9 10.2±2.19.3±1.4 9.7±2.2 8.9±2.1

50% rise time 2.9±1.8 3.1±2.1 4.3±4.26.5±2.2 6.8±2.0 6.5±2.24.2±0.8 5.1±1.1 4.4±0.7

50% decay timefrom offset

2.0±2.0 2.4±3.4 1.8±2.35.6±1.4 5.7±1.6 5.3±1.85.3±1.5 5.7±2.7 4.7±2.5

ROI center(s)of gravity

(2.4, −76.5, 7.5) Left/Right (−33.2, −48.7,−16.0)(−34.9, −51.8, 57.0)/

(33.8, −53.6, 57.0)

29T. Aso et al. / NeuroImage 67 (2013) 25–32

and ended earlier than the BOLD signal in all three separate regions,indicating the precedence of the diffusion response over BOLD is notspecific to visual cortices. Response parameters are summarized inTable 1. The difference in time-to-peak from BOLD response was signif-icant in all three regions by paired t-test across subjects (V1: Pb0.0001,IPS: P=0.005, basal OTJ: P=0.0004). Comparing the left 2 panels,the temporal profile of the DfMRI response was preserved betweenthe two different subject groups.

Response profile mapping

To investigate the spatial variations of the temporal profiles of theresponses we generated maps of the time-to-peak, rise and decaytimes from the group data. Fig. 8A shows the histogram of the voxelpopulation for time-to-peak and 50% decay time. Because we appliedtemporal smoothing before calculating those timing parameters in

A

0.0

0.5

1.0

1.5

2.0

0.0 0.5 1.

Mea

n B

OLD

mag

nitu

de

Mean DfMRI magnitude

Diffusion VS Bold magnitude in BOLD group activation clusters (2706 voxels), 20 subjects

SE-BOLD GRE-BOLD

Fig. 7. A, Comparison of inter-individual response variation measured in all the activatedbetween the diffusion and BOLD (either of SE/GRE) response amplitudes. No correlation is fclusters of corresponding fMRI signals. DfMRI response shows smaller inter-individual variatrespectively).

order to obtain reliable values from each single voxel, time-to-peakappear sometime shorter than the stimulus duration (8 s) for theDfMRI response (Fig. 8A, top). It is also found in the smoothed timecourse plots in the Fig. 5 peaking just before the offset, while the orig-inal signal starts to decrease only after the offset (crosses around thelines, Fig. 5). The histograms show that the DfMRI response is fasterthan BOLD in most voxels. The offset-time distribution appeared largerfor the DfMRI responses than the BOLD responses. From the profilemaps (Fig. 8B), it appears that the BOLD response varies largely be-tween brain regions while the DfMRI response is more homogeneous,with some variations within clusters (arrowhead).

Discussion

The present results confirmed earlier reports that the DfMRI signalresponse precedes the BOLD signal in the visual cortex (Aso et al.,2009; Le Bihan et al., 2006). Further, the temporal precedence seenin the visual cortex during sensory stimulation was also found outside

0.0

0.5

1.0

1.5

2.0

perc

ent m

agni

tude

Diffusion SE-BOLD GRE-BOLD

Highest response magnitudein each individual

B

0

voxels by SE-BOLD fMRI. Each point represents one subject to show the relationshipound (r=0.18, P=0.45). B, Amplitude of the pooled responses in individual activationion. Error bars show confidence interval (n=20, 13 and 7 for DfMRI, SE- and GRE-BOLD

Page 6: Comparison of diffusion-weighted fMRI and BOLD fMRI …cogns.northwestern.edu/cbmg/dfMRI and BOLD NI 2013.pdf · 2013. 1. 22. · Comparison ofdiffusion-weighted fMRI and BOLD fMRI

0 2 4 6 8 second

second second from the offset

Onset rise time Decay time from offset

voxe

l cou

nt

DfMRI SE-BOLD GRE-BOLD

DfMRI SE-BOLD GRE-BOLD DfMRI SE-BOLD GRE-BOLD

Time-to-peak

−53

−95

−92

−89

−62

−86

Y=

SE-BOLDGRE-BOLD

0 2 4 6 8 100

20

40

60

80

100

120

Decay time

4 6 8 10 12 140

50

100

150

200

Time to peak

6 8 10 12

second from onset

DfMRI

A

B

Fig. 8. A, Histograms showing distribution of time-to-peak and offset time among voxels. All voxels that are commonly detected by DfMRI and SE-BOLD group analyses were used.DfMRI response is faster than BOLD in both measures, but with a broad histogram only in decay time. The narrow histograms for time-to-peak (and rise time, not shown) mayindicate homogeneity in source of signal change while offset decay time suffers contamination by BOLD effect. B. Maps showing spatial distribution of the temporal profile. The3 parameters are color-coded in commonly activated voxels (6 s–12 s after onset for time-to-peak and 0 s–8 s for 50% rise/decay time). Within-cluster variation is small in therise time of DfMRI response.

30 T. Aso et al. / NeuroImage 67 (2013) 25–32

the occipital lobe using a cognitive task and survived a random-effectmodel in a similarmanner as BOLD-fMRI. Theway the DfMRI responsevaried across regions and individuals also differed from BOLD, furtherreinforcing the assumption that the physiological surrogate of theDfMRI and BOLD signals is different.

An important feature of the present study is the use of a commonresponse template, the canonical HRF and its temporal derivative, toassure non-biased comparison between methods. This strategy leads,however, to a slight decrease in sensitivity for DfMRI which would bebetter analyzed using a DRF (or DhRF) template (Aso et al., 2009), as

Page 7: Comparison of diffusion-weighted fMRI and BOLD fMRI …cogns.northwestern.edu/cbmg/dfMRI and BOLD NI 2013.pdf · 2013. 1. 22. · Comparison ofdiffusion-weighted fMRI and BOLD fMRI

31T. Aso et al. / NeuroImage 67 (2013) 25–32

suggested by the presence of false negative voxels in DfMRI maps(Fig. 6). The localization of the DfMRI “activated” clusters was in verygood agreement with those revealed by BOLD-fMRI, although the clus-ters remain smaller. The use of a suboptimal template and the lowerSNR of the DfMRI data are certainly responsible this difference in size,but we cannot exclude a differential distribution of the physiologicalsources at the origin of the DfMRI and BOLD signal. This issue shouldawait future studies with higher spatial resolution.

In spite of the similarity in activity localization, the DfMRI andBOLD signal responses differed in many aspects. Regarding the ampli-tude it is particularly interesting to note that the variance of theresponse magnitude was lower for DfMRI than BOLD in spite of rela-tively higher signal SNR in BOLD fMRI (Aso et al., 2009). We alsofailed to find any amplitude correlation between those two methodsover subjects. Those observations are compatible with the fact theDfMRI signal might be more directly related to neuronal responsesthat the BOLD signal, as the degree of neurovascular coupling variesacross brain regions and individuals (Handwerker et al., 2004). Onthe other hand the temporal precedence of diffusion response wasclearly present across subjects in all activated regions at the onset,time-to-peak and offset of the responses. Furthermore, the varianceof those time markers across regions (such as V1 and the parietallobe) was significantly lower for DfMRI than BOLD, again suggestingthat the DfMRI signal is closer to the neuronal response than thehemodynamic BOLD response which is known to be highly variable.Looking more precisely at our DfMRI results, it is important to notethat the variation in the offset-time was larger than in the onset-time across regions. This finding is perfectly in line with our model(Aso et al., 2009) which predicts that the DfMRI response is 100%driven by diffusion at onset and 100% driven by the residual BOLDresponse seconds after the end of the stimulus, especially during thepost-stimulus undershoot.

Interestingly, there was a notable delay in the initial slope ofthe SE-BOLD response relative to GRE-BOLD. To our knowledge thisdifference in time course between SE and GRE BOLD has not beenreported earlier. In fact not many attempts have beenmade to comparein detail the response profile between those sequences, especially atthis level of precision and with our subject sample size. For example,Hulvershorn et al. (2005) showed an earlier response of SE-BOLD thanGRE-BOLD by 1 s in the time-to-peak of an impulse response functionusing only four subjects. Their analysis was based on a linear system as-sumption,whichmay not holdwith longer stimuli (Birn and Bandettini,2005). Also, our TE was significantly longer because we used our DfMRIsequence with b=0, which might shift the BOLD sensitivity balancetowards different pools of blood vessels (Boxerman et al., 1995; vanZijl et al., 1998). In any case, since the SE-BOLD and DfMRI signals wereobtained with the uniform TE the temporal precedence of the DfMRIresponse reported here is, thus, particularly relevant and significant.

Our results clearly show that the BOLD and DfMRI signals differin many aspects, suggesting that both approaches have differentphysiological origins. While the hemodynamic nature of the BOLD re-sponse has been well established, the neuronal origin of the genuinediffusion component of the DfMRI response, and even more its rela-tion with cell swelling, remain, of course, to be demonstrated. Indeed,several groups have tried to reproduce our earlier results, howeverwith mitigated success. For instance, Autio et al. (2011) failed to ob-serve a time difference between BOLD and DfMRI responses. Therewere, however, two major differences in their experimental settingscompared to ours: The magnetic field strength and the brain regioninvestigated (primary somatosensory cortex). It is known that thediffusion-weighted DfMRI signal response includes a residual BOLDcomponent and only the calculation of the ADC from the DfMRI re-sponse would provide a BOLD free response (Le Bihan et al., 2006).While the diffusion component of the response does not depend onmagnetic field, the BOLD response is, by principle, largely sensitiveto field strength. At 7 T the BOLD signal change is expected to be

1.5–2 times larger than at 3 T (Schäfer et al., 2008), and even morewhen intravascular signal is suppressed by diffusion weighting (Yacoubet al., 2003). Hence, the fraction of the BOLD component to the DfMRIsignal at high field is expected to be much higher and may overshadowthe pure diffusion response, so that differences between DfMRI andBOLD responses may be more difficult to observe (Yacoub et al., 2008).As for the primary somatosensory cortex, it is known for exhibiting arather fast and strong onset response followed by a quick habituation(Nangini et al., 2005), as confirmed from electrophysiological recording(Schroeder and Foxe, 2002). Furthermore, fMRI responses in the so-matosensory areas invoked by electric finger nerve stimulation are noto-riously unstable (Spiegel et al., 1999).

In the cat visual cortex Jin et al. (2008) failed to detect uniqueresponse with a gradient-echo based diffusion MRI sequence. Thisis not surprising, as effects of local magnetic field gradients inducedby deoxyhemoglobin are much larger with gradient-echo based se-quences. The abolition of the spin-echo response in the presence ofiron based contrast agents is also not surprising, as it results in astrong negative response largely overwhelming the (positive) BOLDresponse and interfering with the diffusion response (Le Bihan etal., 2012). In summary, in light of those experimental differences,we consider that those earlier reports, although interesting, dot notallow to reject the existence of a pure diffusion component in theDfMRI signal reflecting acute structural changes in the cortex tissue,as acknowledged by those authors.

Indeed, some particular care is necessary to successfully carry outDfMRI experiments, mainly because the MRI signal obtained at highlevel of diffusion sensitization is low. Efforts must be made to obtaina SNR as high as possible (for instance by reducing TE as possible,which implies that gradient pulses are played on 3 axes simulta-neously, by using phased-array coils, parallel imaging and partial Fou-rier acquisition schemes, etc.), while keeping sources of signalfluctuations (head motion, mechanical vibrations (Gallichan et al.,2010), electrical instabilities in the gradient hardware, etc.) as low aspossible. It is always a good practice to check the SNR and the signalstability level (for instance by looking at maps of the signal varianceacross slices) before analyzing DfMRI data.

In conclusion, while it is clear that BOLD fMRI is an easier andmorerobust approach to perform fMRI, especially in human subjects, DfMRIpotentially presents important advantages. If its neuronal origin isconfirmed, DfMRI may overtake BOLD fMRI when neurovascular cou-pling fails, preventing BOLD responses to be seenwhile the underlyingbrain activity remains normal. Such uncoupling has been reported inspecific situation, such as tissue close to arteriovenous malformation(Lehericy et al., 2002), or in the presence of drugs which mightinterfere with the neurovascular coupling, such as antihistaminicagents and more importantly anesthetic drugs used in animal studies(Masamoto et al., 2007). DfMRI may also be a better approach whenresponse time courses need to be known with some accuracy, forinstance to look at synchronization across brain regions during activa-tion or estimate resting state network with higher frequency compo-nent than allowed with BOLD fMRI. In any case, DfMRI would highlybenefit from technical progress to improve SNR, especially dedicatedgradient systems for diffusion-weighting (Feldman et al., 2011).

Supplementary data to this article can be found online at http://dx.doi.org/10.1016/j.neuroimage.2012.11.005.

References

Anderson, A.W., Gore, J.C., 1994. Analysis and correction of motion artifacts in diffusionweighted imaging. Magn. Reson. Med. 32, 379–387.

Anderson, A.W., Zhong, J., Petroff, O.A., Szafer, A., Ransom, B.R., Prichard, J.W., Gore, J.C.,1996. Effects of osmotically driven cell volume changes on diffusion-weighted imagingof the rat optic nerve. Magn. Reson. Med. 35, 162–167.

Andrew, R.D., MacVicar, B.A., 1994. Imaging cell volume changes and neuronal excitationin the hippocampal slice. Neuroscience 62, 371–383.

Ashburner, J., 2007. A fast diffeomorphic image registration algorithm. Neuroimage 38,95–113.

Page 8: Comparison of diffusion-weighted fMRI and BOLD fMRI …cogns.northwestern.edu/cbmg/dfMRI and BOLD NI 2013.pdf · 2013. 1. 22. · Comparison ofdiffusion-weighted fMRI and BOLD fMRI

32 T. Aso et al. / NeuroImage 67 (2013) 25–32

Aso, T., Urayama, S.I., Poupon, C., Sawamoto, N., Fukuyama, H., Bihan, D.L., 2009. Anintrinsic diffusion response function for analyzing diffusion functional MRI timeseries. Neuroimage 47, 1487–1495.

Autio, J.A., Kershaw, J., Shibata, S., Obata, T., Kanno, I., Aoki, I., 2011. High b-valuediffusion-weighted fMRI in a rat forepaw electrostimulation model at 7 T. Neuroimage57, 140–148.

Birn, R.M., Bandettini, P.A., 2005. The effect of stimulus duty cycle and “off” duration onBOLD response linearity. Neuroimage 27, 70–82.

Boxerman, J.L., Bandettini, P.A., Kwong, K.K., Baker, J.R., Davis, T.L., Rosen, B.R.,Weisskoff, R.M., 1995. The intravascular contribution to fMRI signal change:Monte Carlo modeling and diffusion-weighted studies in vivo. Magn. Reson. Med.34, 4–10.

Dietzel, I., Heinemann, U., Hofmeier, G., Lux, H.D., 1980. Transient changes in the size ofthe extracellular space in the sensorimotor cortex of cats in relation to stimulus-induced changes in potassium concentration. Exp. Brain Res. 40, 432–439.

Feldman, R.E., Scholl, T.J., Alford, J.K., Handler, W.B., Harris, C.T., Chronik, B.A., 2011.Results for diffusion-weighted imaging with a fourth-channel gradient insert.Magn. Reson. Med. 66, 1798–1808.

Flint, J., Hansen, B., Vestergaard-Poulsen, P., Blackband, S.J., 2009. Diffusion weightedmagnetic resonance imaging of neuronal activity in the hippocampal slice model.Neuroimage 46, 411–418.

Friston, K.J., Josephs, O., Rees, G., Turner, R., 1998. Nonlinear event-related responses infMRI. Magn. Reson. Med. 39, 41–52.

Friston, K.J., Holmes, A.P., Price, C.J., Buchel, C., Worsley, K.J., 1999. Multisubject fMRIstudies and conjunction analyses. Neuroimage 10, 385–396.

Gallichan, D., Scholz, J., Bartsch, A., Behrens, T.E., Robson, M.D., Miller, K.L., 2010.Addressing a systematic vibration artifact in diffusion-weighted MRI. Hum. BrainMapp. 31, 193–202.

Handwerker, D.A., Ollinger, J.M., D'Esposito, M., 2004. Variation of BOLD hemodynamicresponses across subjects and brain regions and their effects on statistical analyses.Neuroimage 21, 1639–1651.

Hulvershorn, J., Bloy, L., Gualtieri, E.E., Leigh, J.S., Elliott, M.A., 2005. Spatial sensitivityand temporal response of spin echo and gradient echo bold contrast at 3 T usingpeak hemodynamic activation time. Neuroimage 24, 216–223.

Jin, T., Kim, S.G., 2008. Functional changes of apparent diffusion coefficient during visu-al stimulation investigated by diffusion-weighted gradient-echo fMRI. Neuroimage41, 801–812.

Kim, T., Kim, S.G., 2006. Quantification of cerebral arterial blood volume using arterialspin labeling with intravoxel incoherent motion-sensitive gradients. Magn. Reson.Med. 55, 1047–1057.

Kohno, S., Sawamoto, N., Urayama, S.I., Aso, T., Aso, K., Seiyama, A., Fukuyama, H., Le Bihan,D., 2009. Water-diffusion slowdown in the human visual cortex on visual stimulationprecedes vascular responses. J. Cereb. Blood Flow Metab. 29, 1197–1207.

Le Bihan, D., 2007. The ‘wet mind’: water and functional neuroimaging. Phys. Med. Biol.52, R57–R90.

Le Bihan, D., Urayama, S.I., Aso, T., Hanakawa, T., Fukuyama, H., 2006. Direct and fastdetection of neuronal activation in the human brain with diffusion MRI. Proc.Natl. Acad. Sci. U. S. A. 103, 8263–8268.

Le Bihan, D., Joly, O., Aso, T., Uhrig, L., Poupon, C., Tani, N., Iwamuro, H., Urayama, S.,Jarraya, B., 2012. Brain tissue water comes in two pools: evidence from diffusionand R2′measurements with USPIOs in non human primates. Neuroimage 62, 9–16.

Lehericy, S., Biondi, A., Sourour, N., Vlaicu, M., du Montcel, S.T., Cohen, L., Vivas, E.,Capelle, L., Faillot, T., Casasco, A., Le Bihan, D., Marsault, C., 2002. Arteriovenousbrain malformations: is functional MR imaging reliable for studying language reor-ganization in patients? Initial observations. Radiology 223, 672–682.

Lipton, P., 1973. Effects of membrane depolarization on light scattering by cerebral cor-tical slices. J. Physiol. 231, 365–383.

MacVicar, B.A., Hochman, D., 1991. Imaging of synaptically evoked intrinsic optical sig-nals in hippocampal slices. J. Neurosci. 11, 1458–1469.

Masamoto, K., Kim, T., Fukuda, M., Wang, P., Kim, S.-G., 2007. Relationship betweenneural, vascular, and BOLD signals in isoflurane-anesthetized rat somatosensorycortex. Cereb. Cortex 17, 942–950.

Miller, K.L., Bulte, D.P., Devlin, H., Robson, M.D., Wise, R.G., Woolrich, M.W., Jezzard, P.,Behrens, T.E., 2007. Evidence for a vascular contribution to diffusion FMRI at high bvalue. Proc. Natl. Acad. Sci. U. S. A. 104, 20967–20972.

Mishkin, M., Ungerleider, L.G., 1982. Contribution of striate inputs to the visuospatialfunctions of parieto-preoccipital cortex in monkeys. Behav. Brain Res. 6, 57–77.

Nangini, C., Macintosh, B.J., Tam, F., Staines, W.R., Graham, S.J., 2005. Assessing lineartime-invariance in human primary somatosensory cortex with BOLD fMRI usingvibrotactile stimuli. Magn. Reson. Med. 53, 304–311.

Pannasch, U., Vargová, L., Reingruber, J., Ezan, P., Holcman, D., Giaume, C., Syková, E.,Rouach, N., 2011. Astroglial networks scale synaptic activity and plasticity. Proc.Natl. Acad. Sci. U. S. A. 108, 8467–8472.

Reese, T.G., Heid, O., Weisskoff, R.M., Wedeen, V.J., 2003. Reduction of eddy-current-in-duced distortion in diffusion MRI using a twice-refocused spin echo. Magn ResonMed 49, 177–182.

Schäfer, A., van der Zwaag, W., Francis, S.T., Head, K.E., Gowland, P.A., Bowtell, R.W., 2008.High resolution SE-fMRI in humans at 3 and 7 T using a motor task. MAGMA 21,113–120.

Schroeder, C.E., Foxe, J.J., 2002. The timing and laminar profile of converging inputs tomultisensory areas of the macaque neocortex. Cogn. Brain Res. 14, 187–198.

Schummers, J., Yu, H., Sur, M., 2008. Tuned responses of astrocytes and their influenceon hemodynamic signals in the visual cortex. Science 320, 1638–1643.

Sotak, C.H., 2004. Nuclearmagnetic resonance (NMR)measurement of the apparent dif-fusion coefficient (ADC) of tissue water and its relationship to cell volume changesin pathological states. Neurochem. Int. 45, 569–582.

Spiegel, J., Tintera, J., Gawehn, J., Stoeter, P., Treede, R.D., 1999. Functional MRI ofhuman primary somatosensory and motor cortex during median nerve stimula-tion. Clin. Neurophysiol. 110, 47–52.

Tzourio-Mazoyer, N., Landeau, B., Papathanassiou, D., Crivello, F., Etard, O., Delcroix, N.,Mazoyer, B., Joliot, M., 2002. Automated anatomical labeling of activations in SPMusing a macroscopic anatomical parcellation of the MNI MRI single-subject brain.Neuroimage 15, 273–289.

van Zijl, P.C., Eleff, S.M., Ulatowski, J.A., Oja, J.M., Ulug, A.M., Traystman, R.J., Kauppinen,R.A., 1998. Quantitative assessment of blood flow, blood volume and blood oxy-genation effects in functional magnetic resonance imaging. Nat. Med. 4, 159–167.

Yacoub, E., Duong, T.Q., Van De Moortele, P.F., Lindquist, M., Adriany, G., Kim, S.G.,Uğurbil, K., Hu, X., 2003. Spin-echo fMRI in humans using high spatial resolutionsand high magnetic fields. Magn. Reson. Med. 49, 655–664.

Yacoub, E., Uludag, K., Ugurbil, K., Harel, N., 2008. Decreases in ADC observed in tissueareas during activation in the cat visual cortex at 9.4 T using high diffusion sensi-tization. Magn. Reson. Imaging 26, 889–896.