Spatial normalization, bulk motion correction and coregistration for
functional magnetic resonance imaging of the human cervical spinal
cord and brainstem
Patrick W. Stromana,b,c,⁎, Chase R. Figleya, Catherine M. Cahilla,d,e
aCentre for Neuroscience Studies, Queen's University, Kingston, Ontario, Canada K7L 3N6
bDepartment of Diagnostic Radiology, Queen's University, Kingston, Ontario, Canada K7L 3N6
cDepartment of Physics, Queen's University, Kingston, Ontario, Canada K7L 3N6
dDepartment of Pharmacology and Toxicology, Queen's University, Kingston, Ontario, Canada K7L 3N6
eDepartment of Anesthesiology, Queen's University, Kingston, Ontario, Canada K7L 3N6
Received 27 November 2007; accepted 17 January 2008
Functional magnetic resonance imaging (fMRI) of the cortex is a powerful tool for neuroscience research, and its use has been extended
into the brainstem and spinal cord as well. However, there are significant technical challenges with extrapolating the developments that have
been achieved in the cortex to their use in the brainstem and spinal cord. Here, we develop a normalized coordinate system for the cervical
spinal cord and brainstem, demonstrating a semiautomated method for spatially normalizing and coregistering fMRI data from these regions.
fMRI data from 24 experiments in eight volunteers are normalized and combined to create the first anatomical reference volume, and based
on this volume, we define a standardized region-of-interest (ROI) mask, as well as a map of 52 anatomical regions, which can be applied
automatically to fMRI results. The normalization is demonstrated to have an accuracy of less than 2 mm in 93% of anatomical test points.
The reverse of the normalization procedure is also demonstrated for automatic alignment of the standardized ROI mask and region-label map
with fMRI data in its original (unnormalized) format. A reliable method for spatially normalizing fMRI data is essential for analyses of group
data and for assessing the effects of spinal cord injury or disease on an individual basis by comparing with results from healthy subjects.
© 2008 Elsevier Inc. All rights reserved.
Keywords: Spinal cord; Brainstem; Human; Normalization; Methods; fMRI
Functional magnetic resonance imaging (fMRI) has
become one of the most powerful tools for neuroscience
research, and yet its use is limited to studies of the cortex,
with relatively few exceptions [1–4]. This means that studies
of distributed systems, such as those involved with pain, are
limited in their scope and completeness until fMRI can be
applied in the entire central nervous system (CNS). It has
been demonstrated that important neuronal activity involved
with modulation of pain and sensory responses can be
identified in the brainstem with fMRI [4–6]. Similarly, fMRI
of the spinal cord (spinal fMRI) has been demonstrated to
show areas of activity in the cervical andlumbar regions with
high sensitivity and reliability, in response to thermal,
sensory, motor and painful stimuli [3,7–9]. Descending
modulation of activity in the cervical spinal cord from
brainstem regions, corresponding to emotional factors and
focused attention, has also been demonstrated during thermal
sensory stimulation (unpublished results). Together, these
studies demonstrate the feasibility of fMRI spanning the
Despite the advances in fMRI techniques, there are
fundamental problems in extrapolating the technological
Available online at www.sciencedirect.com
Magnetic Resonance Imaging 26 (2008) 809–814
⁎Corresponding author. Departments of Diagnostic Radiology and
Physics, Centre for Neuroscience Studies, Queen's University, Kingston,
Ontario, Canada K7L 3N6. Tel.: +1 613 533 3245; fax: +1 613 533 6840.
E-mail address: firstname.lastname@example.org (P.W. Stroman).
0730-725X/$ – see front matter © 2008 Elsevier Inc. All rights reserved.
developments in the cortex to the brainstem and spinal cord
. These are caused by the unique challenges posed by
fMRI of the brainstem and spinal cord, such as the relatively
small cross-sectional dimensions and large rostral–caudal
(R/C) extent, the variable curvature of the spine between
individuals, the relatively poor local magnetic field homo-
geneity causing image distortion and low signal intensity in
standard T2⁎-weighted fMRI  and the problem of motion
with each heartbeat . These challenges have required
fMRI data acquisition and analysis methods to be adapted
and modified for use in the spinal cord . Here, we present
the first methods for spatial normalization and nonrigid
motion correction of fMRI data spanning the entire
brainstem and cervical spinal cord. These methods enable
voxel-by-voxel comparisons of results across subjects, as
well as automatic labeling of anatomical regions, as is
routinely done for essentially all fMRI studies of the brain.
This development is an essential step toward obtaining fMRI
data that span the entire CNS. It permits group analyses,
enables automatic identification of active regions and
facilitates comparisons with normative data for assessing
injury or disease in the spinal cord or brainstem.
fMRI methods that have been developed for use in the
spinal cord are a logical choice for fMRI of the brainstem
because of the similar challenges that are encountered. Most
fMRI studies of the brainstem to date that have succeeded
in detecting activity with good reliability have been based
on conventional brain fMRI methods [T2⁎-weighted echo-
planar imaging (EPI)] [5,6,12–14]. However, spin-echo
imaging methods are considerably less sensitive to magnetic
field inhomogeneities than gradient-echo methods, and fast
spin-echo methods suffer less distortion than spin-echo
methods with an echo-planar spatial encoding scheme. As a
result, the optimal spinal cord fMRI method is based on a
single-shot fast spin-echo [1,15]. The small transverse
dimensions and large R/C extent of the spinal cord have
been accommodated by imaging thin, contiguous sagittal
slices to obtain three-dimensional (3D) functional image
data with relatively high resolution . Motion artifacts
have been shown to be reduced by means of flow-
compensation gradients and spatial saturation pulses to
eliminate signal arising from anterior to the spine (i.e., heart,
lungs and throat).
The method proposed for the present study is based on an
established spinal fMRI method [15,16], with modifications
to span the entire cervical spinal cord and brainstem. This
choice of imaging method is significant for the proposed
normalization procedure because it produces images with
high spatial accuracy, unlike methods based on EPI spatial
encoding. Here, we define a coordinate system based on the
anatomy of the cord and demonstrate a method for
normalizing the anatomy to consistent dimensions and
posture. The results also enable definition of a standard
normalized reference volume and a standardized map of
anatomical regions. The normalization procedure can also
be applied in reverse, to conform the map of anatomical
regions onto the original image data, regardless of spinal
cord size or curvature.
2.1. Image acquisition
The data used in the present study (for the assessment of
the normalization and motion correction methods) are part
of a larger study of thermal responses in the brainstem and
spinal cord, but the fMRI results themselves are not the
topic of this article. Spinal fMRI studies were carried out in
a 3-T Siemens Magnetom Trio using a phased-array spine
receiver coil with subjects lying supine. Eight healthy
volunteers were studied and provided informed consent
prior to participating. Initial localizer images were acquired
in three planes as a reference for slice positioning for
subsequent fMRI studies. The functional image data were
acquired with a half-Fourier single-shot fast spin-echo
sequence, with the echo time set at 38 ms and with a
repetition time of 1 s per slice, to obtain essentially proton-
density-weighted images. Signal intensity changes observed
upon a change in neuronal activity were the result of signal
enhancement by extravascular water protons, as well as a
contribution from the BOLD effect, as described previously
[17,18]. Sagittal image data were acquired from 14
contiguous sagittal slices, each 2 mm thick, making the
effective repetition time equal to 14 s. Each slice was
centered roughly on the second cervical (C2) vertebra,
spanning from the C7/T1 disc to the top of the thalamus
with a 20 cm×10 cm field of view (FOV) and a 192×96
acquisition matrix. The resulting voxel dimensions were
therefore 1.04 mm×1.04 mm×2.00 mm. As previously
reported, spatial suppression pulses were applied anterior to
the spinal canal (to eliminate signal from this region), and
flow-compensation gradients were applied in the R/C
direction to reduce flow-induced image artifacts.
2.2. Motion correction and spatial smoothing
Correction for bulk motion of the spine and spatial
smoothing are addressed at the same time as spatial
normalization because they are related processes requiring
spatial transformations. The methods are applied in sagittal
planes to avoid combining data from slices acquired at
different times and because the most common bulk motion
occurs in the anterior–posterior (A/P) and R/C directions.
Custom-made software written in MatLab (The Math-
works Inc., Natick, MA) was used to first draw a reference
line along the anterior edge of the cord in a midsagittal
slice. This line was also extended along the brainstem in a
straight line, originating from the rostral end of the medulla
and running tangent to the anterior edge of the thalamus
(Fig. 1) to account for differences in neck flexion and head
position. The caudal edge of the pons and the C7/T1
intervertebral disc were marked along this line as position
810P.W. Stroman et al. / Magnetic Resonance Imaging 26 (2008) 809–814
The data were then smoothed spatially, but only in the
direction parallel to the long axis of the cord (i.e., parallel to
the manually defined reference line), in order to avoid
introducing partial-volume effects. This was achieved by
defining a two-dimensional (2D) kernel to smooth along a
line parallel to one axis of the image and then rotating the
kernel to be parallel to the closest section of the reference
line. The direction of the reference line varied slowly and
smoothly across the imaging FOV in every case; hence,
smoothing was applied piecewise to each image in 15×15
voxel sections to reduce the computation time.
Bulk motion that may have occurred during the acquisition
Processing Toolbox. The process was applied on a slice-by-
slice basis to align each image in the time series to the first
image. First, a set of reference points was defined so as to
overlay the entire spine. This included points along two lines
parallel to the manually drawn reference line (displaced by
5 pixelsanteriorlyand 10pixelsposteriorly fromthe reference
line), as well as points on a regular 2D grid (spaced 10 voxels
apart in each direction) spanning the entire image. This choice
of reference points encompassed the entire image, with a
greater concentration of points along the spine, so that the
alignment of the spine dominated the image transformation.
The “cpcorr” functioninMatLabdeterminedthe displacement
two images. The polynomial transformation required to
optimally align all of the points was then determined with
“cp2tform,” and the image transformation was applied using
the “imtransform” function. The result was nonlinear coalign-
ment of the images that accounts for changes in curvature of
the spine over time, as well as translation in the A/P and R/C
directions but not in the left–right (L/R) direction.
2.3. Spatial normalization
The spatial normalization procedure can be applied as
needed to anatomical images, to the results of fMRI data
analysis or to each volume of the time-series data to enable
group analyses. Applying the normalization after the
analysis enables the results to be displayed in both original
and normalized formats and eliminates the potential of
additional spatial smoothing that may be imposed if the
interpolation in the normalization process was applied before
the analysis. The normalization procedure consists of
aligning the data with a 3D coordinate system with one
dimension parallel to the long axis of the cord and the other
two running A/P and L/R. The 3D image data were first
interpolated to cubic voxels (1 mm in each direction) to make
this transformation, and the volume was resliced transverse
to the reference line every 1 mm along the R/C direction.
A linear transformation was then applied to the volume to
place the reference points — at the caudal edge of the pons
(the pontomedullary junction) and at the C7/T1 disc (shown
in Fig. 1) — at 65 and 205 mm, respectively, from the rostral
end of the volume. The span of 140 mm between the two
reference points was based on typical dimensions observed
in data from a small number of volunteers.
Initially, a reference volume was created by averaging
across all time points of a set of spinal fMRI data in which
the cord and brainstem were centered in L/R and A/P
directions and did not demonstrate any curvature. The data
from each individual subject were then aligned with this
reference volume by means of rigid-body L/R and A/P shifts
at each R/C position, thus fine-tuning the normalization and
correcting for any L/R and A/P curvature. A normalized
reference volume spanning the brainstem and cervical spinal
cord was then constructed by averaging 24 sets of time-series
fMRI data from eight healthy volunteers.
Based on the normalized reference volume, a standard
region-of-interest (ROI) mask was created. This is a binary
Fig. 1. Example of the manually defined reference line (black line) along the
anterior edge of the spinal cord and tangent to the anterior edges of
the medulla and thalamus in a midline slice. The lines that are orthogonal to
the reference line mark the R/C anatomical points of reference at the caudal
edge of the pons and the C7/T1 intervertebral disc. The subsequent spatial
normalization is guided by the reference line and R/C points of reference.
811 P.W. Stroman et al. / Magnetic Resonance Imaging 26 (2008) 809–814
mask with values of 1 in pixels overlying the spinal cord or
brainstem and 0 otherwise. The mask enables automatic
selection of image data from the cord and brainstem and
exclusion of surrounding anatomy (CSF, vertebrae, etc.).
Based on this ROI mask, a map to label anatomical regions
was then constructed by identifying the R/C positions of the
thalamus, midbrain, pons, medulla and approximate limits of
each cervical spinal cord segment. These R/C regions were
then further subdivided into dorsal–ventral and L/R sections,
and in the medulla, a midline region was also defined,
resulting in a total of 52 anatomical zones. A 3D volume was
then constructed with each pixel labeled according to its
corresponding zone, resulting in a standardized anatomical
region-label mask. The region-label mask and the ROI mask
can be used with the normalized data, or the inverse of the
transformation used to normalize the image data can be
applied to the masks, mapping them onto the original
(unnormalized) fMRI data.
2.4. Validation of the normalization procedure
The normalization results were assessed quantitatively by
determining the error between corresponding anatomical
points, a method described by Grachev et al. . Using this
approach, a large number of reference points (n=1000) were
defined on a regular 2D grid spanning the cord and
brainstem, as well as adjacent structures in a midline sagittal
slice. Using the MatLab function “cpcorr,” the optimal
displacement of each reference point (up to 4 pixels along
each axis) was determined to maximize the local correlation
between the normalized image data and the reference
images. The required displacement was taken as the error
in the normalized results at each reference point.
The efficacy of the normalization method across the
cervical spinal cord and brainstem is shown in Fig. 2, as an
averaged image across 24 experiments in the eight
volunteers studied. Anatomical details such as the vertebral
processes and the well-defined edges of the spinal cord and
brainstem structures demonstrate that these features were
precisely aligned across the volunteers.
The accuracy assessment, based on anatomical corre-
spondence, showed that the mean error was 0.3 mm and the
median was 0 mm. Peak errors were approximately 5 mm in
the midline sagittal plane, but only at a few localized sites
(primarily in the CSF near the edges of the thalamus and
brainstem). In these regions, errors may be overestimated
due to signal fluctuations caused by pulsatile CSF flow but
have no effect on the ROI or region-label masks. Fig. 3
shows a plot of the cumulative number of selected points
below a given level of error, demonstrating that approxi-
mately 90% of test points were perfectly aligned and that
Fig. 2. Panel A shows a midline slice of the initial reference template, which
is based on the average of one data set across all time points. An example of
a single normalized midline sagittal slice is shown in Panel B, and Panel C
shows the midline slice averaged over 24 experiments in eight volunteers.
Fig. 3. Cumulative fraction of selected anatomical points as a function of the
optimal calculated shift (i.e., error) to align midline sagittal and coronal
planes with a reference volume. Errors are shown individually along each of
the three axes, as well as the magnitude of error in the sagittal plane. The
magnitudes of errors in the L/R direction in the coronal plane were too small
to be detected with the assessment method used.
812 P.W. Stroman et al. / Magnetic Resonance Imaging 26 (2008) 809–814
essentially all (including those in the CSF) were accurate
within a few millimeters.
Fig. 4 shows a normalized region-label mask in midline
sagittal and coronal planes. The same region-label mask is
also shown after applying the inverse of the normalization
transformation, allowing it to be mapped back onto the
brainstem and spinal cord in the original image data.
The effectiveness of the normalization procedure that we
have developed for the spinal cord and brainstem is
demonstrated by the correspondence of anatomical details
across subjects. This can be directly observed by the fact that
distinct anatomical features remain well defined, even when
image data are averaged across a group of subjects with large
anatomical variability (Fig. 2). Estimation of the normal-
ization accuracy based on selected anatomical points shows
that 90% of the points deviate from the corresponding point
in the reference volume by less than 1 mm in the sagittal
plane and that 93% are within 2 mm (Fig. 3). Normalization
errors in the midcoronal plane were too small to be detected
with this method and are essentially zero. Errors were larger
in the A/P direction than in the L/R or R/C direction,
corresponding to the direction of the largest variation across
subjects (most likely owing to neck tilt and spinal cord
curvature). The reference points with larger errors (N2 mm)
in the sagittal plane tended to occur at localized regions
around the thalamus and in the CSF-filled subarachnoid
space and, thus, may reflect that the signal intensity was
variable in these regions as a result of CSF motion.
While the image data are observed to be aligned to a good
approximation, future studies may attempt to refine the
gray and white matter regions. The alignment between spinal
cord segments and vertebrae may also be accounted for in
future iterations of the normalization, particularly when
extended into the lumbar spinal cord where the alignment is
more variable across people. Nonetheless, the normalization
accuracy in the cervical spinal cord and brainstem is within
the range that is acceptable for brain fMRI  and is
adequate for subsequent application of more precise normal-
ization methods such as ANIMAL (Automated Nonlinear
Image Matching and Anatomical Labeling) .
The ability to apply an automatic ROI mask and define a
standardized set of anatomical region-label maps is demon-
strated for both normalized and original image data (Fig. 4).
Alternative region-label maps can also be constructed to
identify specific nuclei or regions, as needed. This is
essential for automatic and/or objective assessments of
where activity occurs in fMRI of the brainstem and spinal
cord because fMRI data do not provide contrast to delineate
brainstem nuclei or spinal cord gray matter. Reporting fMRI
results, whether for research or for future clinical applica-
tions, requires referencing the areas of activity to a
coordinate system or to anatomical features. Relying on
visual assessments or poorly defined anatomy is subjective
and may be prone to errors or bias. The results of the
normalization procedure we have developed demonstrate
Fig. 4. (A) Midline sagittal and coronal slices of the normalized reference volume with areas outside the spinal cord and brainstem masked out and with each of
52 anatomical regions labeled with a unique number. (B) The midline sagittal slice through the region-label mask after applying the reverse of the normalization
transform to map it onto the anatomy shown in Panel C. Panel D shows the region-label mask overlying the original image, demonstrating the reliability of the
normalization and reverse-mapping procedures.
813P.W. Stroman et al. / Magnetic Resonance Imaging 26 (2008) 809–814
that we can now objectively report the locations of areas of Download full-text
activity with an accuracy of a few millimeters. Therefore, it
is now possible to create normative activity maps in the
brainstem and spinal cord with a consistent format, allowing
us to compare results within or between patient populations
(with spinal cord injury or disease) and healthy volunteers.
The method that we have developed for 3D normalization
and temporal alignment of spinal cord and brainstem fMRI
data has been demonstrated to be effective. The accuracy of
the normalization and the temporal alignment has been
demonstrated to be within 2 mm or less in 93% of test points,
with a mean accuracy of 0.3 mm. This advance permits
unbiased comparisons of results across subjects, regardless
of cross-subject variability in spinal cord size or curvature.
Also, a standardized ROI mask has been created and voxel-
by-voxel region labels have been defined, either of which
can be displayed in a normalized orientation or mapped back
onto the original fMRI image data. Although further
refinements may still be required to precisely align gray
matter and white matter structures, and spinal cord segments
at all levels of the cord, this development is a significant step
toward making spinal fMRI a practical and reliable tool for
research or clinical assessments of spinal cord function.
We thank Sharon David for technical support for this
project, and we are grateful for the support provided by the
International Spinal Research Trust (U.K.) and the Canada
Research Chairs Program.
 Stroman PW. Magnetic resonance imaging of neuronal function in the
spinal cord: spinal FMRI. Clin Med Res 2005;3(3):146–56.
 Govers N, Beghin J, Van Goethem JW, Michiels J, van den HL,
VandervlietE,et al. Functional MRIof thecervicalspinalcordon1.5 T
with fingertapping: to what extent is it feasible? Neuroradiology 2007;
Functional responses in the human spinal cord during willed motor
actions: evidence for side- and rate-dependent activity. J Neurosci
 Fairhurst M, Wiech K, Dunckley P, Tracey I. Anticipatory brainstem
activity predicts neural processing of pain in humans. Pain 2007;128
 Zambreanu L, Wise RG, Brooks JC, Iannetti GD, Tracey I. A role for
the brainstem in central sensitisation in humans. Evidence from
functional magnetic resonance imaging. Pain 2005;114(3):397–407.
 Dunckley P, Wise RG, Fairhurst M, Hobden P, Aziz Q, Chang L, et al.
A comparison of visceral and somatic pain processing in the human
brainstem using functional magnetic resonance imaging. J Neurosci
 Stroman PW, Kornelsen J, Bergman A, Krause V, Ethans K, Malisza
KL, et al. Noninvasive assessment of the injured human spinal cord by
means of functional magnetic resonance imaging. Spinal Cord 2004;42
 Stroman PW, Cahill CM. Functional magnetic resonance imaging of
the human spinal cord and brainstem during heat stimulation.
Proceedings of the International Society for Magnetic Resonance in
Medicine 2006;14th Annual Meeting, Seattle, USA, May 6–12; 2006.
 Kornelsen J, Stroman PW. Detection of the neuronal activity occurring
caudal to the site of spinal cord injury that is elicited during lower limb
movement tasks. Spinal Cord 2007.
 Wilmink JT, Backes WH, Mess WH. Functional MRI of the spinal
cord: will it solve the puzzle of pain? JBR-BTR 2003;86(5):293–4.
 Figley CR, Stroman PW. Investigation of human cervical and upper
thoracic spinal cord motion: implications for imaging spinal cord
structure and function. Magn Reson Med 2007;58(1):185–9.
 Fairhurst M, Wiech K, Dunckley P, Tracey I. Anticipatory brainstem
activity predicts neural processing of pain in humans. Pain 2006.
 Brooks J, Tracey I. From nociception to pain perception: imaging the
spinal and supraspinal pathways. J Anat 2005;207(1):19–33.
 Tracey I, Ploghaus A, Gati JS, Clare S, Smith S, Menon RS, et al.
Imaging attentional modulation of pain in the periaqueductal gray in
humans. J Neurosci 2002;22(7):2748–52.
 Stroman PW, Kornelsen J, Lawrence J. An improved method for spinal
functional MRI with large volume coverage of the spinal cord. J Magn
Reson Imaging 2005;21:520–6.
 Stroman PW. Discrimination of errors from neuronal activity in
functional magnetic resonance imaging in the human spinal cord by
means of general linear model analysis. Magn Reson Med 2006;56:
 Stroman PW, Krause V, Malisza KL, Frankenstein UN, Tomanek B.
Extravascular proton-density changes as a non-BOLD component of
contrast in fMRI of the human spinal cord. Magn Reson Med 2002;48
 Stroman PW, Tomanek B, Krause V, Frankenstein UN, Malisza KL.
Mapping of neuronal function in the healthy and injured human spinal
cord with spinal fMRI. NeuroImage 2002;17:1854–60.
 Grachev ID, Berdichevsky D, Rauch SL, Heckers S, Kennedy DN,
Caviness VS, et al. A method for assessing the accuracy of intersubject
registration of the human brain using anatomic landmarks. Neuro-
 Collins DL, Neelin P, Peters TM, Evans AC. Automatic 3D
intersubject registration of MR volumetric data in standardized
Talairach space. J Comput Assist Tomogr 1994;18(2):192–205.
814P.W. Stroman et al. / Magnetic Resonance Imaging 26 (2008) 809–814