fmri image preprocessing - mcgill university · 2011-11-21 · fmri image preprocessing rick hoge,...
TRANSCRIPT
fMRI Image Preprocessing
Rick Hoge, Ph.D.Laboratoire de neuroimagerie vasculaire (LINeV)Centre de recherche de l’institut universitaire de
gériatrie de Montréal, Université de Montréal
Outline
• Motion correction
• Spatial filtering
• Distortion correction
• Physiological noise correction
Motion Correction
• serial realignment of all images to a target volume
• average over all volumes
• a single early or middle volume
• motion parameters can be used in subsequent temporal filtering
Image series with motion
Translations
0 100 200 300 400-2
-1
0
1
2
Time (s)
Signa
l (au
)
Signal
LR
AP
HF
Rotations
0 100 200 300 400-2
-1
0
1
2
Time (s)
Signa
l (au
)
Signal
pitch
roll
yaw
Realigned Series
MRM 31:283-291 (1994)
Stimulus-correlated motion
Artifactual “activation”
fMRI Bite Bar
Moana-Filho et al. BMC Neuroscience 2010
Software Support
• all major fMRI software packages provide motion correction
• FSL
• SPM
• AFNI
• etc...
Spatial Filtering
• random noise in fMRI data has a fairly high amplitude, comparable to functional changes we seek to detect
• averaging adjacent voxels can help increase the signal-to-noise ratio
• typically a 3D Gaussian smoothing kernel with width of around 5-6 mm is applied
Noise in fMRI data
2 mm in-plane resolution
Dependence of SNR on spatial resolution
4 mm in-plane resolution
2 mm
4 mm
Noise drives residual error in GLM
0 40 80 120 160 200 240 280
708
712
716
720
724
Time (s)
Signa
l (au
)
β: 6.038 5.333 714.469 -2.685 -1.052 0.092 (1.6%)
0 100 200 300 400 5000
200
400
600
800
1000
1200
Time (s)
Sig
nal (a
u)
β: 787.607 -0.671 -0.522 0.609
0 100 200 300 400 5000
200
400
600
800
1000
1200
Time (s)
Sig
nal (a
u)
β: 830.730 -1.029 0.699 0.726
2 mm
4 mm
Effect of spatial smoothing on physiological noise inhigh-resolution fMRI
Christina Triantafyllou, Richard D. Hoge, and Lawrence L. Wald*
MGH/MIT/HMS A.A. Martinos Center for Biomedical Imaging, Massachusetts General Hospital, Dept. of Radiology, Mailcode 2301, Bldg. 149, 13th Street,
Charlestown, MA 02129, USA
Received 6 December 2005; revised 14 March 2006; accepted 5 April 2006
Available online 30 June 2006
Physiological noise dominates the SNR of the fMRI time-course at
commonly used spatial resolutions at field strengths of 3 T and above.
Operating in this physiological noise dominated regime limits some
benefits of high field acquisition since increases in image SNR produce
only modest increases in time-course SNR. Although previous studies
have shown that the physiological noise dominance can be mitigated
by using higher spatial resolutions, not all functional studies require
voxel sizes smaller than the thickness of the human cortex. In this
study, we examine the effect of acquiring high spatial resolution,
thermal noise dominated time-courses and spatially smoothing the
images to lower resolutions, which would otherwise be physiological
noise dominated. At high field strengths, where physiological noise is
most problematic, this strategy lowered the overall time-course
variance compared to direct acquisition at commonly used spatial
resolution. At 7 T for example, 5 ! 5 ! 3 mm3 resolution images
derived from smoothing 1.5 ! 1.5 ! 3 mm3 data improved time-course
SNR by a factor of 1.89 compared to a time-series acquired at 5 ! 5 !3 mm3. Presumably, this effect was derived from the reduced
physiological-to-thermal noise ratio in the high spatial resolution data
followed by a smoothing operation that improves SNR without adding
physiological noise. Our findings demonstrate that in contrast to
conventional SNR penalties associated with spatially smoothing Four-
ier data, the time-course SNR of smoothed high-resolution data can be
improved compared to direct acquisition at the desired resolution.
D 2006 Elsevier Inc. All rights reserved.
Keywords: Physiological noise; fMRI; High field; Spatial resolution; SNR;
Magnetic field strength; 7 Tesla
Introduction
Sensitivity and specificity of BOLD fMRI maps are expected toimprove with increasing field strength due to favorable increases in
BOLD contrast, BOLD localization and improved image sensitiv-ity. The increase in image sensitivity with higher field, however, is
not fully utilized at conventional spatial resolutions since the fMRItime-course noise is dominated by physiological noise (Krueger
and Glover, 2001; Triantafyllou et al., 2005). Since the physio-logical fluctuations appear as a multiplicative modulation of theimage signal (Krueger and Glover, 2001; Krueger et al., 2001),
their amplitude scales with the MR image intensity. In contrast, thethermal noise sources can be represented by the addition of a fixedamount of Gaussian noise power whose amplitude is determined
primarily by the coil loading. Thus, as higher image sensitivity isachieved through the use of better RF coils, stronger magnets orlower spatial resolution, the physiological noise increases propor-
tionally. When the time-course noise is dominated by thesephysiological effects, as it is for high field fMRI at commonlyused image resolutions (27 mm3), increases in image SNR result inonly marginal increases in the SNR of the fMRI time-course
(tSNR) (Krueger and Glover, 2001; Triantafyllou et al., 2005).Our previous work showed that by acquiring the functional
images at higher spatial resolutions, it is possible to operate in a
regime where thermal image noise dominates physiological noiseeven for 7 T fMRI (Triantafyllou et al., 2005). In this regime,increased image SNR (e.g., from 7 T acquisition) still translates
into improved time-course SNR, while the tSNR of lowerresolution acquisitions only marginally benefited from theincreased sensitivity of 7 T. While high spatial resolutionfunctional maps are intrinsically desirable for some studies, it is
not clear that voxel dimensions smaller than the cortical ribbonthickness are beneficial for applications where large areas of cortexare activated. We propose a method of allowing lower spatial
resolution studies to avoid some of the physiological noise penaltyby acquiring the data at high spatial resolution where thermal noisedominates, and spatially smoothing the data to the desired (lower)
spatial resolution.In this study, we examine the effect of spatially smoothing high
spatial resolution EPI images down to conventional resolutions.
The goal is to obtain the best of both worlds; thermal noise-dominated images at acquisition whose tSNR benefits from thesensitivity of high field, and further increases in image andtemporal SNR from the spatial smoothing operation. We hypoth-
esize that the time-course SNR from images with spatially
1053-8119/$ - see front matter D 2006 Elsevier Inc. All rights reserved.
doi:10.1016/j.neuroimage.2006.04.182
* Corresponding author. Fax: +1 617 726 7422.
E-mail address: [email protected] (L.L. Wald).
Available online on ScienceDirect (www.sciencedirect.com).
www.elsevier.com/locate/ynimg
NeuroImage 32 (2006) 551 – 557
Image noise vs. temporal noise
Image Temporal
0 40 80 120 160 200 240 280
708
712
716
720
724
Time (s)
Signa
l (au
)
β: 6.038 5.333 714.469 -2.685 -1.052 0.092 (1.6%)
are best described by 1.33, 1.25 and 1.13 mm FWHM
Gaussians at 1.5 T, 3 T and 7 T, respectively. This sinc-to-Gaussian conversion factor was then used in all subsequentanalysis.
Fig. 2 shows the time series SNR of the smoothed data as afunction of the smoothing parameter V for each of the 3 fieldstrengths as well as the theoretical values expected from Eq. (4)
(spatially uncorrelated rp) and Eq. (7) (spatially correlated rp).The measured tSNR increase on smoothing lies between the two
theoretical cases for all field strengths, consistent with partialspatial correlation in the physiological noise. Fig. 3 illustrates
the dependence of the time series SNR on the image SNR fordifferent spatial resolutions where the spatial resolution resultedfrom either smoothing (circles) or from direct acquisition with
lower spatial resolution (squares). As shown in our previousstudy (Triantafyllou et al., 2005), the tSNR approaches anasymptote for the data acquired at progressively courser spatialresolution. For the time series obtained by smoothing the 1.5-
Fig. 2. Time series SNR (tSNRsmooth) in gray matter as a function of
degree of smoothing for 1.5 T (A), 3 T (B) and 7 T (C). The top dashed
curves show the expected theoretical relationship assuming spatially
uncorrelated physiological noise Eq. (4), whereas the lower curves
represent Eq. (7) (which assumes completely spatially correlated
physiological noise).
Fig. 1. Image SNR in gray and white matter ROIs as a function degree of
smoothing for 1.5 T (A), 3 T (B) and 7 T (C). The voxel volume fraction is
defined as the volume of the smoothed data’s voxel divided by the volume
of the original high-resolution data set from which the smoothed images are
derived. The dashed curves show the expected theoretical relationship
SNR0smooth !!!!!Vp
SNR0highres.
C. Triantafyllou et al. / NeuroImage 32 (2006) 551–557554
are best described by 1.33, 1.25 and 1.13 mm FWHM
Gaussians at 1.5 T, 3 T and 7 T, respectively. This sinc-to-Gaussian conversion factor was then used in all subsequentanalysis.
Fig. 2 shows the time series SNR of the smoothed data as afunction of the smoothing parameter V for each of the 3 fieldstrengths as well as the theoretical values expected from Eq. (4)
(spatially uncorrelated rp) and Eq. (7) (spatially correlated rp).The measured tSNR increase on smoothing lies between the two
theoretical cases for all field strengths, consistent with partialspatial correlation in the physiological noise. Fig. 3 illustrates
the dependence of the time series SNR on the image SNR fordifferent spatial resolutions where the spatial resolution resultedfrom either smoothing (circles) or from direct acquisition with
lower spatial resolution (squares). As shown in our previousstudy (Triantafyllou et al., 2005), the tSNR approaches anasymptote for the data acquired at progressively courser spatialresolution. For the time series obtained by smoothing the 1.5-
Fig. 2. Time series SNR (tSNRsmooth) in gray matter as a function of
degree of smoothing for 1.5 T (A), 3 T (B) and 7 T (C). The top dashed
curves show the expected theoretical relationship assuming spatially
uncorrelated physiological noise Eq. (4), whereas the lower curves
represent Eq. (7) (which assumes completely spatially correlated
physiological noise).
Fig. 1. Image SNR in gray and white matter ROIs as a function degree of
smoothing for 1.5 T (A), 3 T (B) and 7 T (C). The voxel volume fraction is
defined as the volume of the smoothed data’s voxel divided by the volume
of the original high-resolution data set from which the smoothed images are
derived. The dashed curves show the expected theoretical relationship
SNR0smooth !!!!!Vp
SNR0highres.
C. Triantafyllou et al. / NeuroImage 32 (2006) 551–557554
are best described by 1.33, 1.25 and 1.13 mm FWHM
Gaussians at 1.5 T, 3 T and 7 T, respectively. This sinc-to-Gaussian conversion factor was then used in all subsequentanalysis.
Fig. 2 shows the time series SNR of the smoothed data as afunction of the smoothing parameter V for each of the 3 fieldstrengths as well as the theoretical values expected from Eq. (4)
(spatially uncorrelated rp) and Eq. (7) (spatially correlated rp).The measured tSNR increase on smoothing lies between the two
theoretical cases for all field strengths, consistent with partialspatial correlation in the physiological noise. Fig. 3 illustrates
the dependence of the time series SNR on the image SNR fordifferent spatial resolutions where the spatial resolution resultedfrom either smoothing (circles) or from direct acquisition with
lower spatial resolution (squares). As shown in our previousstudy (Triantafyllou et al., 2005), the tSNR approaches anasymptote for the data acquired at progressively courser spatialresolution. For the time series obtained by smoothing the 1.5-
Fig. 2. Time series SNR (tSNRsmooth) in gray matter as a function of
degree of smoothing for 1.5 T (A), 3 T (B) and 7 T (C). The top dashed
curves show the expected theoretical relationship assuming spatially
uncorrelated physiological noise Eq. (4), whereas the lower curves
represent Eq. (7) (which assumes completely spatially correlated
physiological noise).
Fig. 1. Image SNR in gray and white matter ROIs as a function degree of
smoothing for 1.5 T (A), 3 T (B) and 7 T (C). The voxel volume fraction is
defined as the volume of the smoothed data’s voxel divided by the volume
of the original high-resolution data set from which the smoothed images are
derived. The dashed curves show the expected theoretical relationship
SNR0smooth !!!!!Vp
SNR0highres.
C. Triantafyllou et al. / NeuroImage 32 (2006) 551–557554
1.5 Tesla 3 Tesla 7 Tesla
Image SNR and voxel volume
Temporal SNR and voxel volume
are best described by 1.33, 1.25 and 1.13 mm FWHM
Gaussians at 1.5 T, 3 T and 7 T, respectively. This sinc-to-Gaussian conversion factor was then used in all subsequentanalysis.
Fig. 2 shows the time series SNR of the smoothed data as afunction of the smoothing parameter V for each of the 3 fieldstrengths as well as the theoretical values expected from Eq. (4)
(spatially uncorrelated rp) and Eq. (7) (spatially correlated rp).The measured tSNR increase on smoothing lies between the two
theoretical cases for all field strengths, consistent with partialspatial correlation in the physiological noise. Fig. 3 illustrates
the dependence of the time series SNR on the image SNR fordifferent spatial resolutions where the spatial resolution resultedfrom either smoothing (circles) or from direct acquisition with
lower spatial resolution (squares). As shown in our previousstudy (Triantafyllou et al., 2005), the tSNR approaches anasymptote for the data acquired at progressively courser spatialresolution. For the time series obtained by smoothing the 1.5-
Fig. 2. Time series SNR (tSNRsmooth) in gray matter as a function of
degree of smoothing for 1.5 T (A), 3 T (B) and 7 T (C). The top dashed
curves show the expected theoretical relationship assuming spatially
uncorrelated physiological noise Eq. (4), whereas the lower curves
represent Eq. (7) (which assumes completely spatially correlated
physiological noise).
Fig. 1. Image SNR in gray and white matter ROIs as a function degree of
smoothing for 1.5 T (A), 3 T (B) and 7 T (C). The voxel volume fraction is
defined as the volume of the smoothed data’s voxel divided by the volume
of the original high-resolution data set from which the smoothed images are
derived. The dashed curves show the expected theoretical relationship
SNR0smooth !!!!!Vp
SNR0highres.
C. Triantafyllou et al. / NeuroImage 32 (2006) 551–557554
are best described by 1.33, 1.25 and 1.13 mm FWHM
Gaussians at 1.5 T, 3 T and 7 T, respectively. This sinc-to-Gaussian conversion factor was then used in all subsequentanalysis.
Fig. 2 shows the time series SNR of the smoothed data as afunction of the smoothing parameter V for each of the 3 fieldstrengths as well as the theoretical values expected from Eq. (4)
(spatially uncorrelated rp) and Eq. (7) (spatially correlated rp).The measured tSNR increase on smoothing lies between the two
theoretical cases for all field strengths, consistent with partialspatial correlation in the physiological noise. Fig. 3 illustrates
the dependence of the time series SNR on the image SNR fordifferent spatial resolutions where the spatial resolution resultedfrom either smoothing (circles) or from direct acquisition with
lower spatial resolution (squares). As shown in our previousstudy (Triantafyllou et al., 2005), the tSNR approaches anasymptote for the data acquired at progressively courser spatialresolution. For the time series obtained by smoothing the 1.5-
Fig. 2. Time series SNR (tSNRsmooth) in gray matter as a function of
degree of smoothing for 1.5 T (A), 3 T (B) and 7 T (C). The top dashed
curves show the expected theoretical relationship assuming spatially
uncorrelated physiological noise Eq. (4), whereas the lower curves
represent Eq. (7) (which assumes completely spatially correlated
physiological noise).
Fig. 1. Image SNR in gray and white matter ROIs as a function degree of
smoothing for 1.5 T (A), 3 T (B) and 7 T (C). The voxel volume fraction is
defined as the volume of the smoothed data’s voxel divided by the volume
of the original high-resolution data set from which the smoothed images are
derived. The dashed curves show the expected theoretical relationship
SNR0smooth !!!!!Vp
SNR0highres.
C. Triantafyllou et al. / NeuroImage 32 (2006) 551–557554
are best described by 1.33, 1.25 and 1.13 mm FWHM
Gaussians at 1.5 T, 3 T and 7 T, respectively. This sinc-to-Gaussian conversion factor was then used in all subsequentanalysis.
Fig. 2 shows the time series SNR of the smoothed data as afunction of the smoothing parameter V for each of the 3 fieldstrengths as well as the theoretical values expected from Eq. (4)
(spatially uncorrelated rp) and Eq. (7) (spatially correlated rp).The measured tSNR increase on smoothing lies between the two
theoretical cases for all field strengths, consistent with partialspatial correlation in the physiological noise. Fig. 3 illustrates
the dependence of the time series SNR on the image SNR fordifferent spatial resolutions where the spatial resolution resultedfrom either smoothing (circles) or from direct acquisition with
lower spatial resolution (squares). As shown in our previousstudy (Triantafyllou et al., 2005), the tSNR approaches anasymptote for the data acquired at progressively courser spatialresolution. For the time series obtained by smoothing the 1.5-
Fig. 2. Time series SNR (tSNRsmooth) in gray matter as a function of
degree of smoothing for 1.5 T (A), 3 T (B) and 7 T (C). The top dashed
curves show the expected theoretical relationship assuming spatially
uncorrelated physiological noise Eq. (4), whereas the lower curves
represent Eq. (7) (which assumes completely spatially correlated
physiological noise).
Fig. 1. Image SNR in gray and white matter ROIs as a function degree of
smoothing for 1.5 T (A), 3 T (B) and 7 T (C). The voxel volume fraction is
defined as the volume of the smoothed data’s voxel divided by the volume
of the original high-resolution data set from which the smoothed images are
derived. The dashed curves show the expected theoretical relationship
SNR0smooth !!!!!Vp
SNR0highres.
C. Triantafyllou et al. / NeuroImage 32 (2006) 551–557554
1.5 Tesla 3 Tesla 7 Tesla
Temporal Filtering
• typically carried out as part of statistical modelling
• low frequency drift
• residual motion effects
• physiological noise
Temporal filtering example
response terms
motion terms
drift terms
Physiological Noise
• motion
• cardiac pulsation
• respiratory movement
Image-Based Method for Retrospective Correction ofPhysiological Motion Effects in fMRI: RETROICOR
Gary H. Glover,1* Tie-Qiang Li,1 and David Ress2
Respiration effects and cardiac pulsatility can induce signalmodulations in functional MR image time series that increasenoise and degrade the statistical significance of activation sig-nals. A simple image-based correction method is describedthat does not have the limitations of k-space methods thatpreclude high spatial frequency correction. Low-order Fourierseries are fit to the image data based on time of each imageacquisition relative to the phase of the cardiac and respiratorycycles, monitored using a photoplethysmograph and pneu-matic belt, respectively. The RETROICOR method is demon-strated using resting-state experiments on three subjects andcompared with the k-space method. The method is found toperform well for both respiration- and cardiac-induced noisewithout imposing spatial filtering on the correction. Magn Re-son Med 44:162–167, 2000. © 2000 Wiley-Liss, Inc.Key words: functional magnetic resonance imaging; physiolog-ical motion; retrospective motion correction
Functional Magnetic Resonance Imaging (fMRI) is basedon changes in the signal resulting from blood oxygenationlevel dependence (BOLD) contrast (1,2) or blood flowchanges (2). Pulsatility of blood flow in the brain andrespiration-induced magnetic field changes or motion cancause appreciable modulation of the signal (3,4), which inturn causes undesired perturbations in the images thatinclude intensity fluctuations and other artifacts. In manycases the added noise induced by these physiological pro-cesses can be comparable to the desired signal, whichdegrades the statistical significance of activation signals orotherwise compromises event-related analyses. Dagli et al.(5) have found that the effects of cardiac function tend tobe rather localized in the brain as a result of vessel-depen-dent brain pulsatility. Respiration effects, which originatefrom thoracic modulation of the magnetic field in the heador from bulk motions of the head, are often more spatiallyglobal (3). However, we show here that respiration effectscan also be localized.
Several methods have been developed for reducing suchphysiological noise in fMRI time series. Navigator meth-ods correct k-space data using either an auxiliary echo (6)or the scan data themselves (7,8). Although navigatormethods can be effective, they sample either a projectionof the brain or the entire slice and therefore lack specificityin localizing the source of motion, which in turn can cause
incomplete correction or can introduce fluctuations in qui-escent parts of the brain. Biswal et al. (9) introduced theuse of notch filters to remove components of the time-series spectrum at the cardiac and respiratory frequencies;this method fails, however, if the noise spectra alias intothe region of the task spectrum. A retrospective correctionmethod was introduced by Hu et al. (10) that fits a low-order Fourier series to the k-space time-series data basedon the phase of the respiratory or cardiac cycle during eachacquisition. This method was found to work particularlywell for correcting respiratory effects and has been usedfor event-related experiments such as observation of theprompt response (11). However, only the data near thek-space origin have adequate signal-to-noise ratio (SNR)for the Fourier fit to be robust, and thus only low-orderspatial corrections can be made; this, in turn, introducescorrelation between pixels in the correction. If only afraction of the k-space values needed to represent a local-ized region such as the area surrounding a vessel arecorrected, the noise will be partially reduced only in thevessel region and spurious modulation will be introducedinto other regions of the brain.
In this work we describe a retrospective correction tech-nique similar to the method of Hu et al. (10), but operatingin the image domain (dubbed RETROICOR). Image-basedcorrection provides advantages in that spatial-frequencyfiltering is not imposed on the correction as with thek-space method, and therefore the correction functionsequally well for both global and localized cardiac andrespiratory noise. The concept was first introduced byJosephs et al. (12), who used SPM to generate maps of thephysiological motion components. This approach was em-ployed to diminish physiological noise in fMRI by model-ing the electrocardiogram (ECG) and respiratory-phasewaveforms as confounding signals (13). In many cases,however, correcting the data apart from performing a sta-tistical analysis is desired, and thus the independentRETROICOR method was developed. The techniqueis demonstrated with resting-state brain data and com-pared with the k-space method of Hu et al. (10), calledRETROKCOR for this work; their software implementa-tion, which can perform corrections in k-space or imagespace, is available from http://www.cmrr.umn.edu/soft-ware/physioFix_userGuide.html.
MATERIALS AND METHODS
RETROICOR
The correction method assumes that the time series ofintensities y(t) in a pixel is corrupted by additive noiseresulting from cardiac and respiratory functions. The car-diac and respiratory states are monitored during the scanusing a photoplethysmograph and a pneumatic belt placed
1Department of Radiology, Stanford University School of Medicine, Center forAdvanced MR Technology at Stanford, Stanford, California.2Department of Psychology, Stanford University, Stanford, California.Grant sponsor: National Institutes of Health; Grant number: P41 RR 09784;Grant sponsors: The Lucas Foundation; GE Medical Systems.Tie-Qiang Li’s current affiliation is Dept. of Psychology, Princeton University,Princeton, NJ 08544.*Correspondence to: Gary H. Glover, Stanford University School of Medicine,Department of Diagnostic Radiology, Lucas MR Center, Stanford, CA 94305-5488. E-mail: [email protected] 22 December 1999; revised 3 March 2000; accepted 6 March 2000.
Magnetic Resonance in Medicine 44:162–167 (2000)
© 2000 Wiley-Liss, Inc. 162
istics. Cardiac pulsatility is often localized to edges of thebrain such as near sulci or in tissue regions close to vesselssuch as the superior sagittal sinus. Some respiration-in-duced fluctuations result from longer-range effects such assmall bulk movement of the head or magnetic field mod-ulations from the changing state of the thoracic cavity.Image noise from this respiration component thereforetends to span the entire brain. In this case, noise associatedwith respiratory function occupies a smaller extent ink-space than circulatory-induced noise. However, asshown in Fig. 4, many regions of the brain have localizedmotion components tied to the respiratory cycle, perhaps
through brainstem motion. These effects are localized in afashion similar to that of cardiac motion and thus occupya similarly broad extent in k-space.
Retrospective correction methods that operate ink-space are limited to those spatial frequencies for whichthe SNR is adequate to ensure a good fit of the Fourierseries to the data. This region includes only componentsclose to the k-space origin, so that correlations in imagespace are introduced by the correction. This is not harmfulfor global respiration noise because of its low spatial fre-quency distribution, but can be detrimental for cardiac-induced noise or localized respiratory noise, since there
FIG. 3. RETROICOR method applied to ROI time-series data acquired at TR ! 250 msec. (a) Raw data (") and y# cardiac fit (*) plotted vs.phase in cardiac cycle; (b) same data plotted vs. phase of respiratory cycle (") and corresponding respiratory y# fit (*). Only one-fourth ofthe 750 data points are plotted for clarity. Spectra of time series (c) without correction; (d) with cardiac correction alone; (e) with respiratorycorrection alone; (f) with both corrections. In this case the cardiac and respiratory spectra are resolved with peaks near 0.8 and 0.15 Hz,respectively.
FIG. 4. Left: Maps of noise distributions for image data acquired at TR ! 250 msec corresponding to Fig. 3, showing (top) cardiaccomponents and (bottom) respiratory components. The three columns depict maps that are uncorrected, corrected with RETROKCOR, andcorrected with RETROICOR, respectively. In this case the cardiac-related noise is highly localized, whereas the respiratory noise is morediffuse but shows some focal noise foci medially. Right: Localizer showing slice location, and T*2-weighted image.
166 Glover et al.
TR = 250 ms
Physiological noise in short-TR acquisition
both correction methods. As may be seen, the image-basedmethod was effective in reducing the noise to a greaterextent than the k-space method for all subjects, althoughthere are insufficient data for statistical significance of thedifferences. However, the ROIs were deliberately chosenas worst-case; in most brain regions the residual corrected
physiological noise is essentially unmeasurable over thebackground, as shown in Fig. 4.
DISCUSSIONThe noise induced in fMRI time series by cardiac andrespiratory functions can have different spatial character-
FIG. 1. RETROICOR method applied to ROI time-series data acquired at TR ! 1000 msec. (a) Raw data (") and y# cardiac fit (*) plottedvs. phase in cardiac cycle; (b) same data plotted vs. phase of respiratory cycle (") and corresponding respiratory y# fit (*). Spectra of timeseries (c) without correction; (d) with cardiac correction alone; (e) with respiratory correction alone; (f) with both corrections. In this case thecardiac fluctuation spectrum (strong peak near 0.2 Hz) was aliased (heart rate 47 bpm) and partially overlapped with the respiratoryspectrum (0.1–0.15 Hz).
FIG. 2. Time series without (top) and with (bottom) RETROICOR correction corresponding to Fig. 1. Ordinate values are expressed aspercentage of mean values.
Retrospective Motion Correction 165
TR = 1 s
Physiological noise in long-TR acquisition
both correction methods. As may be seen, the image-basedmethod was effective in reducing the noise to a greaterextent than the k-space method for all subjects, althoughthere are insufficient data for statistical significance of thedifferences. However, the ROIs were deliberately chosenas worst-case; in most brain regions the residual corrected
physiological noise is essentially unmeasurable over thebackground, as shown in Fig. 4.
DISCUSSIONThe noise induced in fMRI time series by cardiac andrespiratory functions can have different spatial character-
FIG. 1. RETROICOR method applied to ROI time-series data acquired at TR ! 1000 msec. (a) Raw data (") and y# cardiac fit (*) plottedvs. phase in cardiac cycle; (b) same data plotted vs. phase of respiratory cycle (") and corresponding respiratory y# fit (*). Spectra of timeseries (c) without correction; (d) with cardiac correction alone; (e) with respiratory correction alone; (f) with both corrections. In this case thecardiac fluctuation spectrum (strong peak near 0.2 Hz) was aliased (heart rate 47 bpm) and partially overlapped with the respiratoryspectrum (0.1–0.15 Hz).
FIG. 2. Time series without (top) and with (bottom) RETROICOR correction corresponding to Fig. 1. Ordinate values are expressed aspercentage of mean values.
Retrospective Motion Correction 165
TR = 1 s
istics. Cardiac pulsatility is often localized to edges of thebrain such as near sulci or in tissue regions close to vesselssuch as the superior sagittal sinus. Some respiration-in-duced fluctuations result from longer-range effects such assmall bulk movement of the head or magnetic field mod-ulations from the changing state of the thoracic cavity.Image noise from this respiration component thereforetends to span the entire brain. In this case, noise associatedwith respiratory function occupies a smaller extent ink-space than circulatory-induced noise. However, asshown in Fig. 4, many regions of the brain have localizedmotion components tied to the respiratory cycle, perhaps
through brainstem motion. These effects are localized in afashion similar to that of cardiac motion and thus occupya similarly broad extent in k-space.
Retrospective correction methods that operate ink-space are limited to those spatial frequencies for whichthe SNR is adequate to ensure a good fit of the Fourierseries to the data. This region includes only componentsclose to the k-space origin, so that correlations in imagespace are introduced by the correction. This is not harmfulfor global respiration noise because of its low spatial fre-quency distribution, but can be detrimental for cardiac-induced noise or localized respiratory noise, since there
FIG. 3. RETROICOR method applied to ROI time-series data acquired at TR ! 250 msec. (a) Raw data (") and y# cardiac fit (*) plotted vs.phase in cardiac cycle; (b) same data plotted vs. phase of respiratory cycle (") and corresponding respiratory y# fit (*). Only one-fourth ofthe 750 data points are plotted for clarity. Spectra of time series (c) without correction; (d) with cardiac correction alone; (e) with respiratorycorrection alone; (f) with both corrections. In this case the cardiac and respiratory spectra are resolved with peaks near 0.8 and 0.15 Hz,respectively.
FIG. 4. Left: Maps of noise distributions for image data acquired at TR ! 250 msec corresponding to Fig. 3, showing (top) cardiaccomponents and (bottom) respiratory components. The three columns depict maps that are uncorrected, corrected with RETROKCOR, andcorrected with RETROICOR, respectively. In this case the cardiac-related noise is highly localized, whereas the respiratory noise is morediffuse but shows some focal noise foci medially. Right: Localizer showing slice location, and T*2-weighted image.
166 Glover et al.
Raw K-Spacecorrection
Image-Spacecorrection
Reduction of residual error through physiological noise “correction”
Physiological noise modelling for spinal functional magneticresonance imaging studies
Jonathan C.W. Brooks,a,⁎ Christian F. Beckmann,e Karla L. Miller,b Richard G. Wise,c
Carlo A. Porro,d Irene Tracey,a,b and Mark Jenkinsonb
aPaIN Group, Department of Physiology Anatomy and Genetics, Le Gros Clark Building, South Parks Road, Oxford OX1 3QX, UKbCentre for Functional Magnetic Resonance Imaging of the Brain, Department of Clinical Neurology, John Radcliffe Hospital, UKcCUBRIC, School of Psychology, Cardiff University, UKdDepartment of Biomedical Sciences, University of Modena and Reggio Emilia, Modena, ItalyeDivision of Neuroscience and Mental Health, Imperial College London, Hammersmith Campus, Du Cane Road, London, UK
Received 18 January 2007; revised 6 September 2007; accepted 10 September 2007Available online 20 September 2007
Spinal cord functional imaging allows assessment of activity in primarysynaptic connections made by sensory neurons relaying informationabout the state of the body. However, reported human data based ongradient-echo techniques have been largely inconsistent, with no clearpatterns of activation emerging. One reason for this variability is theinfluence of physiological noise, which is typically not corrected for. Byacquiring single-slice resting data from the spinal cord with aconventional gradient-echo EPI pulse sequence at TR=200 ms(critically sampled) and TR=3 s (under-sampled), we have char-acterised various sources of physiological noise. In 8 healthy subjects,the presence of physiologically dependent signal was explored usingprobabilistic independent component analysis (PICA). Based on theinsights provided by PICA, we defined a new physiological noise model(PNM) based on retrospective image correction (RETROICOR),which uses independent physiological measurements taken from thesubject to model sources of noise. Statistical significance of individualcomponents included in the PNM was assessed by F-tests, whichdemonstrated that the optimal PNM included cardiac, respiratory,interaction and low-frequency regressors. In a group of 10 healthysubjects, activation data were acquired from the cervical spinal region(T1 to C5) during painful thermal stimulation of the right and lefthands. The improvement obtained when using a PNM in estimatingspinal cord activation was reflected in a reduction of false-positiveactivation (active voxels in the CSF space surrounding the cord), whencompared to conventional GLM modelling without a PNM.© 2007 Elsevier Inc. All rights reserved.
Introduction
The spinal cord is the first site at which sensory informationreceived from the body is processed (Willis and Coggeshall, 2004).Considering the pain system, primary afferent fibres will typicallymake their first synaptic connection in the dorsal portion of thespinal cord, the “dorsal horn”. Neurons within the dorsal horn arearranged in layers (“laminae”), which broadly correspond to thedifferent types of afferent fibre entering the cord (Willis andCoggeshall, 2004). These connections are subject to top–down andlocal neuronal influences that will modulate the transmission ofnociceptive signals to the brain, and thus have a great impact onpain sensation (Suzuki et al., 2004; Tracey et al., 2002). Indeed,neuropathic pain (Hansson et al., 2001) is thought to arise, at leastin part, due to dysfunction (e.g. rearrangement, disinhibition, cellloss) at the primary synaptic connection in the dorsal horn (Woolfand Salter, 2000). Pain-related activity of rat spinal neurons hasbeen assessed using autoradiography (Coghill et al., 1991; Porroand Cavazzuti, 1993), and animal spinal fMRI studies arebeginning to be performed (Lawrence et al., 2004; Lilja et al.,2006; Majcher et al., 2006; Malisza and Stroman, 2002; Porszaszet al., 1997). There is a clear need for the development oftechniques to record human spinal cord function non-invasively.The close connection between vasculature and neurons within thespinal cord (Giove et al., 2004), should make this structureamenable to the same techniques used to study the brain: e.g. usingblood oxygenation level-dependent (BOLD) contrast to inferneuronal activation (Ogawa et al., 1990).
The main difficulties facing researchers attempting to image thespinal cord are its small cross-sectional area, its proximity tostructures (vertebrae, intervertebral discs) giving rise to largevariations in magnetic susceptibility, the influence of physiologicaleffects that displace the cord or generate signal intensity changes
www.elsevier.com/locate/ynimgNeuroImage 39 (2008) 680–692
⁎ Corresponding author. Fax: +44 1865 222717.E-mail address: [email protected] (J.C.W. Brooks).Available online on ScienceDirect (www.sciencedirect.com).
1053-8119/$ - see front matter © 2007 Elsevier Inc. All rights reserved.doi:10.1016/j.neuroimage.2007.09.018
was seen in the signal component located in the CSF surrounding thespinal cord (at the cardiac frequency). Respiratory signal wasobserved in the critically sampled data, with frequency !0.33 Hz,and was located primarily in the connective tissue (e.g. neckmuscles), andmay reflect movement or susceptibility induced signalchange. Interestingly, an interaction between cardiac and respiratorysignals was also apparent (Fig. 2(C)) in the TR=200 ms data: thecharacteristic “fingerprint” of amplitude modulation (AM) wasvisible, with both respiratory and cardiac frequencies present in thepower spectrum, plus AM sidebands located either side of thecardiac (carrier) frequency, separated by ±0.33 Hz (i.e. therespiratory frequency). Another component was frequently ob-served across the group, and was at very low frequency with thecorresponding spectrum containing power below 0.1 Hz.
DiscussionThe application of exploratory data analysis allowed visual
identification of physiologically driven sources of structured noise.The technique typically identified 16 to 22 independent compo-nents (in critically sampled data), which by visual inspection couldbe assigned to 4 groups: cardiac, respiratory, interaction and lowfrequency. The detection of cardiac and respiratory components inspinal EPI data via exploratory analysis methods, confirmsprevious findings by Friese et al. (2004b), but also demonstrated
an interaction between these two signals. In this study, thisinteraction manifested itself as amplitude modulation (AM) of thecardiac waveform, and characteristic AM sidebands in the powerspectrum of the measured signal (see Fig. 2(C)). As expected, thisinteraction effect was localised in the CSF space surrounding thecord, but was not observed for all subjects. The source of thisinteraction is thought to be due to the effect of intrathoracicpressure on venous return to the heart and consequently the cardiacstroke volume (Frank et al., 2001; Lin, 1999), which would havethe effect of increasing/decreasing CSF flow according to positionwithin the respiratory cycle. Low-frequency signals componentswere also identified in the MELODIC output (e.g. Fig 2(D)). Low-frequency components in CSF flow time courses have previouslybeen reported (Friese et al., 2004a) with characteristic frequenciesin the range 0.01 to 0.1 Hz. These so-called B- or C-waves arethought to stem from variations in parasympathetic and sympa-thetic excitation of the heart which are translated to variations inneurovascular blood pressure and hence to the CSF (Auer andSayama, 1983). These observations suggest that a conventionalRETROICOR approach to modelling physiological noise would beinadequate to explain all sources of structured noise in/around thespinal cord, particularly those effects which, when using theexploratory approach, are identified as interaction effects. We willdemonstrate henceforth that a purely model-based approach to
Fig. 2. Main ICA-derived signal components (and associated power spectra) detected with critically sampled (TR=200 ms) spinal cord EPI data: (A) cardiac, (B)respiratory, (C) interaction of cardiac and respiratory, and (D) low (b0.1 Hz) frequency. The time courses and spatial distributions are estimated from resting-fMRI data, and reveal clear cardiac related (!1 Hz) signal around the major vessels (carotids and vertebral arteries) plus cardiac driven CSF flow effects in thesubarachnoid space surrounding the spinal cord. Respiratory effects (0.33 Hz) were found in the muscle of the neck, but were also found to interact with cardiaceffects, and were distributed around the spinal cord. Low-frequency components, due to either uncorrected motion or B-/C-waves were also detected. For clarityonly the first 120 s of each time course are shown.
683J.C.W. Brooks et al. / NeuroImage 39 (2008) 680–692
TR = 250 ms
Image distortion and dropout
• Microscopic:
• deoxygenated hemoglobin
• Macroscopic
• air-filled sinuses
Field Mapping
• image magnetization pattern at different echo times
• allows calculation of field offset based on phase accrual per unit time
• can be used to correct for distortion, but not dropout
� = tan�1 My
Mx|Mxy|
MRI Data is “Complex”
Magnitude (used) Phase (discarded)
Phase image - short TE
Phase image - long TE
Distortion vs. Dropout
• distortion is associated with large echo-spacing values in EPI readouts
• dropout is associated with large voxel dimensions
• the following slides illustrate that they are independent processes (even though both are caused by field inhomogeneities)
2 mm
EPI overMPRAGE
3 mm
4 mm
5 mm
2 mm
EPI overMPRAGE
3 mm
4 mm
5 mm
2 mm
EPI overMPRAGE
3 mm
4 mm
5 mm
Distortion Correction
• use of parallel imaging techniques to minimize EPI readout duration
• always acquire a field map
• only use 128 matrix EPI scans if you really need them
Avoiding Dropout
• simplest way to minimize dropout is by reducing voxel dimensions
• will require more smoothing to recover SNR
• other advanced techniques such as Z-shim may be used
• always check your EPI coverage by overlaying raw EPI scans on an MPRAGE
Typical Order of Operations
• motion-correction
• spatial smoothing
• linear modeling
• temporal filtering of drift and and residual motion as nuisance regressors
• distortion correction applied to effect-size estimates etc. prior to group GLM
Questions?