Documentos de Académico
Documentos de Profesional
Documentos de Cultura
Cerebral Cortex
doi:10.1093/cercor/bhs008
Analysis of Time and Space Invariance of BOLD Responses in the Rat Visual System
Christopher J. Bailey1,2,3,4, Basavaraju G. Sanganahalli1,2,5, Peter Herman1,2,5, Hal Blumenfeld2,6,7, Albert Gjedde3,4 and
Fahmeed Hyder1,2,5,8
1
Magnetic Resonance Research Center (MRRC) and 2Core Center for Quantitative Neuroscience with Magnetic Resonance
(QNMR), Yale University, New Haven, CT 06520, USA, 3Center of Functionally Integrative Neuroscience, Aarhus University,
DK-8000 Aarhus, Denmark, 4Pathophysiology and Experimental Tomography Center, Aarhus University Hospitals, DK-8000 Aarhus,
Denmark, 5Department of Diagnostic Radiology and 6Department of Neurology and 7Department of Neurobiology and 8Department
of Biomedical Engineering, Yale University, New Haven, CT 06520, USA
Address correspondence to D. S. Fahmeed Hyder, N143 TAC (MRRC), 300 Cedar Street, Yale University, New Haven, CT 06520, USA. Email:
fahmeed.hyder@yale.edu.
Neuroimaging studies of functional magnetic resonance imaging time series from human experiments, both with and without
(fMRI) and electrophysiology provide the linkage between neural task (Frackowiak 2004; Huettel et al. 2004). When applied in
activity and the blood oxygenation level--dependent (BOLD) response. the most common context of activation detection, the GLM
The Author 2012. Published by Oxford University Press. All rights reserved.
For permissions, please e-mail: journals.permissions@oup.com
a single-trial level. Although neural and BOLD responses to adapted for the measurements. A standard block stimulation protocol was
blue/green light generally agreed with each other, responses used in all experiments with at least 300 s between consecutive trials.
from V1 and SC were noticeably different. We find that time-
invariance holds for a wide range of flash rates in SC but not in Electrophysiology Experiments
V1, whereas IRFSC and IRFV1 are both space invariant. These Rats were mounted on a stereotaxic frame (David Kopf Instruments,
results suggest caution in using GLM of BOLD responses Tujunga, CA) placed on a vibration-free table inside a Faraday cage. The
because of the potential to misconstrue underlying neural scalp and the ‘‘galea aponeurotica’’ were removed and small burr holes
activity for varying stimulus rates. were drilled for insertion of high-impedance tungsten microelectrodes
(2--4 MX; FHC, Bowdoinham, ME) that measured neural electrical
signals. The coordinates for electrode placement were guided by our
Materials and Methods fMRI data and the atlas of Paxinos and Watson (1998). With reference
to bregma and the sagittal midline plane, the electrodes were placed in
Animal Preparation the following coordinates: V1 (–6.5 mm posterior, 3.5 mm lateral, ~1
All procedures were performed in accordance with protocols approved mm from dura) and SC (–7.0 mm posterior, 0.8 mm lateral, 2.4--2.6 mm
by the Yale University Institutional Animal Care and Use Committee, in from dura). Electrical signals were digitized at 20 kHz with CED l1401
agreement with the National Institutes of Health Guide for the Care and using Spike2. LFP and MUA were obtained by splitting the electrical
Use of Laboratory Animals. Experiments were conducted on artificially signal into low ( <300 Hz) and high (400 Hz--10 kHz) frequency bands
ventilated (70% N2O/30% O2) adult male rats (Long Evans; 220--350 g; using an online analogue Butterworth filter (24 dB/oct attenuation).
Charles River, Wilmington, MA). A femoral artery was cannulated (PE-50) Off-line data analysis was performed using custom scripts in Matlab
Page 2 of 13 Dynamic and Regional BOLD Impulse Response Functions in Rat Visual System d
Bailey et al.
estimating h(t) from BOLD-MUA pairs is ill-conditioned, we used the colors is practically complete, with fewer significantly modu-
gamma variate function (Madsen 1992) as the IRF lated voxels during red than blue/green stimulation, particularly
in the SC (Supplementary Fig. 2). We, therefore, analyzed trials
t 1ttp a
hðt Þ=A e ; ð2Þ with green and blue light, whereas red light trials are not
tp
included in subsequent analysis.
where A is a scaling constant, tp is the peak time, and a is related to the
width of the function. We used standard nonlinear optimization (Matlab
function lsqcurvefit with trust region reflective algorithm) to find the
Differential V1 and SC Response Characteristics
parameters {A,tp,a} that
minimize the squared difference between Because V1 and SC signals were measured simultaneously,
measured BOLDiR ðt Þ and modeled BOLD h R ðt Þ5MUAjR ðt Þ for i,j = whether they are in fMRI studies with the rat inside the magnet
{1,. . . ,11} (121 pairs). The final parameter estimates were obtained by or electrophysiology recordings with the rat on the bench, we
averaging over all fits after regularizing against outliers by removing the quantified BOLD and neural response reproducibility over trials
5% of the largest and smallest tp 3 a-products (corresponding to late
and animals by calculating the pairwise cross correlations of V1
sharp and early broad peaks, respectively).
and SC time courses in 11 representative trials in 6 animals and
plotted the squared correlation coefficient (Pearson R2) value
GLM of BOLD Response for each V1 3 V1, SC 3 SC, and V1 3 SC trial pair (gray scale
A model-based analysis was performed on a subset of the data to coded 11 3 11 matrix). We used the 1-Hz flash rate data for this
examine frequency-dependent effects. A single gamma variate was analysis. The darker shade in the correlation matrices in
chosen as the canonical HRF with tp =3.23 s and a = 11.33. The boxcar
flash rate are summarized in Table 1. For both of these features modeled BOLD responses based on the convolution between
observed in SC and V1, it can be argued that the neurovascular MUA and the IRF. We used the 1 Hz data shown in Figure 2 for
uncoupling is maximum at 10-Hz flash rate, but the trend is the estimation of h (t), leading to 121 pairs of MUA and BOLD
partly detectable at 5 Hz (but absent at 1 Hz) flash rate. signals in both V1 and SC. For numerical tractability, we limited
Supplementary Figures 4 and 5 represent the time-frequency the optimization to the 3 parameters (amplitude, peak, and
decompositions of the 5- and 10-Hz LFP data and illustrate the width) of the gamma variate family of functions. Furthermore,
good agreement between the MUA data shown in Figure 3 and the fitting time of each data set was limited to [–5, 40] s (i.e., to
the 80--200 Hz band LFP signals. mainly the stimulation period; for details on the estimation
procedure, see Materials and Methods).
IRFV1 and IRFSC for 1-Hz Flash Rate The estimated IRFs for V1 and SC data at 1-Hz flash rate are
To quantify the extent of neurovascular coupling in V1 and SC, shown in Figure 4. The peak and width parameters of the IRFs
we estimated the parameters of the IRF (i.e., h (t) in equations are of similar magnitude and estimates statistically indistin-
(1) and (2)) that lead to optimal fits between the measured and guishable (overlapping SDs): IRFV1 and IRFSC peak at 2.5 ± 0.5 s
Page 4 of 13 Dynamic and Regional BOLD Impulse Response Functions in Rat Visual System d
Bailey et al.
Downloaded from http://cercor.oxfordjournals.org/ at Yale University on February 3, 2012
Figure 2. BOLD and neural temporal response profiles in V1 and SC differ significantly. (A) Trial-by-trial reproducibility analysis of 11 examples of BOLD and neural responses to 1-
Hz flash stimulation (data from 6 animals). Each matrix, BOLD signal left, and MUA right, illustrates the R2-values for pairwise cross correlations between single trial time courses
from the same region (V1 3 V1 and SC 3 SC, i.e., intraregional) as well as those between time courses from different regions (V1 3 SC, i.e., interregional). The adjacent bar
graphs summarize these results, indicating higher correlations within a given region than between regions, due to different signal time courses in V1 and SC. (B and C) Average
time courses (±SD) of BOLD signal and MUA measured from V1 (black) and SC (gray) for 1-Hz flash stimulation. The time courses of BOLD and MUA from V1 show a prominent
onset peak, and a weaker but noticeable offset transient (arrows). The gray horizontal bar represents the stimulation period (30 s). For illustration of the corresponding LFP data,
see Supplementary Figure 3.
and 2.0 ± 1.1 s, respectively, with widths at half height of 1.5 ± Temporal Invariance of the IRF
0.4 s and 1.4 ± 0.3 s. These large variations of the peaks and We tested the temporal invariance of the IRFs estimated for the
widths of the IRFs are due to intersubject variations. Thus, we 1-Hz flash rate (Fig. 4) by applying them to data collected
conclude that there is no evidence supporting different during 5-Hz and 10-Hz flash rates. Thus, we convolved the MUA
neurovascular (BOLD-MUA) couplings in V1 and SC during 1- signals recorded during stimulation at these higher flash rates
Hz flash stimulation. We repeated the estimation procedure with the region-specific IRFs shown in Figure 4 (i.e., IRFV1 for
using the power in the 80--200 Hz band of the LFP signal as the V1 data and IRFSC for SC data) to provide comparisons of
neural input function (for equations (1) and (2)). The IRFs
modeled and measured BOLD responses (Fig. 5). Assumption of
generated for the ‘‘high-gamma’’ LFP and BOLD responses are in
linearity (and temporal invariance) holds if the modeled
close agreement to those obtained for BOLD and MUA
responses (data not shown). We also calculated IRFs for SC BOLD responses can be scaled to match the measured BOLD
and V1 at 5-Hz flash rate and found negligible differences from responses at all flash rates. Comparisons of MUA and BOLD
IRFs with 1-Hz flash rate (data not shown). However (linear) responses (modeled and measured single trials) are shown for
IRFs at 10-Hz flash rate are not defined due to the inherent 1-, 5-, and 10-Hz flash rates in Figures 5A,B,C, respectively. The
neurovascular uncouplings observed in V1 during the sustained predictions for SC are accurate within experimental noise for
response (MUA > BOLD) and in SC during the offset response all flash rates, whereas those for V1 concur with measurements
(BOLD > MUA) (see Table 1 and Fig. 3). only for 1- and 5-Hz flash rates, but not for flash rate of 10 Hz.
Table 1
Summary of neurovascular coupling and uncoupling observed during 10-Hz flash rate
(0--2 s) and offset (31--33 s) transients, a sustained component
during stimulation (0--30 s), and a poststimulus baseline shift
V1 SC (31--90 s). Each of these 4 components of the model was
Onset (~3 s in duration) Coupling observed Coupling observed convolved with a gamma variate-based canonical HRF using
Sustained ([25 s in duration) Coupling NOT observed Coupling observed parameters set to the average of IRFV1 and IRFSC (tp = 3.23 s, a =
Offset (~3 s in duration) Coupling observed Coupling NOT observed
11.33). Results for a single representative animal are shown in
Figure 7. An identical analysis was performed for 5 other data
Therefore, the IRFSC (in Fig. 4) is time invariant at all applied sets, and the group results are summarized in Table 2.
flash rates, whereas the IRFV1 (Fig. 4) is time invariant except at The left column of Figure 7 illustrates thresholded statistical
the 10-Hz flash rate. We repeated the analysis shown in Figure 5 parametric maps (left panel) for each of the 4 model
for single trials on the whole data set, and plotted the modeled components for 1- (Fig.7A), 5- (Fig. 7B), and 10-Hz (Fig. 7C)
versus measured BOLD responses in the period from –5 s to flash rates, respectively. The first observation is that the V1
+40 s. Consistent with the above observations, the data points transient and the 2 sustained components (when present) are
for both SC and V1 comparisons lie close to the line of identity, colocalized, that is, all 3 response components are generated by
except for the case of 10-Hz flash rate stimulus in V1, where the same region of V1 and presumably in the same neuronal
the prediction consistently overestimates the measured BOLD population from which the electrical signals (i.e., MUA and LFP)
signal (Supplementary Fig. 6A,B). were recorded. Compared with the sustained SC response, the
poststimulus signal level elevation observed in SC is more
Spatial Invariance of the IRF restricted to midline structures. The right column of Figure 7
demonstrates the GLM fits for each of the 4 model components
The demonstration of space-invariant neurovascular coupling is
for 1- (Fig. 7A), 5- (Fig. 7B), and 10-Hz (Fig. 7C) flash rates. Good
shown in Figure 6, in which IRFV1 is applied on the data
model fits (dark solid lines) are observed in both SC and V1 at all
obtained form the SC and IRFSC on the data obtained from V1.
stimulus rates. Table 2 lists the estimated component amplitudes
Assumption of linearity (and spatial invariance) holds if the
for both regions and all flash rates (data from 6 animals).
modeled BOLD responses can be scaled to match the measured
Each phase of the V1 response shows strong dependence on
BOLD responses from both regions. As can be seen in Figure
flash rate, whereas the SC response is generally less frequency
6A,B,C, respectively, for 1-, 5-, and 10-Hz flash rates, the effect
dependent. In V1, the onset and sustained response compo-
of interchanging the IRF of one region with that of the other is
nents diminish as a function of increasing frequency, whereas
minimal and within experimental error at all flash rates
the opposite is true for the offset transient. The sustained
(compare Fig. 5 with Fig. 6). We repeated the analysis on all
response attenuates to the extreme when the flash rate reaches
trials and plotted modeled BOLD responses against measured
10 Hz leading to potential misinterpretation of the underlying
BOLD responses (Supplementary Fig. 6C,D). We conclude that
neural responses. By contrast, the SC responses are dominated
IRFSC and IRFV1 (Fig. 4) are space invariant within experimental
by sustained signal at all frequencies, with smaller contribu-
error.
tions from the transients. A small onset transient that decreases
in amplitude with higher flash rates is observed in some SC
GLM of the BOLD Response traces. The offset transient in SC seems to be moderately
We constructed a GLM for the rat visual system that embodies increasing with flash rate, but unlike in V1, it does not reach in
the main features observed in V1 and SC BOLD responses: onset amplitude the signal level reached during sustained stimulation.
Page 6 of 13 Dynamic and Regional BOLD Impulse Response Functions in Rat Visual System d
Bailey et al.
Figure 4. Average IRFs for V1 and SC (i.e., IRFV1 and IRFSC represented by solid and dashed lines, respectively) obtained by deconvolution between single-trial BOLD and MUA
data measured during 1 Hz stimulation (data from 6 animals). See text for details on the gamma variate function. IRFV1 and IRFSC peak at 2.5 ± 0.5 s and 2.0 ± 1.1 s,
respectively, with widths at half height of 1.5 ± 0.4 s and 1.4 ± 0.3 s. The gray shading represents the SD of the estimated IRFs.
The poststimulus signal level elevation is only observed in SC at et al. 1996; Engel et al. 1997). GLM of the BOLD response relies
the higher flash rates. on this assumption and can only accommodate certain special
types of nonlinearities with uncertain physiological relevance
(Friston et al. 1998). Because interpreting fMRI data in an
Discussion acceptable manner requires that we know how stable neuro-
Neurovascular coupling is commonly modeled as a linear time- vascular coupling is for varying stimuli and across brain regions,
invariant system that is fully characterized by its IRF (Boynton using the rat visual system as a model, we evaluated temporal
and spatial variances of IRFs and assessed their implications, if 10-Hz flash rate for the offset response (i.e., where BOLD >
any, on GLM of BOLD responses. MUA). Both IRFV1 and IRFSC are space invariant within
experimental accuracy (Figs 4--6). We illustrate how the use
Summary of Present Findings of a suitably designed GLM will accurately represent the
We built a visual stimulus system to allow change of light color hemodynamics at all stimulus rates in both regions, but will
and/or flash rate of full-field binocular stimuli. Our recording lead to the wrong conclusion regarding the underlying neural
setups for fMRI (with multiple slices) and electrophysiology activity (Fig. 7).
(with multiple electrodes) enabled reproducible measure-
ments of evoked responses from multiple loci (i.e., V1 and
SC). Both in SC and V1, reliable and reproducible single-trial Comparison with Previous Imaging Studies of the Rat
responses were obtained with green and blue colors, whereas Visual System
red color induced only weaker responses. SC time courses Neuroimaging of the rat visual system has recently been
showed a rapid rise to peak and moderate decrease during performed by optical imaging and fMRI techniques. Gias et al.
sustained stimulation, whereas V1 exhibited onset/offset (2005) demonstrated retinotopic organization in upper layers
transients in addition to the sustained signal during the of rat V1 with optical imaging. To reach deeper cortical layers
stimulus. These salient features were depicted in both neural in the rat, van Camp et al. (2006) used 7.0T fMRI to image
and BOLD responses, albeit at different time scales. Overall, we functional responses in V1/V2, SC, and parts of the accessory
found that BOLD and neural responses of V1 and SC, which are optic system of pigmented Long-Evans rats under medetomi-
color sensitive, differ greatly in their temporal evolutions and dine anesthesia. Although the relatively weak BOLD responses
dependencies on flash rates (Figs 1--3). While IRFV1 and IRFSC necessitated long stimulus blocks (2 min) and low temporal
were found to be similar at the 1-Hz flash rate, we determined resolution (25 s), the most salient feature of their results was
that in the SC temporal invariance generally holds for the flash the opposite frequency dependence of the sustained V1 versus
rate range examined (1--10 Hz), whereas this assumption SC responses, a finding which the present study confirms.
breaks down significantly in V1 at 10 Hz mainly due to strong Furthermore, the authors point to the observation that at high
attenuation of the sustained BOLD component (i.e., where flash rates the SC responses exceed the stimulation period,
MUA > BOLD). However, there is a small uncoupling in SC at which is in line with our own observations.
Page 8 of 13 Dynamic and Regional BOLD Impulse Response Functions in Rat Visual System d
Bailey et al.
Downloaded from http://cercor.oxfordjournals.org/ at Yale University on February 3, 2012
Figure 7. Comparison between experimentally measured BOLD responses and modeled BOLD responses created from GLM. Time courses (right column) of measured BOLD
responses (pink shade) from V1 (top trace) and SC (bottom trace) for (A) 1-, (B) 5-, and (C) 10-Hz flash rates are juxtaposed with the BOLD activation maps (left column) for the 4
phases of the response: onset, sustained, offset, and poststimulation. The modeled BOLD responses were generated by summing the time courses generated from GLM of the 4
components, where the canonical HRF used for GLM was the average of IRFSC and IRFV1 (in Fig. 4), but very similar results are obtained if either of IRFSC and IRFV1 is used (data
not shown). The gray horizontal bar represents the stimulation period (30 s). Data from a single representative animal; for summary statistics for 6 animals, see Table 2.
poststimulation.
Table 2
Values of parameter estimates of the four phases GLM following 1-, 5-, and 10-Hz flash rate stimulation (n 5 6)
V1 (% ± SD) SC (% ± SD)
1 Hz 5 Hz 10 Hz 1 Hz 5 Hz 10 Hz
Onset 2.01 (±0.59) 0.95 (±0.33) 0.76 (±0.46) 0.59 (±0.70) 0.46 (±0.64) 0.18 (±0.42)
Sustained 0.79 (±0.19) 0.43 (±0.24) 0.05 (±0.47) 1.66 (±0.29) 1.80 (±0.40) 1.79 (±0.45)
Offset 0.69 (±0.29) 1.78 (±0.44) 2.25 (±0.78) 0.51 (±0.60) 0.77 (±0.52) 0.99 (±0.66)
Poststimulation 0.16 (±0.12) 0.01 (±0.20) 0.19 (±0.23) 0.07 (±0.29) 0.70 (±0.50) 1.06 (±0.36)
At 9.4T, Pawela et al. (2008) were able to considerably refractory nonlinearities), exhibits poorest performance in
improve the temporal resolution (2 s) and, presumably using visual cortex at the higher (5 and 10 Hz) flash rates. We
a much larger RF coil, covered extensively the posterior brain extend this finding by showing that measured neural activity
regions and thereby detected BOLD responses from DLG, V1/ becomes increasingly decoupled from the BOLD response as
V2, and SC in albino Sprague-Dawley rats under medetomidine a function of increasing flash rate, the greatest extent of which
anesthesia. Signal averaging was necessary across trials to reveal was observed at 10-Hz flash rate.
transient responses (i.e., at onset/offset of stimuli) in V1/V2 In a recent paper, Lau et al. report on BOLD modulations by
compared with SC/DLG, a finding which we are also able to short (1-s duration) blocks of 10-Hz flash rate stimulation in SC
corroborate. We note that the modeling approach taken by and thalamus of albino Sprague-Dawley rats under isoflurane
Pawela et al., which attempts to explain measured BOLD signals anesthesia (Lau et al. 2011). The IRFSC estimated in the present
as a function of assumed underlying neural dynamics (including work for 30 s blocks of 1-Hz flash rate stimulation predicts well
Page 10 of 13 Dynamic and Regional BOLD Impulse Response Functions in Rat Visual System d
Bailey et al.
2010). It is important to consider that neither intracortical modeling, as differences in experimental noise across the
measures of electrical network activity (in present study) nor regions have not been fully assessed. In the present study, we
extracranial recordings (in the Gutschalk study) can unambig- found that BOLD transfer functions for SC and V1 are
uously distinguish between membrane fluctuations of excitatory interchangeable. The implication for GLM of BOLD responses
versus inhibitory cells. Rather, the power recorded by a micro- is that the use of a single canonical HRF for whole-brain
electrode (in present study) or magnetic sensor (in the evaluations is likely to be accurate within experimental error,
Gutschalk study) is a weighted sum of many subpopulations of once temporal dynamics of the responses are sufficiently
neurons and possibly even glia. Thus, it remains to be shown accounted for by the model structure.
whether the weaker BOLD response in V1 at high flash rates is
due to a large metabolic demand or a small blood flow response.
Preliminary results using surface laser-Doppler flowmetry probes Conclusions
over V1 indicate that blood flow time courses closely mimic the Interpretation of fMRI data mandates that we know how stable
BOLD response at all flash rates examined (data now shown), neurovascular coupling is for varying stimuli and across brain
thereby suggesting that it might be blood flow that is uncoupled regions (Boynton et al. 1996; Engel et al. 1997). This is achieved
from neural activity during the sustained response at high flash by experimentally measuring IRFs that are calculated from
rates. While these preliminary results are not as comprehensive measured neural and BOLD responses. The underlying assump-
as our BOLD and electrophysiological data, we believe that with tion of linearity and time invariance should not be taken for
future improvements in fMRI sensitivity, we can conduct reliable
Page 12 of 13 Dynamic and Regional BOLD Impulse Response Functions in Rat Visual System d
Bailey et al.
Ohki K, Chung S, Ch’ng YH, Kara P, Reid RC. 2005. Functional imaging Sheth SA, Nemoto M, Guiou M, Walker M, Pouratian N, Toga AW. 2004.
with cellular resolution reveals precise micro-architecture in visual Linear and nonlinear relationships between neuronal activity,
cortex. Nature. 433:597--603. oxygen metabolism, and hemodynamic responses. Neuron.
Ostwald D, Porcaro C, Bagshaw AP. 2010. An information theoretic 42:347--355.
approach to EEG-fMRI integration of visually evoked responses. Smith AJ, Blumenfeld H, Behar KL, Rothman DL, Shulman RG, Hyder F.
Neuroimage. 49:498--516. 2002. Cerebral energetics and spiking frequency: the neurophysi-
Pawela CP, Hudetz AG, Ward BD, Schulte ML, Li R, Kao DS, Mauck MC, ological basis of fMRI. Proc Natl Acad Sci U S A. 99:10765--10770.
Cho YR, Neitz J, Hyde JS. 2008. Modeling of region-specific fMRI Stark E, Abeles M. 2007. Predicting movement from multiunit activity.
BOLD neurovascular response functions in rat brain reveals residual J Neurosci. 27:8387--8394.
differences that correlate with the differences in regional evoked Ueki M, Linn F, Hossmann KA. 1988. Functional activation of cerebral
potentials. Neuroimage. 41:525--534. blood flow and metabolism before and after global ischemia of rat
Paxinos G, Watson C. 1998. The rat brain in stereotaxic coordinates. San brain. J Cereb Blood Flow Metab. 8:486--494.
Diego (CA): Academic Press. Ueki M, Mies G, Hossmann KA. 1992. Effect of alpha-chloralose,
Prusky GT, Harker KT, Douglas RM, Whishaw IQ. 2002. Variation in halothane, pentobarbital and nitrous oxide anesthesia on metabolic
visual acuity within pigmented, and between pigmented and albino coupling in somatosensory cortex of rat. Acta Anaesthesiol Scand.
rat strains. Behav Brain Res. 136:339--348.
36:318--322.
Sanganahalli BG, Herman P, Blumenfeld H, Hyder F. 2009a. fMRI and
Ureshi M, Matsuura T, Kanno I. 2004. Stimulus frequency dependence
electrophysiological studies with a-chloralose and domitor anes-
of the linear relationship between local cerebral blood flow and
thesia. J Cereb Blood Flow Metab. 29:S616
field potential evoked by activation of rat somatosensory cortex.
Sanganahalli BG, Herman P, Blumenfeld H, Hyder F. 2010. BOLD
Neurosci Res. 48:147--153.