SCT: Spinal Cord Toolbox, an open-source software

for processing spinal cord MRI data

Benjamin De Leener1, Simon Lévy1,2, Sara M. Dupont1, Vladimir S. Fonov3, Nikola Stikov1,4, D.
Louis Collins3, Virginie Callot5,6, Julien Cohen-Adad1,2*
1
NeuroPoly Lab, Institute of Biomedical Engineering, Polytechnique Montreal, Montreal, QC,
Canada
2
Functional Neuroimaging Unit, CRIUGM, Université de Montréal, Montreal, QC, Canada
3
Montreal Neurological Institute, McGill University, Montreal, QC, Canada
4
Montreal Heart Institute, Montreal, QC, Canada
5
Aix-Marseille Université, CNRS, CRMBM UMR 7339, Marseille, France
6
AP-HM, Hopital de la Timone, Pôle d’imagerie médicale, CEMEREM, Marseille, France

*
Corresponding author. Department of Electrical Engineering, Polytechnique Montreal, 2700,
chemin de la Tour, Montréal, QC, H3T 1J4, Canada, jcohen@[Link]

Abstract
For the past 25 years, the field of neuroimaging has witnessed the development of several
software packages for processing multi-parametric magnetic resonance imaging (mpMRI) to
study the brain. These software packages are now routinely used by researchers and clinicians,
and have contributed to important breakthroughs for the understanding of brain anatomy and
function. However, no software package exists to process mpMRI data of the spinal cord.
Despite the numerous clinical needs for such advanced mpMRI protocols (multiple sclerosis,
spinal cord injury, cervical spondylotic myelopathy, etc.), researchers have been developing
specific tools that, while necessary, do not provide an integrative framework that is compatible
with most usages and that is capable of reaching the community at large. This hinders cross-
validation and the possibility to perform multi-center studies. In this study we introduce the
Spinal Cord Toolbox (SCT), a comprehensive software dedicated to the processing of spinal cord
MRI data. SCT builds on previously-validated methods and includes state-of-the-art MRI
templates and atlases of the spinal cord, algorithms to segment and register new data to the
templates, and motion correction methods for diffusion and functional time series. SCT is
tailored towards standardization and automation of the processing pipeline, versatility,
modularity, and it follows guidelines of software development and distribution. Preliminary
applications of SCT cover a variety of studies, from cross-sectional area measures in large
databases of patients, to the precise quantification of mpMRI metrics in specific spinal pathways.
We anticipate that SCT will bring together the spinal cord neuroimaging community by
establishing standard templates and analysis procedures.

Keywords

Spinal cord, MRI, software, template, atlas, open-source

1
1. Introduction
Pathologies of the spinal cord can result from neurodegenerative and vascular diseases, disc
degeneration, trauma and cancer, all of which can induce severe functional disabilities and
neuropathic pain (Adams and Salam-Adams, 1991; Rowland et al., 2008). Precise assessment of
the structural (e.g., extent of the lesion) and functional damage to the spinal cord is critical for
informing on the prognosis and for guiding the intervention therapy program (Bozzo et al., 2011;
van Middendorp et al., 2011). Moreover, the development of novel therapeutic approaches
require objective biomarkers that can help assessing the efficiency and specific effect of these
treatments, e.g., regeneration and axonal growth in spinal cord injury (Bradbury and McMahon,
2006) or remyelination in multiple sclerosis (Harlow et al., 2015; Luessi et al., 2014). While
conventional Magnetic Resonance Imaging (MRI) based on relaxation parameters and proton
density provides useful structural information that complement clinical and neurophysiological
exams (Cadotte et al., 2011; Kearney et al., 2015), it can sometimes miss subtle pathological
events such as Wallerian degeneration (Mac Donald et al., 2007; Zhang et al., 2009) or diffusely
abnormal white matter in multiple sclerosis (Laule et al., 2011; West et al., 2014). The ongoing
development of multi-parametric MRI (mpMRI), e.g., diffusion tensor imaging (DTI),
magnetization transfer ratio (MTR) and functional MRI (fMRI) (Cohen-Adad and Wheeler-
Kingshott, 2014; Tofts, 2003) provides a closer look at white matter microstructure and neuronal
function and thus can more precisely characterize the pathological spinal cord (Allan R. Martin
et al., 2016).
While mpMRI has been used in the brain for several decades now, spinal cord imaging in
research and clinics is still largely underutilized (Stroman et al., 2014; Wheeler-Kingshott et al.,
2014). One reason is the difficulty in acquiring good quality data due to the numerous artifacts
and the small cross-sectional size of the spinal cord. For the past few years though, researchers
have developed methods to overcome these challenges, such as advanced coil designs (Cohen-
Adad et al., 2011) and pulse sequences (Dowell et al., 2009; Finsterbusch, 2009; Wilm et al.,
2007). A second reason is that, in contrast to the brain, fewer tools exist that are dedicated to
processing of spinal cord images (Stroman et al., 2014). Unlike popular software packages for
brain neuroimaging (e.g., FSL, SPM, BrainVoyager, FreeSurfer, AFNI, MINC Toolkit), spinal
cord researchers have only been developing specific tools, e.g., for spinal cord segmentation (De
Leener et al., 2016). Moreover, these single tools don’t provide an integrative framework that is
compatible with most usages (i.e., adapted to a large variety of mpMRI protocols) and that can
reach the community at large (i.e., open source, licence-free scripting environment, extensive
documentation, support forum, continuous integration service for integrity testing). At the same
time, brain software packages are not optimized for spinal cord images because (i) the spine is an
articulated structure, therefore standard motion-correction algorithms assuming rigid or affine
transformation are inadequate, (ii) brain extraction and segmentation tools are not adapted to the
spinal cord because of different shape, contrast-to-noise ratio, etc., (iii) common brain MRI
templates (Evans et al., 1992; Fonov et al., 2011) and atlases (Brodmann, 1909) are de facto not

2
usable for spinal cord imaging, (iv) useful features are de facto missing, such as spinal cord
cross-sectional area (CSA) measurement. The lack of a standard processing platform has had
negative impacts on the spinal cord neuroimaging community as it has limited the ability of
researchers to compare and reproduce published results, as well at to conduct collaborative and
multi-center studies.
In this paper we introduce the Spinal Cord Toolbox (SCT), a software package specifically
designed to process spinal cord mpMRI data and to perform atlas-based analysis. SCT includes
state-of-the-art tools, which have been validated in this manuscript (see Appendix) and/or are
already published (see list of publications in the next section). SCT is compatible with any
scanner brand as it uses NIfTI format and is designed to accept a variety of sequences, modalities
and contrasts. Being based on Python language, SCT ensures cross-platform compatibility, is
free and open source. The article is organized as follows. First, an overview of the software is
presented, along with a description of the main features. Then, technical details are presented
(e.g., coding language, installation) followed by an example application and a discussion.

2. Main features
SCT is a comprehensive and open-source library of analysis tools for multi-parametric MRI of
the spinal cord. The primary objective of SCT is to provide a common pre-processing platform,
which supplements what is missing from the common brain software package. Hence, the goal of
SCT is not to replace entirely what other software already offers (e.g., first- and second-level
analysis of fMRI data), but to provide the necessary tools to pre-process mpMRI data, to perform
group analysis within standard space, and/or to perform cord-specific quantification such as
quantification of cross-sectional area across vertebral levels. A typical user-case example
consists of registering the spinal cord of several patients to a common template, then quantifying
DTI metrics along vertebral levels C2 to C5 within the corticospinal tract. This can be done
automatically with SCT, thanks to some of the following features:
- Automatic segmentation of the spinal cord (De Leener et al., 2014)
- Automatic segmentation of the spinal canal (which includes the cord and CSF) (De
Leener et al., 2015)
- Automatic segmentation of the white and gray matter using multi-atlas algorithm
(Dupont et al. 2016)
- Automatic vertebral labeling (Ullmann et al., 2014)
- MRI template of the human spinal cord (Fonov et al., 2014)
- Probabilistic template of white and gray matter (Taso et al., 2014)
- Atlas of white matter tracts (Lévy et al., 2015)
- Probabilistic map of spinal levels (Cadotte et al., 2014)
- Pipeline for registering data with the template (Fonov et al., 2014)
- Robust motion correction methods for diffusion and functional MRI time series (Cohen-
Adad et al., 2015).
A complete list of tools is available at: [Link]
The following section will detail some of the main features.
2.1. Segmentation
2.1.1. Spinal cord
The delineation of the spinal cord contour is a common procedure for quantifying CSA or spinal
cord volume (De Leener et al., 2016). As described in section “Template and atlases”,
segmentation can also be used to improve the quality of registration. Here we have implemented
a fast, robust and accurate automatic segmentation algorithm for spinal cord MRI called PropSeg
(De Leener et al., 2015, 2014). The algorithm consists of (i) detecting the center of the spinal
cord using ellipse detection and information from the body symmetry, then (ii) propagating a
tubular surface along the spinal cord edge using deformable models. The outputs include a
binary mask of the spinal cord, a 3D parametric surface and CSA measure for each axial slice
along the spinal cord. PropSeg has been validated against multiple contrasts (T1-, T2- and T2*-
weighted) (De Leener et al., 2014) and was shown to be robust in pathological cases, including
in patients with multiple sclerosis (MS) (Yiannakas et al., 2015).
While PropSeg is fully automatic, it can sometimes fail if contrast between the CSF and the cord
is low, if strong artifacts are present (e.g., caused by metallic implants or large cord compression
in trauma) or if the FOV only includes a small portion of the spinal cord (e.g., brain scan, with
only small portion of the cord visible). In such cases it is possible to adjust the parameters to
better fit the type of data (e.g., specifying that the spinal cord is only visible in the inferior
portion of the FOV) or to add manual initialization (e.g., three points in the spinal cord, or
several points along the spinal cord centerline, or “attraction points” which will force the 3D
mesh to be attracted to these points), similar to the method proposed by (Horsfield et al., 2010).
Manual initialization is handled by a window which pops up and shows axial slices of the cord:
the user can then quickly add labels on this window.

2.1.2. Gray matter

Gray matter segmentation in SCT builds upon the groupwise multi-atlas segmentation method
from Asman et al. (Asman et al., 2014) and includes additional improvements (Dupont et al.,
2016) (see Appendix). Firstly, prior knowledge of vertebral level (obtained from automatic
vertebral labeling) helps to select better candidates from the dictionary of gray matter shapes,
resulting in more accurate segmentation. Secondly, while the original method requires the input
image to have similar contrast to images from the dictionary, gray matter segmentation in SCT
includes a tissue-specific robust linear intensity normalization (obtained from the pre-registration
of the gray matter atlas) which makes it possible to input any image contrast, such as T 2*-
weighted, magnetization transfer or fractional anisotropy map from DTI. In addition to providing
clinically relevant information, gray matter segmentation also proves to be useful for improving
the accuracy of atlas-based registration, as will be described in the next section.

2.2. Template and atlases

SCT includes the MNI-Poly-AMU1 template, which is a collection of MRI volumes that forms a
common anatomical space reference, covering C1 to T6 vertebral levels sampled at 0.5 mm

1
The acronym stands for the Montreal Neurological Institute (MNI), Polytechnique Montreal (Poly) and Aix-
Marseille Université (AMU).
isotropic. This template includes an averaged T2-weighted volume (Fonov et al., 2014), a binary
mask of the spinal cord and surrounding cerebrospinal fluid, a labeling of vertebral levels,
probabilistic labeling of spinal levels (Cadotte et al., 2014), a probabilistic atlas of the white and
gray matter (Taso et al., 2015) and an atlas of spinal white matter tracts with partial volume
information (Lévy et al., 2015). The template was constructed from straightened spinal cord for
facilitating registration and visualization of results. Note that the T 2w template and the atlas of
white and gray matter resulted from the fusion of 15 T2-w and 16 T2*-w images from different
subjects and different scanners. Figure 1 illustrates the MNI-Poly-AMU template. The next
sections describe how to register a new subject with the template.
Figure 1. Template and atlases included in SCT: straight T 2-weighted template with vertebral
levels and spinal cord/CSF segmentation (top left), probabilistic atlases of white and gray
matter (top center), probabilistic map of spinal levels according to vertebral levels (top right)
and white matter atlas (bottom).

2.2.1. Registering new data with the template

One motivation for registering new data with the template is to perform atlas-based analysis of
mpMRI data. Registration is best done using an anatomical dataset (e.g., T1- or T2-weighted)
with isotropic resolution (e.g., 1mm3). The main steps are:
1. Spinal cord segmentation (e.g., using PropSeg)
2. Identification of two vertebral levels. Each of these two landmarks consists of one
voxel located in the middle of the spinal cord, at the level of the corresponding mid-
vertebral body. The value of each voxel corresponds to the vertebral level, e.g., value=2
for C2 and value=10 for T3. Note that this vertebral labeling can be done automatically
(Ullmann et al., 2014) using function sct_label_vertebrae, assuming the data are
acquired in a consistent manner (e.g., FOV centered at the C3/C4 disc).
3. Registration to the template. Firstly, the spinal cord image is straightened to match the
shape of the template. Straightening is achieved by automatically generating a series of
labels (cross-shape) along the spinal cord centerline, and then registering these labels to a
straight centerline using landmark-based affine and nonlinear transformations as
implemented in ITK (McCormick et al., 2014). Secondly, local deformations are
performed using multi-step nonlinear registration constrained in the axial plane and
regularized in the inferior-superior direction (for more details, see section “Multi-modal
registration”).
4. Warping the template back into the native space of the subject by applying the
transformation to each template object (e.g., spinal cord mask, vertebral and spinal
labeling, white matter atlas, etc.).
The registration procedure outputs forward (warp_anat2template.[Link]) and backward
(warp_template2anat.[Link]) transformations in ITK format, which can be concatenated to
subsequent transformations if needed. For example, if other mpMRI data (e.g., DTI) from the
same subject need to be registered to the template, these data are first registered to the anatomical
data, then warping fields are concatenated to yield a single direct transformation between the
template and these additional data. This approach ensures minimum interpolations and hence
preserves the effective spatial resolution of the data. If these mpMRI data also present enough
gray/white contrast, gray matter segmentation can be performed to further improve the
registration of the white matter atlas. Figure 2 illustrates the procedure. The next section
describes how registration is performed in SCT.
Figure 2. Overview of the template-based analysis pipeline. Firstly, anatomical data are
registered to the template (blue arrows). Additional multi-parametric data acquired during the
same scan session (e.g., DTI, MTR, fMRI) are registered to the anatomical data, then template
objects are warped to the multi-parametric data (green arrows). To improve accuracy of
template registration, it is possible to add a step where gray matter is segmented (manually or
automatically) and then warped to the gray matter template in order to update the warping fields
(purple arrows). Subsequently, mpMRI metrics can be quantified within the spinal cord or within
specific tracts of the white matter at specific vertebral levels (red arrows). Cord and gray matter
CSA and cord volume can also be computed from the mpMRI data, or from the anatomical data
in the first place.

2.3. Multi-modal registration

2.3.1. Slice-by-slice regularized registration (SliceReg)
In typical mpMRI protocols, there is a need to co-register multiple data from one subject
together, e.g., T1-, T2-, T2*-, diffusion, MT-weighted images and fMRI time series. While
common registration methods for the brain employ unconstrained affine or non-rigid
registrations (Klein et al., 2009), multimodal registration in spinal cord MRI data is often
performed using affine transformations constrained in the axial plane (Figure 3, left panel),
assuming that subject motion in the inferior-superior direction is negligible because subjects are
usually lying in supine position. Although fairly robust, affine transformations often lack
accuracy because of the articulated nature of the spine, which can produce non-rigid
deformations. To overcome this problem, planar translations and/or rotations can be estimated
for each slice in the axial plane (“Slice-by-slice” in Figure 3). While this method has become
popular in the spinal cord imaging community (Summers et al., 2014), correcting each slice
independently is sub-optimal because it does not consider the smooth nature of the deformation
along z, hence reducing robustness compared to volume-based registration. SCT offers a new
registration method that estimates slice-by-slice translations while ensuring regularization
constraints along z (SliceReg in Figure 3). Regularization is achieved using a polynomial
function (order to be specified by the user) and is part of the convergence algorithm of the cost
function (i.e., it is not applied after the estimation but during). SliceReg was shown to be more
accurate than rigid-body transformations and offers more robustness than non-regularized slice-
by-slice registration (Cohen-Adad et al., 2015). This algorithm is part of the generalized
procedure for multi-modal registration, as described in the next section.

Figure 3. Illustration of methods for registration of spinal cord MRI.

2.3.2. Multi-step registration

[Link]. Purpose

The purpose of sct_register_multimodal is to register one image (source) to another

image (target), such as for instance, an MT_off to an MT_on, or a T2 anatomical to a mean b=0.
As suggested by these examples, the source and target images do not need to be in the same
space. The function features several methods for affine and non-rigid registration, as listed in
Table 1. All these methods are constrained in the axial plane (with or without regularizations
along z), hence the target volume is assumed to be in the radiological orientation2.

An important feature of sct_register_multimodal is that several registration steps can be

prescribed in a single command line. This allows to perform multi-scale registration while using

2 Radiological orientation refers to the order of data storage for a volume, with x, y and z coordinates corresponding
to Right→Left, Posterior→Anterior and Inferior→Superior, respectively.
different algorithms at each step. Typically, the first step would consist in addressing large
deformations while latter steps would deal with cord shape and fine anatomical adjustments.
Multiple registration steps are specified by the flag “-param” as shown in the following example

sct_register_multimodal -i [Link] -d [Link] -iseg

image_seg.[Link] -dseg dest_seg.[Link] -param
step=1,type=seg,algo=slicereg,metric=MeanSquares,smooth=5,poly=5:step=2,type
=seg,algo=bsplinesyn,metric=MeanSquares,smooth=1,iter=5:step=3,type=im,algo=
syn,metric=CC,iter=5

A list of parameters available for each step is given in Table 1, along with explanations and
suggestions for best usage. The accuracy and robustness of registration has been addressed in
(Fonov et al., 2014). An example of mpMRI registration with the template is illustrated in Figure
4.

Table 1. Parameters available for multi-step registration.

Parameter Value Description

step int Step identification number (starts with 1).

type im: image Type of information used for registration. If
seg: cord segmentation cord segmentation is available, it could be
used for robust gross alignment (first step).
In some cases (e.g., if images are corrupted
by artifacts), cord segmentation can also be
used at later steps as a more reliable
definition of cord contour.
algo translation: Tx, Ty (2 dof) Type of transformation. For large-scale
rigid: Tx, Ty, Rz (3 dof) deformations (e.g., gross realignment of
affine: Tx, Ty, Rz, Sx, Sy (5 dof) cord centerlines between source and target
syn: (Avants et al., 2008) image), rigid/affine transformations based
bsplinesyn: (Tustison and on cord segmentations are prefered. For
Avants, 2013) small-scale deformations, the algorithm
slicereg: slice-wise regularized SyN based on the image is prefered. To
rigid (Cohen-Adad et al., 2015) maintain the consistency of the internal
structure (e.g., gray matter and surrounding
white matter tracts), users can prefer the B-
spline-regularized version of SyN (Tustison
and Avants, 2013).

slicewise int 0: volume-based registration

1: slice-by-slice registration.
Only for algo={translation, rigid, affine}.
metric CC: cross-correlation (slow, Similarity metric used for registration. In
recommended for type=im) general, for type=im we recommend CC
MI: mutual information (performs well on multi-modal images but
MeanSquares: mean square (fast, slow), while for type=seg we recommend
recommended for type=seg) MeanSquares (fast). The metric MI (mutual
information) can be used for both images
and segmentations, but requires enough
image overlap and voxel count to compute
histogram (otherwise it fails).
iter positive int Number of iterations in current registration
step.
shrink strictly positive int (>0) Shrink factor, which specifies sub-sampling
factor of the destination volume for faster
computation. Default=1. Only for
algo={syn, bsplinesyn}
smooth positive int Smooth factor, which specifies in mm the
sigma of the Gaussian smoothing kernel
applied to the source and destination
volumes before registration. Recommended
for type=seg. Default=0. Only for
algo={syn, bsplinesyn}
gradstep strictly positive float Gradient descent optimization. The higher
the more deformation. Default=0.5. Only
for algo={syn, bsplinesyn}
poly positive int Polynomial degree. Only for algo=slicereg.
Figure 4. Example of co-registration between T1-weighted MPRAGE, T2-weighted fast spin echo,
T2*-weighted gradient echo, diffusion-weighted image, T2*-weighted EPI and the MNI-Poly-
AMU template. Note that in the “native space” box, the category “template” refers to the
warped template, and the category “atlas” refers to the warped atlas. Despite the different
contrasts and levels of distortions across sequences, all images were successfully registered to
the template, as assessed qualitatively. The red grid overlaid on images and the checkerboard
images enable to appreciate the spatial correspondence within each sequence (between native
and template), and the slight deformation/motion inherent to each sequence. Reproduced with
permission from (Fonov et al., 2014).

2.4. Atlas-based analysis of mpMRI metrics

Once mpMRI data are registered with the template it is possible to extract relevant metrics (e.g.,
DTI fractional anisotropy) using pre-existing atlases, without the need to manually draw ROIs.
This analysis can either be performed on the template space (after warping the mpMRI data to
the template using the function sct_apply_transfo) or on the mpMRI space (after warping
the template objects to the mpMRI data using the function sct_warp_template). The former
approach is preferred for obtaining group maps of mpMRI results, while the latter approach is
preferred for minimizing interpolation errors during metric extractions.

Metrics can be quantified automatically using the function sct_extract_metric and an

atlas. The default atlas (Lévy et al., 2015) is the perfect candidate as it includes Ntracts=30
different tracts of the white matter as well as the gray matter and the surrounding CSF.
A notable feature of metric quantification in SCT is the possibility to account for partial volume
effect using Gaussian mixture modeling, which provides a more accurate estimate of mpMRI
metrics within the spinal cord regions in comparison to the traditional average within binary
mask (Lévy et al., 2015). In brief, the signal in each voxel is assumed to be a linear combination
of signals from adjacent regions (e.g., gray matter, CSF) and weighted by the partial volume
information of each region. The simplest form of estimation is the maximum likelihood (ml).
Let’s call n the total number of voxels in the template, y the metric data rearranged in vector
form (n×1), X the atlas containing partial volume values (n×Ntracts) and β the ‘true’ metric
within each tract (Ntracts×1). The least-square solution of β is:

þ = (Xt ⋅ X)–1 ⋅ Xt ⋅ y
A limitation of the maximum likelihood method however is its poor robustness to noise, hence
values estimated within small tracts can be aberrant. To mitigate this poor robustness, constraints
can be added to ‘force’ the estimated β needs to be close to a prior β 0. This can be done using the
maximum a posteriori (map) approach. The equation now reads:

þ = þO + (Xt ⋅ X + I)–1 ⋅ Xt ⋅ (y − X ⋅ þO)

with I being the identity matrix. Here, a pragmatic approach was chosen to estimate β0 a priori,
by considering that tracts within the white matter have similar values, therefore the white matter
can be considered as a single class with a global mean. In order to robustly and accurately
estimate β0 in the white matter, maximum likelihood estimation was employed, by considering
the following three classes that represent adjacent regions: white matter, gray matter and CSF.
Following estimation of β0, the map method can be applied.
More traditional quantification methods are also available in SCT, including the binary average
(bin) and the weighted average (wa). The accuracy and precision of each method was assessed
using simulations (Lévy et al., 2015), and suggest highest performance for the map method.
Further simulations on the effect of spatial resolution are available in Annex 9.8. Note that
metrics can be estimated in single tracts, in a group of tracts (e.g., right and left fasciculus
gracilis and cuneatus) or in the entire white matter. It is also possible to specify the range of
slices or the vertebral levels to extract the metrics from (e.g., C2 to C5). Also note that extraction
of metric in the gray matter is possible with correction for partial volume effect (using maximum
likelihood). However, the current version of the gray matter template does not include sub-
structure (e.g., motor neurons, dorsal nucleus, etc.), therefore the extracted value will correspond
to the total gray matter at a given slice (or group of slices).

2.5. Miscellaneous tools

2.5.1. Motion correction for diffusion and functional MRI time series
Diffusion and functional MRI datasets consist of a collection of multiple volumes whose number
can range from 10-20 to several hundreds. These data usually take several minutes to acquire. If
the subject moves between volumes, this can compromise the analysis done on the whole dataset
(e.g., tensor estimation or statistical analysis in fMRI). It is thus common procedure to perform
motion correction, which consists in registering all volumes to a target volume (e.g., the first
one). However, as mentioned previously in section “2.3. Multi-modal registration”, spinal cord
movement can be non-affine (e.g., if subject tilts the head) therefore in contrast to the brain,
affine transformations cannot be employed (or at the risk of poor accuracy). It is commonly
accepted to perform motion correction using slice-wise translations in axial plane (Cohen-Adad
et al., 2010; Duval et al., 2015; El Mendili et al., 2014; Grussu et al., 2015; Iglesias et al., 2015;
Kong et al., 2012; Mohammadi et al., 2013; Smith et al., 2010), which in SCT can be performed
using regularization along the z direction (SliceReg algorithm).
In addition to this feature, motion correction in SCT includes the methods of Xu et al. for
grouping successive volumes, thereby improving the robustness of registration in high b-value
diffusion MRI data (Xu et al., 2013). For example, if one volume is acquired with diffusion-
weighted gradients along the spinal cord, signal might be too low to estimate a meaningful
transformation. In such cases it can be preferred to rely on volumes adjacent in time and estimate
the transformation from groups of n successive volumes (typically 3 to 5) averaged together.
This approach is particularly successful for correcting slow drifts in long acquisitions, but fails in
presence of strong Eddy-current distortions, which affect each diffusion-weighted volume
independently from its next predecessor and successor. In that case, n=1 would be preferable.
Note that the target volume is iteratively updated by averaging the registered volumes with the
target volume.

Another feature that improves motion correction is the possibility to mask out undesired regions.
It is known in spinal cord times series that other structures can move independently from the
spinal cord, such as the muscles or the chest (in thoracic imaging). Given that these structures
usually contain more voxels than the spinal cord, estimating a global transformation on the entire
volume can lead to spurious displacement of the spinal cord (because the metric of the cost
function will be weighted towards larger structures). Hence, a mask of the spinal cord can be
input for motion correction to overcome this problem. This mask can be obtained automatically
from the spinal cord segmentation using the function sct_create_mask.
2.5.2. Smoothing along spinal cord centerline
Denoising an MRI image of the spinal cord with a classical isotropic Gaussian smoothing kernel
leads to undesired partial volume effect with CSF, therefore decreasing the accuracy of spinal
cord measurements. Even a unidirectional kernel (e.g., in the inferior-superior direction) can lead
to CSF contamination in some parts of the spinal cord due to its variable curvature. To
circumvent this issue, SCT features an algorithm that can smooth a dataset along the spinal cord
centerline, as illustrated in Figure 5. This approach is analogous to the 2D surface smoothing that
is popular for brain cortical analysis (Hagler et al., 2006; Kiebel et al., 1999). Our algorithm
works by straightening the cord, applying a 1-D Gaussian smoothing kernel (size can be adjusted
by the user) and then un-straightening the cord back into the original space. Centerline
smoothing can be useful for improving the sensitivity for spinal cord fMRI experiments or for
structural assessment of degenerative pathology occurring along the superior-inferior direction.
Further demonstration about the usefulness of this algorithm is required.
Figure 5. Methods to smooth the spinal cord. From left to right: no smoothing; smoothing using
isotropic Gaussian kernel (classical approach); smoothing using anisotropic Gaussian kernel;
smoothing using an adaptive Gaussian kernel oriented along the spinal cord centerline.

3. Technical considerations
3.1. Licence, language and dependences
SCT is an open-source project that falls under the MIT license3. SCT is written in Python and has
been designed in an object-oriented programming fashion in order to improve modularity and
extensibility. SCT tools are available via two interfaces: (i) a command-line software, meaning
that tools can be called within a Unix terminal, and (ii) a Python library called within Python
code.
Dependent Python libraries include: nibabel4 for reading/writing NIFTI files, numpy5 and scipy6
for scientific computation and dipy7 for processing diffusion-weighted imaging. The only
external dependent software is ANTs (Avants et al., 2014), however the required binaries are
already included in SCT packages therefore users don’t need to install it. The image format that
is currently supported by SCT is NIfTI 8. In order to ensure stability and reproducibility towards
Python dependences, SCT comes with an integrated light distribution of Python, called
Miniconda9. When called via command-line, each script calls a Python launcher that sets a

3 [Link]

4 [Link]

5 [Link]

6 [Link]

7 [Link]

8 [Link]

9 [Link]
virtual environment for the correct Python distribution and then runs the script. This prevents
SCT from interfering with the user’s default Python distribution.

4. Example applications
4.1. Analysis of mpMRI data
An adult subject was scanned at 3T (TIM Trio, Siemens Healthcare) using the following
sequences: (i) 3D T2-weighted fast spin echo, (ii) gradient echo FLASH with and without
magnetization transfer and (iii) diffusion-weighted EPI (b-value = 800 s/mm 2). For details on
sequence parameters see (Fonov et al., 2014). Processing included spinal cord segmentation,
vertebral labeling, gray matter segmentation, cross-sectional measurement, registration to
template, motion correction using SliceReg and metrics extraction using maximum a posteriori.
Figure 6a shows vertebral labeling, CSA overlaid on the anatomical data and the corresponding
CSA plot from C1 to T6. Figure 6b shows the gray (red) and white matter (blue) segmentation
with the corresponding CSA plot between C1 and C7. Figure 6c shows magnetization transfer
ratio (MTR), fractional anisotropy (FA), radial diffusivity (RD) and mean diffusivity (MD)
images with overlay of four spinal tracts: left and right lateral corticospinal tracts and cuneatus.
The table shows the maximum a posteriori estimations within each tract.
Figure 6. a: Following spinal cord segmentation and vertebral labeling, cross-sectional area
(CSA) is overlaid on the anatomical data and values can be extracted at specific vertebral levels
(here, between C1 and T6). b: Following gray and white matter segmentation on the T2*-
weighted data, CSA values can be computed. Note that output segmentations include partial
volume information, therefore CSA estimation is more precise than when using binary masks.
Here also, vertebral labeling can be used to extract CSA at specific vertebral levels (here,
between C1 and C7). c: Following registration of FLASH and diffusion data to the template,
magnetization transfer ratio (MTR), radial diffusivity (RD), fractional anisotropy (FA) and mean
diffusivity (MD) were computed within the gray matter (not displayed for clarity purpose) and
four white matter tracts: corticospinal tracts left (red) and right (yellow), cuneatus left (green)
and right (blue). Values were extracted between C1 and C4 levels. Note that there is an apparent
overlap between the cuneatus ROIs and the dorsal horns of the gray matter, which is related to
the partial volume encoding of the ROIs (thresholded at 0.3 for visual purpose).

4.2. Applications at other field strengths and in pathologies

While SCT has mostly been validated and used in healthy data from 3T scanners, it can also be
applied to lower/higher field strength as well as in patients, providing data quality is sufficient.
Figure 7 shows example applications at 1.5T, 3T and 7T, in a variety of spinal cord pathologies
Figure 7. Example application in T2*-weighted data from 1.5T, 3T and 7T, different vendors and
different pathologies. Here, SCT was used to segment the spinal cord, registering the MNI-Poly-
AMU to the data, segmenting the gray and white matter, adjusting the warping field based on
gray matter shape, and finally warping the white matter atlas to the template. The white/gray
matter is shown in blue/yellow respectively, and is thresholded at 0.4 for visual purpose. The
left/right corticospinal and left/right gracilis are respectively shown in yellow/blue and
green/red and are thresholded at 0.4. Parameters for each subject are listed hereafter. HC
@1.5T: GE Signa HDxt, neurovascular coil, 2D MERGE sequence, TR/TE=607/16ms,
matrix=320×320, resolution=0.8×0.8×4mm3 (interpolated to 0.4mm in-plane using zero
padding), R=2 acceleration, bandwidth=244Hz/pixel. HC @7T: Siemens 7T whole-body
scanner with SC72 gradients and 4ch-Tx/19ch-Rx coil (Zhao et al., 2014), multi-echo FLASH
(all echoes were averaged), TR/TE=512/[7.8,15,22,29.4]ms, matrix=384×384,
3
resolution=0.35×0.35×3mm , R=2 acceleration, bandwidth=195Hz/pixel. Syringo @3T:
Siemens TIM Trio, 4ch neck coil, multi-echo FLASH (all echoes were averaged),
TR/TE=539/[5.41,12.56,19.16]ms, matrix=320×320, resolution=0.5×0.5×5mm3, R=2
acceleration, bandwidth=195Hz/pixel. CSM @3T: GE Signa HDxt, neurovascular coil, 2D
MERGE sequence, TR/TE=550/13ms, matrix=320×320, resolution=0.6×0.6×4mm3, R=1
acceleration, bandwidth=244Hz/pixel. Data courtesy of Drs. Allan Martin and Michael Fehlings
from Toronto Western Hospital. MS @7T: Siemens 7T whole-body scanner with SC72 gradients
and 4ch-Tx/19ch-Rx coil, multi-echo FLASH, TR/TE=500/[7.8,13.7,18.4]ms, matrix=480×534,
resolution=0.4×0.4×3mm3, R=2 acceleration, bandwidth=199Hz/pixel. Data courtesy of Dr.
Caterina Mainero from Massachusetts General Hospital. ALS @7T: Siemens 7T whole-body
scanner with SC72 gradients and 4ch-Tx/19ch-Rx coil, multi-echo FLASH,
TR/TE=514/[7.8,15,22,29.4]ms, matrix=534×484, resolution=0.35×0.35×3mm3, R=2
acceleration, bandwidth=195Hz/pixel. Data courtesy of Dr. Nazem Atassi from Massachusetts
General Hospital. SCI @7T: Siemens 7T whole-body scanner with SC72 gradients and 4ch-
Tx/19ch-Rx coil, multi-echo FLASH, TR/TE=514/[7.8,15,22,29.4]ms, matrix=534×480,
resolution=0.37×0.37×3mm3, R=2 acceleration, bandwidth=195Hz/pixel. Data courtesy of Dr.
Anne-Louise Oaklander from Massachusetts General Hospital.

5. Discussion
SCT is an open-source image processing software dedicated to spinal cord mpMRI data. SCT
includes state-of-the-art MRI templates and atlases of the spinal cord internal structure, robust
methods to register new data to the template and motion correction methods for diffusion and
functional time series.

5.1. Atlas-based analysis

An important feature of SCT is the MNI-Poly-AMU template, which provides a spatial reference
for quantifying mpMRI data within the spinal cord and/or in specific white matter tracts (e.g, the
left corticospinal tract). One limitation of the current white matter atlas is that it is built from a
single drawing showing the C4 mid cervical level (Standring, 2008), which is then extrapolated
to superior and inferior levels using non-linear warping (Lévy et al., 2015). While the warping is
regularized using BSplineSyN function (Tustison and Avants, 2013) and takes into account the
gray matter shape, the exact location of tracts might change across levels and across subjects,
especially in diseased patients where only specific tracts degenerate (e.g., amyotrophic lateral
sclerosis) thereby altering the overall distribution of tracts in the cross-sectional plane. While we
believe these discrepancies might induce inaccuracies when quantifying metrics within certain
tracts (or group of tracts), there is currently no known solution as to how find the exact location
of spinal tracts in vivo. Brain researchers face similar issues when warping subject’s brain onto a
common template based on MRI-visible anatomical features (gyri/sulci, ventricles, etc.), without
knowing the exact internal structure of white matter tracts or functional clusters (Frost and
Goebel, 2012; Mori et al., 2009). We thus believe that although it has flaws, the proposed
approach for quantifying tracts in the cord is still the least biased in comparison with the current
state-of-the-art, which consists in manually drawing an ROI in “what appears to be” the tract of
interest based on the user’s knowledge of spinal cord anatomy (Ciccarelli et al., 2007; Cohen-
Adad et al., 2008; Gullapalli et al., 2006; Klawiter et al., 2011; Lindberg et al., 2010; Narayana
et al., 2004; Onu et al., 2010; Qian et al., 2011; Smith et al., 2010; Xu et al., 2013). This
approach has major flaws: (i) the identification of the tract is biased by the user experience and
knowledge of the anatomy, (ii) the manual delineation of ROIs is long and tedious, especially if
several slices, tracts and subjects are involved and (iii) ROIs consist of binary masks and hence
do not account for partial volume effect. The approach offered in SCT solves all these issues.
Firstly, since the atlas is aligned with the MNI-Poly-AMU T2-weighted template, registration of
any mpMRI data with the template ensures that the atlas is also aligned with the subject’s spinal
cord. Moreover, this procedure is done automatically allowing processing of a large database
without user bias. Secondly, each voxel of the atlas represents the fractional volume of a specific
tract, normalized by the probability to be in the white matter (Taso et al., 2014). Hence, partial
volume effect can be accounted for during metric extraction using Gaussian mixture modeling
(maximum likelihood, maximum a posteriori).
While quantifying metrics within large tracts (e.g., lateral corticospinal) was shown to be fairly
robust using typical sequences and resolutions (Lévy et al., 2015), quantifying metrics in small
tracts that sometimes only cover a fraction of a voxel in the cross-sectional plane is more prone
to noise and thus requires caution and further consideration. However, when considering the
superior-inferior dimension, the cumulative volume fraction can be higher than a voxel. An
important aspect of metric quantification in the cord is to take advantage of the elongated nature
of the tract, while assuming homogeneous metric value across a certain rostro-caudal distance
(which of course might not always be true, as discussed later). As an example, let’s take the left
lateral reticulospinal tract across C3 level. The volume of the tract 16 at this level is about 4.6
mm3. With a typical resolution of 0.8×0.8×4 mm3 (= 2.56 mm3), the voxel fraction would be
approximately 1.8 (4.6/2.56), which might be sufficient for quantifying a metric, depending on
the underlying noise and extraction method. To further investigate the effect of resolution on
metric extraction, simulations have been conducted and are presented in Annex 9.8. Despite the

10To calculate this volume the command is:

source sct_env; sct_process_segmentation -i $SCT_DIR/data/atlas/WMtract [Link] -p

csa -vert 3 -vertfile $SCT_DIR/data/template/MNI-Poly-AMU_level.[Link]
encouraging results of this simulation, suggesting that metrics can be quantified within small
tracts with minimum error (<2%) if sufficient SNR (>10) and voxel fraction (>30) are available,
we would like to stress that metric extraction from real data also suffers from image artifacts
(e.g., chemical shift, distortions, dropout, motion), mis-registration and subject-to-subject
differences in the underlying white matter anatomy. To summarize, the important aspects to
consider for robust metric estimation are: the number of voxels covering the tract (in voxel
fraction) and the level of noise. There is no universal approach as to what minimum tract size
people can quantify, as it not only depends on the robustness of estimation (which, as mentioned
before, depends on tract voxel fraction and noise level), but also on the scientific question that is
being addressed. For example, if the goal is to detect a metric difference between two
populations, then the anticipated effect size should be part of the statistical power calculation.
We therefore encourage people to choose the metric extraction method in light of the scientific
hypotheses. For instance, if the goal is to detect subtle Wallerian degeneration above a
compressed cord, then there is likely a diffuse abnormality that spans several vertebral levels
above, and therefore estimation should encompass all levels above in order to increase the
robustness of the estimation. Grouping tracts together (e.g., right and left fasciculus gracilis and
cuneatus to form the dorsal column) is another trick to gain robustness, which again should
comply with the underlying hypotheses.
Another aspect of the MNI-Poly-AMU template that warrants caution is the probabilistic spinal
level location based on (Cadotte et al., 2014). We would like to stress that this location is based
on 20 adult subjects and that further validation (e.g., from fMRI studies) is needed.

5.2. Modularity and cross-compatibility with other software

SCT was designed to be modular, i.e., to facilitate the addition of new functionalities, template or
atlases. For example, if researchers want to use an external template (e.g., for pediatric
population), they can specify the location of this external template when calling the function for
template registration. Another aspect of the modularity is the possibility to reuse part of SCT into
other spinal cord imaging packages. Like most brain software packages, SCT is composed of
callable functions with explicit input/output and proper documentation, therefore some of them
can easily be wrapped into other software packages.
A notable aspect that facilitates cross-compatibility is that the generated transformations (affine
or warping fields) are compatible with ITK and ANTs protocols, but can also be used with other
software such as FSL and SPM. A typical example of fMRI analysis would be to preprocess the
fMRI time series with SCT (e.g., motion correction, registration to the template) and then use the
estimated transformations to perform the first and second level analysis of the general linear
models implemented in FSL, as was shown in a recent publication (Kong et al., 2014).

5.3. Applications
Since its first release (June 2014), SCT has been used in several studies such as resting-state
fMRI of the spinal cord (Kong et al., 2014; Eippert et al., 2016), fMRI study of cord activity
during an isometric upper extremity motor task (Weber et al., 2015) and thermal stimulation
(Weber et al., 2016), investigation of microstructure in aging (Taso et al., 2016),
adrenomyeloneuropathy (Castellano et al., 2016) and infectious myelitis (Talbott et al., 2016),
automated white matter/gray matter segmentation and tensor-based morphometry (Taso et al.,
2015), mapping of fiber orientation in ex vivo human (Foxley et al., 2015), mapping of axonal
diameter (Duval et al., 2015) and myelin g-ratio (Duval et al., 2016) in vivo in humans,
correlation of cord atrophy with ambulation (Ljungberg et al., 2015), assessment of cord atrophy
in multiple sclerosis (Yiannakas et al., 2015), quantification of axon degeneration in cervical
spondylotic myelopathy (Grabher et al., 2016; A. R. Martin et al., 2016), multi-site DTI study of
the cervical spinal cord (Samson et al., 2016), mapping of T1, T2 and T2* relaxation times and
diffusion tensor metrics from 7T data (Massire et al., 2016).
With the emergence of international initiatives for testing new drugs for the spinal cord, the
standardization of mpMRI acquisition and processing methods for the spinal cord, combined
with common templates and atlases, will facilitate collaborative work and will provide guidelines
for future multi-center studies and clinical trials. For example, having a standard frame of
reference will make it possible to share results of unthresholded statistical maps of fMRI
experiments. Public repositories already exist for such initiatives17.
SCT is primarily developed and tested on human data, although some of its functions can also be
useful for non-human MRI data of the spinal cord. For example, the segmentation and the cord
straightening modules have already been successfully applied to animal data (unpublished
material).

5.4. Perspectives
Development of SCT is strongly tailored to researchers needs. As such, the MNI-Poly-AMU
template is being expanded to the brainstem and full spinal cord and now includes multiple MRI
contrasts: T1-, T2- and T2*-weighted. The spinal cord detection module (part of PropSeg) is in
constant evolution to be more robust and compatible with other MR contrast (e.g., diffusion- and
T2*-weighted). The two-landmark based registration to the template will be expended to higher
number of landmarks to allow more accurate correspondence with vertebral levels. These
improvements would lead to faster and more robust analysis of large databases.
Regarding white/gray matter segmentation, further tests are required in different pathological
cases (e.g., MS, spinal cord injury). The research community would greatly benefit from a shared
repository of pathological cases in order to make such validations.
The function-specific integrative testing implemented in SCT was designed to be modular, so
that researchers developing methods for spinal cord can easily reuse the proposed framework.
For example, if someone is developing another gray matter segmentation method, the same
testing framework can be applied as that already used for the gray matter segmentation method
implemented in SCT. In the long run, we hope this approach will facilitate cross-validation of
other methods by comparing the same metrics with the same database and gold standards.
Finally, we would like to stress that image processing is only half of the story, and that the
acquisition of good quality data and the standardization of protocols across vendors remains a
challenging task, which contributes to the difficulty to perform multi-center studies (Stroman et
al., 2014). The standardization of both acquisition and processing protocols for multi-parametric
MRI of the spinal cord is being tackled at the international level (Samson et al., 2016), and will

11[Link]
likely open promising avenues for cross-validating research outputs and for testing new
therapeutic strategies (Wheeler-Kingshott et al., 2014).

6. Conclusion
SCT is a comprehensive software package dedicated to the processing of mpMRI data of the
spinal cord. SCT is tailored towards standardization and automation of processing pipeline
(intuitive batch scripts), versatility (user-oriented development of new features) modularity
(possibility to reuse some SCT functions or to contribute to new features) and wide distribution
(open-source, extensive testing framework, active support via forum). Preliminary applications
of SCT cover a variety of studies, from cross-sectional area measures in a large database of
patients to the precise quantification of mpMRI metrics in specific spinal pathways. It is hoped
that SCT will contribute to bringing together the spinal cord neuroimaging community by
establishing standard templates and analysis procedures, similarly to when MRI brain software
were first introduced in the early 90’.

Acknowledgments
The authors would like to acknowledge all core contributors of SCT: Tanguy Duval, Charley
Gros, Pierre-Olivier Quirion, Julien Touati, Augustin Roux, Tanguy Magnan, Olivier Comtois,
Geoffrey Lévêque, Marc Benhamou and all collaborators who contributed to some of the tools:
Drs. Michael Fehlings, Allan Martin, David Cadotte, Adam Cadotte, Brian Avants, Manuel
Taso, Arnaud Le Troter and Michael Sdika. The following people are acknowledged for useful
discussions: Drs. Pierre Bellec, Eleftherios Garyfallidis, Maxime Descoteaux, Satrajit Ghosh,
Demian Wassermann, Yves Goussard and Serge Rossignol. Finally we are grateful to all beta-
testers and users who are contributing to the improvement of SCT.
The development of SCT is supported by the Canada Research Chair in Quantitative Magnetic
Resonance Imaging, the Canadian Institute of Health Research [CIHR FDN-143263], the Fonds
de Recherche du Québec - Santé [28826], the Fonds de Recherche du Québec - Nature et
Technologies [2015-PR-182754], the Natural Sciences and Engineering Research Council of
Canada [435897-2013], the Sensorimotor Rehabilitation Research Team (SMRRT), the
Functional Neuroimaging Unit (CRIUGM, Université de Montréal) and the Quebec BioImaging
Network.
References
Adams, R.D., Salam-Adams, M., 1991. Chronic nontraumatic diseases of the spinal cord.
Neurol. Clin. 9, 605–623.
Asman, A.J., Bryan, F.W., Smith, S. a., Reich, D.S., Landman, B. a., 2014. Groupwise multi-
atlas segmentation of the spinal cord’s internal structure. Med. Image Anal. 18, 460–471.
Avants, B.B., Epstein, C.L., Grossman, M., Gee, J.C., 2008. Symmetric diffeomorphic image
registration with cross-correlation: evaluating automated labeling of elderly and
neurodegenerative brain. Med. Image Anal. 12, 26–41.
Avants, B.B., Tustison, N.J., Song, G., Cook, P.A., Klein, A., Gee, J.C., 2011. A reproducible
evaluation of ANTs similarity metric performance in brain image registration. Neuroimage
54, 2033–2044.
Avants, Brian B., Nicholas J. Tustison, Michael Stauffer, Gang Song, Baohua Wu, and James C.
Gee. 2014. “The Insight ToolKit Image Registration Framework.” Frontiers in
Neuroinformatics 8 (April): 44.
Bozzo, A., Marcoux, J., Radhakrishna, M., Pelletier, J., Goulet, B., 2011. The role of magnetic
resonance imaging in the management of acute spinal cord injury. J. Neurotrauma 28,
1401–1411.
Bradbury, E.J., McMahon, S.B., 2006. Spinal cord repair strategies: why do they work? Nat.
Rev. Neurosci. 7, 644–653.
Brodmann, K., 1909. Vergleichende Lokalisationslehre der Grosshirnrinde in ihren Prinzipien
dargestellt auf Grund des Zellenbaues. Barth.
Cadotte, D.W., Cadotte, A., Cohen-Adad, J., Fleet, D., Livne, M., Wilson, J.R., Mikulis, D.,
Nugaeva, N., Fehlings, M.G., 2014. Characterizing the location of spinal and vertebral
levels in the human cervical spinal cord. AJNR Am. J. Neuroradiol. 36, 803–810.
Cadotte, D.W., Wilson, J.R., Mikulis, D., Stroman, P.W., Brady, S., Fehlings, M.G., 2011.
Conventional MRI as a diagnostic and prognostic tool in spinal cord injury: a systemic
review of its application to date and an overview on emerging MRI methods. Expert Opin.
Med. Diagn. 5, 121–133.
Castellano, A., Papinutto, N., Cadioli, M., Brugnara, G., Iadanza, A., Scigliuolo, G., Pareyson,
D., Uziel, G., Köhler, W., Aubourg, P., Falini, A., Henry, R.G., Politi, L.S., Salsano, E.,
2016. Quantitative MRI of the spinal cord and brain in adrenomyeloneuropathy: in vivo
assessment of structural changes. Brain. doi:10.1093/brain/aww068
Chang, H.-H., Zhuang, A.H., Valentino, D.J., Chu, W.-C., 2009. Performance measure
characterization for evaluating neuroimage segmentation algorithms. Neuroimage 47, 122–
135.
Ciccarelli, O., Wheeler-Kingshott, C.A., McLean, M.A., Cercignani, M., Wimpey, K., Miller,
D.H., Thompson, A.J., 2007. Spinal cord spectroscopy and diffusion-based tractography to
assess acute disability in multiple sclerosis. Brain 130, 2220–2231.
Cohen-Adad, J., Descoteaux, M., Rossignol, S., Hoge, R.D., Deriche, R., Benali, H., 2008.
Detection of multiple pathways in the spinal cord using q-ball imaging. Neuroimage 42,
739–749.
Cohen-Adad, J., Gauthier, C.J., Brooks, J.C.W., Slessarev, M., Han, J., Fisher, J.A., Rossignol,
S., Hoge, R.D., 2010. BOLD signal responses to controlled hypercapnia in human spinal
cord. Neuroimage 50, 1074–1084.
Cohen-Adad, J., Lévy, S., Avants, B., 2015. Slice-by-slice regularized registration for spinal cord
 MRI: SliceReg, in: Proceedings of the 23th Annual Meeting of ISMRM, Toronto, Canada.
Toronto, p. 5553.
Cohen-Adad, J., Mareyam, A., Keil, B., Polimeni, J.R., Wald, L.L., 2011. 32-channel RF coil
optimized for brain and cervical spinal cord at 3 T. Magn. Reson. Med. 66, 1198–1208.
Cohen-Adad, J., Wheeler-Kingshott, C. (Eds.), 2014. Quantitative MRI of the Spinal Cord.
Elsevier.
De Leener, B., Cohen-Adad, J., Kadoury, S., 2015. Automatic Segmentation of the Spinal Cord
and Spinal Canal Coupled With Vertebral Labeling. IEEE Trans. Med. Imaging 34, 1705–
1718.
De Leener, B., Kadoury, S., Cohen-Adad, J., 2014. Robust, accurate and fast automatic
segmentation of the spinal cord. Neuroimage 98, 528–536.
De Leener, B., Taso, M., Cohen-Adad, J., Callot, V., 2016. Segmentation of the human spinal
cord. MAGMA 29, 125–153.
Dowell, N.G., Jenkins, T.M., Ciccarelli, O., Miller, D.H., Wheeler- Kingshott, C.A.M., 2009.
Contiguous-slice zonally oblique multislice (CO-ZOOM) diffusion tensor imaging:
Examples of in vivo spinal cord and optic nerve applications. J. Magn. Reson. Imaging 29,
454–460.
Duval, T., McNab, J.A., Setsompop, K., Witzel, T., Schneider, T., Huang, S.Y., Keil, B.,
Klawiter, E.C., Wald, L.L., Cohen-Adad, J., 2015. In vivo mapping of human spinal cord
microstructure at 300mT/m. Neuroimage 118, 494–507.
Duval, T., S. Lévy, N. Stikov, J. Campbell, A. Mezer, T. Witzel, B. Keil, et al. 2016. “G-Ratio
Weighted Imaging of the Human Spinal Cord in Vivo.” NeuroImage, September.
doi:10.1016/[Link].2016.09.018
Eippert, Falk, Yazhuo Kong, Anderson M. Winkler, Jesper L. Andersson, Jurgen Finsterbusch,
Christian Buchel, Jonathan C. W. Brooks, and Irene Tracey. 2016. “Investigating Resting-
State Functional Connectivity in the Cervical Spinal Cord at 3T.” bioRxiv.
doi:10.1101/073569.
El Mendili, M.M., Cohen-Adad, J., Pelegrini-Issac, M., Rossignol, S., Morizot-Koutlidis, R.,
Marchand-Pauvert, V., Iglesias, C., Sangari, S., Katz, R., Lehericy, S., Benali, H., Pradat,
P.F., 2014. Multi-parametric spinal cord MRI as potential progression marker in
amyotrophic lateral sclerosis. PLoS One 9, e95516.
Evans, A.C., Marrett, S., Neelin, P., Collins, L., Worsley, K., Dai, W., Milot, S., Meyer, E., Bub,
D., 1992. Anatomical mapping of functional activation in stereotactic coordinate space.
Neuroimage 1, 43–53.
Finsterbusch, J., 2009. High-resolution diffusion tensor imaging with inner field-of-view EPI. J.
Magn. Reson. Imaging 29, 987–993.
Fonov, V., Evans, A.C., Botteron, K., Almli, C.R., McKinstry, R.C., Collins, D.L., Brain
Development Cooperative Group, 2011. Unbiased average age-appropriate atlases for
pediatric studies. Neuroimage 54, 313–327.
Fonov, V.S., Le Troter, A., Taso, M., De Leener, B., Lévêque, G., Benhamou, M., Sdika, M.,
Benali, H., Pradat, P.-F., Collins, D.L., Callot, V., Cohen-Adad, J., 2014. Framework for
integrated MRI average of the spinal cord white and gray matter: The MNI-Poly-AMU
template. Neuroimage 102P2, 817–827.
Foxley, S., Mollink, J., Ansorge, O., Scott, C., Jbabdi, S., Yates, R., De Luca, G., Miller, K.,
2015. Whole post-mortem spinal cord imaging with diffusion-weighted steady state free
precession at 7T. Proceedings of the 23th Annual Meeting of ISMRM, Toronto, Canada
 4429.
Frost, M.A., Goebel, R., 2012. Measuring structural-functional correspondence: spatial
variability of specialised brain regions after macro-anatomical alignment. Neuroimage 59,
1369–1381.
Grabher, P., Mohammadi, S., Trachsler, A., Friedl, S., David, G., Sutter, R., Weiskopf, N.,
Thompson, A.J., Curt, A., Freund, P., 2016. Voxel-based analysis of grey and white matter
degeneration in cervical spondylotic myelopathy. Sci. Rep. 6, 24636.
Grussu, F., Schneider, T., Zhang, H., Alexander, D.C., Wheeler-Kingshott, C.A.M., 2015.
Neurite orientation dispersion and density imaging of the healthy cervical spinal cord in
vivo. Neuroimage 111, 590–601.
Gullapalli, J., Krejza, J., Schwartz, E.D., 2006. In vivo DTI evaluation of white matter tracts in
rat spinal cord. J. Magn. Reson. Imaging 24, 231–234.
Hagler, D.J., Jr, Saygin, A.P., Sereno, M.I., 2006. Smoothing and cluster thresholding for
cortical surface-based group analysis of fMRI data. Neuroimage 33, 1093–1103.
Harlow, D.E., Honce, J.M., Miravalle, A.A., 2015. Remyelination Therapy in Multiple Sclerosis.
Front. Neurol. 6, 257.
Horsfield, M.A., Sala, S., Neema, M., Absinta, M., Bakshi, A., Sormani, M.P., Rocca, M.A.,
Bakshi, R., Filippi, M., 2010. Rapid semi-automatic segmentation of the spinal cord from
magnetic resonance images: application in multiple sclerosis. Neuroimage 50, 446–455.
Iglesias, C., Sangari, S., El Mendili, M.-M., Benali, H., Marchand-Pauvert, V., Pradat, P.-F.,
2015. Electrophysiological and spinal imaging evidences for sensory dysfunction in
amyotrophic lateral sclerosis. BMJ Open 5, e007659.
Jenkinson, M., Bannister, P., Brady, M., Smith, S., 2002. Improved optimization for the robust
and accurate linear registration and motion correction of brain images. Neuroimage 17,
825–841.
Kearney, H., Miller, D.H., Ciccarelli, O., 2015. Spinal cord MRI in multiple sclerosis—
diagnostic, prognostic and clinical value. Nat. Rev. Neurol. 11, 327–338.
Kiebel, S.J., Poline, J.B., Friston, K.J., Holmes, A.P., Worsley, K.J., 1999. Robust smoothness
estimation in statistical parametric maps using standardized residuals from the general
linear model. Neuroimage 10, 756–766.
Klawiter, E.C., Schmidt, R.E., Trinkaus, K., Liang, H.-F., Budde, M.D., Naismith, R.T., Song,
S.-K., Cross, A.H., Benzinger, T.L., 2011. Radial diffusivity predicts demyelination in ex
vivo multiple sclerosis spinal cords. Neuroimage 55, 1454–1460.
Klein, A., Andersson, J., Ardekani, B.A., Ashburner, J., Avants, B., Chiang, M.-C., Christensen,
G.E., Collins, D.L., Gee, J., Hellier, P., Song, J.H., Jenkinson, M., Lepage, C., Rueckert, D.,
Thompson, P., Vercauteren, T., Woods, R.P., Mann, J.J., Parsey, R.V., 2009. Evaluation of
14 nonlinear deformation algorithms applied to human brain MRI registration. Neuroimage
46, 786–802.
Kong, Y., Eippert, F., Beckmann, C.F., Andersson, J., Finsterbusch, J., Büchel, C., Tracey, I.,
Brooks, J.C.W., 2014. Intrinsically organized resting state networks in the human spinal
cord. Proc. Natl. Acad. Sci. U. S. A. 111, 18067–18072.
Kong, Y., Jenkinson, M., Andersson, J., Tracey, I., Brooks, J.C.W., 2012. Assessment of
physiological noise modelling methods for functional imaging of the spinal cord.
Neuroimage 60, 1538–1549.
Laule, C., Vavasour, I.M., Leung, E., Li, D.K.B., Kozlowski, P., Traboulsee, A.L., Oger, J.,
Mackay, A.L., Moore, G.R.W., 2011. Pathological basis of diffusely abnormal white
 matter: insights from magnetic resonance imaging and histology. Mult. Scler. 17, 144–150.
Lévy, S., Benhamou, M., Naaman, C., Rainville, P., Callot, V., Cohen-Adad, J., 2015. White
matter atlas of the human spinal cord with estimation of partial volume effect. Neuroimage
119, 262–271.
Lindberg, P.G., Feydy, A., Maier, M.A., 2010. White matter organization in cervical spinal cord
relates differently to age and control of grip force in healthy subjects. J. Neurosci. 30, 4102–
4109.
Ljungberg, E., Kolind, S., Tam, R., Freedman, M., Li, D.K., Traboulsee, A., 2015. Correcting
Cervical Spinal Cord Area for Cord Length Strengthens Correlation Between Atrophy and
Ambulation. doi:10.13140/RG.2.1.4225.4163
Luessi, F., Kuhlmann, T., Zipp, F., 2014. Remyelinating strategies in multiple sclerosis. Expert
Rev. Neurother. 14, 1315–1334.
Mac Donald, C.L., Dikranian, K., Bayly, P., Holtzman, D., Brody, D., 2007. Diffusion tensor
imaging reliably detects experimental traumatic axonal injury and indicates approximate
time of injury. J. Neurosci. 27, 11869–11876.
Martin, A.R., Aleksanderek, I., Cohen-Adad, J., Tarmohamed, Z., Tetreault, L., Smith, N.,
Cadotte, D.W., Crawley, A., Ginsberg, H., Mikulis, D.J., Fehlings, M.G., 2016. Translating
state-of-the-art spinal cord MRI techniques to clinical use: A systematic review of clinical
studies utilizing DTI, MT, MWF, MRS, and fMRI. Neuroimage Clin 10, 192–238.
Martin, A.R., De Leener, B., Aleksanderek, I., Cohen-Adad, J., Cadotte, D.W., Kalsi-Ryan, S.,
Tetreault, L., Crawley, A.P., Ginsberg, H., Mikulis, D., Fehlings, M.G., 2016. A
Prospective Longitudinal Study in Degenerative Cervical Myelopathy Using Quantitative
Microstructural MRI with Tract-Specific Metrics. Proceedings of the 24th Annual Meeting
of ISMRM, Singapore 0853.
Massire, A., Taso, M., Guye, M., Ranjeva, J.P., Callot, V., 2016. High-resolution quantitative
magnetic resonance imaging of the human cervical spinal cord at 7T. Proceedings of the
24th Annual Meeting of ISMRM, Singapore 1130.
McCormick, M., Liu, X., Jomier, J., Marion, C., Ibanez, L., 2014. ITK: enabling
reproducible research and open science. Front. Neuroinform. 8, 13.
Mohammadi, S., Freund, P., Feiweier, T., Curt, A., Weiskopf, N., 2013. The impact of post-
processing on spinal cord diffusion tensor imaging. Neuroimage 70, 377–385.
Mohammadi, S., Möller, H.E., Kugel, H., Müller, D.K., Deppe, M., 2010. Correcting eddy
current and motion effects by affine whole-brain registrations: evaluation of three-
dimensional distortions and comparison with slicewise correction. Magn. Reson. Med. 64,
1047–1056.
Mori, S., Oishi, K., Faria, A.V., 2009. White matter atlases based on diffusion tensor imaging.
Curr. Opin. Neurol. 22, 362–369.
Narayana, P.A., Grill, R.J., Chacko, T., Vang, R., 2004. Endogenous recovery of injured spinal
cord: longitudinal in vivo magnetic resonance imaging. J. Neurosci. Res. 78, 749–759.
Onu, M., Gervai, P., Cohen-Adad, J., Lawrence, J., Kornelsen, J., Tomanek, B., Sboto-
Frankenstein, U.N., 2010. Human cervical spinal cord funiculi: investigation with magnetic
resonance diffusion tensor imaging. J. Magn. Reson. Imaging 31, 829–837.
Qian, W., Chan, Q., Mak, H., Zhang, Z., Anthony, M., Yau, K.K., Khong, P., Chan, K.H., Kim,
M., 2011. Quantitative assessment of the cervical spinal cord damage in neuromyelitis
optica using diffusion tensor imaging at 3 Tesla. J. Magn. Reson. Imaging 33, 1312–1320.
Rowland, J.W., Hawryluk, G.W., Kwon, B., Fehlings, M.G., 2008. Current status of acute spinal
 cord injury pathophysiology and emerging therapies: promise on the horizon. Neurosurg.
Focus 25, E2.
Samson, R.S., Lévy, S., Schneider, T., Smith, A.K., Smith, S.A., Cohen-Adad, J., Gandini
Wheeler-Kingshott, C.A.M., 2016. ZOOM or Non-ZOOM? Assessing Spinal Cord
Diffusion Tensor Imaging Protocols for Multi-Centre Studies. PLoS One 11, e0155557.
Smith, S.A., Jones, C.K., Gifford, A., Belegu, V., Chodkowski, B., Farrell, J.A.D., Landman,
B.A., Reich, D.S., Calabresi, P.A., McDonald, J.W., van Zijl, P.C.M., 2010. Reproducibility
of tract- specific magnetization transfer and diffusion tensor imaging in the cervical spinal
cord at 3 tesla. NMR Biomed. 23, 207–217.
Smith, S.M., Jenkinson, M., Woolrich, M.W., Beckmann, C.F., Behrens, T.E., Johansen-Berg,
H., Bannister, P.R., De Luca, M., Drobnjak, I., Flitney, D.E., Niazy, R.K., Saunders, J.,
Vickers, J., Zhang, Y., De Stefano, N., Brady, J.M., Matthews, P.M., 2004. Advances in
functional and structural MR image analysis and implementation as FSL. Neuroimage 23
Suppl 1, S208–19.
Standring, S., 2008. Spinal cord: internal organization. Gray’s Anatomy, 40th Edition 257–274.
Stroman, P.W., Wheeler-Kingshott, C.A.M., Bacon, M., Schwab, J.M., Bosma, R., Brooks, J.,
Cadotte, D., Carlstedt, T., Ciccarelli, O., Cohen-Adad, J., Curt, A., Evangelou, N., Fehlings,
M.G., Filippi, M., Kelley, B.J., Kollias, S., Mackay, A., Porro, C.A., Smith, S., Strittmatter,
S.M., Summers, P., Tracey, I., 2014. The current state-of-the-art of spinal cord imaging:
methods. Neuroimage 84, 1070–1081.
Summers, P.E., Brooks, J., Cohen-Adad, J., 2014. Spinal Cord fMRI, in: Cohen-Adad, J.,
Wheeler-Kingshott, C.A.M. (Eds.), Quantitative MRI of the Spinal Cord. Elsevier, pp. 221–
236.
Talbott, J.F., Narvid, J., Chazen, J.L., Chin, C.T., Shah, V., 2016. An Imaging-Based Approach
to Spinal Cord Infection. Seminars in Ultrasound, CT and MRI.
doi:10.1053/[Link].2016.05.006
Taso, Manuel, Olivier M. Girard, Guillaume Duhamel, Arnaud Le Troter, Thorsten Feiweier,
Maxime Guye, Jean-Philippe Ranjeva, and Virginie Callot. 2016. “Tract-Specific and Age-
Related Variations of the Spinal Cord Microstructure: A Multi-Parametric MRI Study
Using Diffusion Tensor Imaging (DTI) and Inhomogeneous Magnetization Transfer
(ihMT).” NMR in Biomedicine 29 (6): 817–32.
Taso, M., Le Troter, A., Sdika, M., Cohen-Adad, J., Arnoux, P.-J., Guye, M., Ranjeva, J.-P.,
Callot, V., 2015. A reliable spatially normalized template of the human spinal cord -
Applications to automated white matter/gray matter segmentation and tensor-based
morphometry (TBM) mapping of gray matter alterations occurring with age. Neuroimage
117, 20–28.
Taso, M., Le Troter, A., Sdika, M., Ranjeva, J.-P., Guye, M., Bernard, M., Callot, V., 2014.
Construction of an in vivo human spinal cord atlas based on high-resolution MR images at
cervical and thoracic levels: preliminary results. Magn. Reson. Mater. Phys. Biol. Med. 27,
257–267.
Tofts, P., 2003. Quantitative MRI of the Brain: Measuring Changes Caused by Disease.
Tustison, N.J., Avants, B.B., 2013. Explicit B-spline regularization in diffeomorphic image
registration. Front. Neuroinform. 7, 39.
Ullmann, E., Paquette, J.F.P., Thong, W.E., Cohen-Adad, J., 2014. Automatic labeling of
vertebral levels using a robust template-based approach. Int. J. Biomed. Imaging 2014,
719520.
van Middendorp, J.J., Goss, B., Urquhart, S., Atresh, S., Williams, R.P., Schuetz, M., 2011.
Diagnosis and prognosis of traumatic spinal cord injury. Global Spine J 1, 1–8.
Weber, K.A., 2nd, Chen, Y., Wang, X., Kahnt, T., Parrish, T.B., 2015. Lateralization of cervical
spinal cord activity during an isometric upper extremity motor task with functional
magnetic resonance imaging. Neuroimage 125, 233–243.
Weber, Kenneth A., 2nd, Yufen Chen, Xue Wang, Thorsten Kahnt, and Todd B. Parrish. 2016.
“Functional Magnetic Resonance Imaging of the Cervical Spinal Cord During Thermal
Stimulation Across Consecutive Runs.” NeuroImage, September.
doi:10.1016/[Link].2016.09.015
West, J., Aalto, A., Tisell, A., Leinhard, O.D., Landtblom, A.-M., Smedby, Ö., Lundberg, P.,
2014. Normal appearing and diffusely abnormal white matter in patients with multiple
sclerosis assessed with quantitative MR. PLoS One 9, e95161.
Wheeler-Kingshott, C.A., Stroman, P.W., Schwab, J.M., Bacon, M., Bosma, R., Brooks, J.,
Cadotte, D.W., Carlstedt, T., Ciccarelli, O., Cohen-Adad, J., Curt, A., Evangelou, N.,
Fehlings, M.G., Filippi, M., Kelley, B.J., Kollias, S., Mackay, A., Porro, C.A., Smith, S.,
Strittmatter, S.M., Summers, P., Thompson, A.J., Tracey, I., 2014. The current state-of-the-
art of spinal cord imaging: applications. Neuroimage 84, 1082–1093.
Wilm, B.J., Svensson, J., Henning, A., Pruessmann, K.P., Boesiger, P., Kollias, S.S., 2007.
Reduced field-of-view MRI using outer volume suppression for spinal cord diffusion
imaging. Magn. Reson. Med. 57, 625–630.
Xu, J., Shimony, J.S., Klawiter, E.C., Snyder, A.Z., Trinkaus, K., Naismith, R.T., Benzinger,
T.L., Cross, A.H., Song, S.K., 2013. Improved in vivo diffusion tensor imaging of human
cervical spinal cord. Neuroimage 67, 64–76.
Yiannakas, M.C., Kearney, H., Samson, R.S., Chard, D.T., Ciccarelli, O., Miller, D.H., Wheeler-
Kingshott, C. a. M., 2012. Feasibility of grey matter and white matter segmentation of the
upper cervical cord in vivo: a pilot study with application to magnetisation transfer
measurements. Neuroimage 63, 1054–1059.
Yiannakas, M.C., Mustafa, A.M., De Leener, B., Kearney, H., Tur, C., Altmann, D.R., De
Angelis, F., Plantone, D., Ciccarelli, O., Miller, D.H., Cohen-Adad, J., Wheeler-Kingshott,
C.A.M., 2015. Fully automated segmentation of the cervical cord from T1-weighted MRI
using PropSeg: Application to multiple sclerosis. NeuroImage: Clinical.
doi:10.1016/[Link].2015.11.001
Zhang, J., Jones, M., DeBoy, C.A., Reich, D.S., Farrell, J.A., Hoffman, P.N., Griffin, J.W.,
Sheikh, K.A., Miller, M.I., Mori, S., Calabresi, P.A., 2009. Diffusion tensor magnetic
resonance imaging of Wallerian degeneration in rat spinal cord after dorsal root axotomy. J.
Neurosci. 29, 3160–3171.
Zhao, W., Cohen-Adad, J., Polimeni, J.R., Keil, B., Guerin, B., Setsompop, K., Serano, P.,
Mareyam, A., Hoecht, P., Wald, L.L., 2014. Nineteen-channel receive array and four-
channel transmit array coil for cervical spinal cord imaging at 7T. Magn. Reson. Med. 72,
291–300.
9. Appendix
The data used for validation were acquired at two sites (UNF-Montreal and CRMBM-Marseille)
on a 3T system (TIM Trio, Siemens Healthcare). Acquisition parameters are: slab-selective fast
spin echo, TR = 1500 ms, TE = 119 ms, flip angle = 140°, bandwidth = 723 Hz/voxel, voxel size
= 1x1x1 mm3, acquisition time = 6 min. Coverage: brainstem → T5. Note that subjects used for
validation were different than subjects used to construct the template and GM model.

9.1. Validation of spinal cord segmentation (sct_propseg)

Spinal Cord Toolbox (version dev-9af778d77ef6527478a1ad3f8f2179f6e277235c)
Running: /Users/benjamindeleener/code/spinalcordtoolbox/scripts/isct_test_function.py
-f sct_propseg -d /Volumes/data_shared/sct_testing/article_sct/ -p "-i t2/[Link] -c
t2"
OS: osx (Darwin-15.0.0-x86_64-i386-64bit), Hostname: bendeleener, CPU available/used:
8/8, RAM: 16.00 gigabytes. Testing... (started on: 2016-05-30 [Link])

subject dice_segmentation duration [s] status

0 ALT 0.974909 59.192263 0
1 AM 0.979136 25.009005 0
2 ED 0.974385 30.344065 0
3 errsm_03 0.978930 30.219893 0
4 errsm_04 0.980217 33.304414 0
5 errsm_05 0.984458 28.966360 0
6 errsm_09 0.985376 31.513813 0
7 errsm_10 0.985270 30.059773 0
8 errsm_11 0.963750 33.417223 0
9 errsm_12 0.978226 28.130072 0
10 errsm_13 0.982187 34.205332 0
11 errsm_14 0.970628 29.322435 0
12 errsm_16 0.975682 33.403359 0
13 errsm_24 0.984579 31.423410 0
14 errsm_25 0.972632 33.938798 0
15 errsm_31 0.979778 31.034228 0
16 errsm_35 0.978907 26.739745 0
17 errsm_36 0.948414 34.814216 0
18 errsm_37 0.967442 26.488759 0
19 errsm_44 0.973496 31.830878 0
20 GB 0.972811 52.091329 0
21 HB 0.937498 24.502725 0
22 JW 0.981558 56.317613 0
23 MT 0.970099 37.002365 0
24 PA 0.958671 23.217906 0
25 pain_pilot_1 0.981062 33.690780 0
26 pain_pilot_2 0.979456 31.673408 0
27 pain_pilot_3 0.981869 27.859988 0
28 pain_pilot_4 0.970748 31.189391 0
29 pain_pilot_7 0.970450 31.905474 0
30 sct_001 0.981768 31.928909 0
31 sct_002 0.974058 23.013328 0
32 T045 0.951027 34.562992 0
33 T047 0.983665 10.713219 0
34 VC 0.972095 19.746288 0
Mean 0.973864 31.793536
STD 0.010873 9.033228
Passed: 35/35
Total duration: 168s
Status legend: 0: Passed, 1: Crashed, 99: Failed, 200: File(s) missing
Table 3. Results of validation for sct_propseg. Metric (with pass/fail threshold indicated in
brackets) is: dice_segmentation (>0.9): Dice coefficient between manual cord segmentation and
automatic cord segmentation. Note that the Dice coefficients are overall larger than in the
previous studies (B. De Leener et al., 2015, 2014) (0.97 vs. 0.9), which is explained by the
difference in coverage between the two studies (C1-T5 here vs. C1-T12). Since PropSeg is based
on a propagation of a 3d mesh, it is subject to propagation errors which accumulate over a large
distance. Moreover, given that the cord diameter decreases at the caudal levels, the Dice
coefficient is more sensitive to discrepancies caused by binarization (i.e., inclusion/exclusion of
cord voxel in manual and automatic segmentation), as pointed out in (Chang et al., 2009).

9.2. Validation of spinal cord straightening (sct_straighten_spinalcord)

Spinal Cord Toolbox (version dev-9af778d77ef6527478a1ad3f8f2179f6e277235c)
Running: /Users/benjamindeleener/code/spinalcordtoolbox/scripts/isct_test_function.py
-f sct_straighten_spinalcord -d /Volumes/data_shared/sct_testing/article_sct/ -p "-i
t2/[Link] -s t2/t2_seg_manual.[Link]"
OS: osx (Darwin-15.0.0-x86_64-i386-64bit), Hostname: bendeleener, CPU available/used:
8/8, RAM: 16.00 gigabytes. Testing... (started on: 2016-05-30 [Link])

subject dice dist_max duration rmse status

0 ALT 0.964091 1.210000 138.035741 0.770000 0
1 AM 0.964272 1.240000 129.986078 0.750000 0
2 ED 0.940310 1.220000 136.400379 0.570000 0
3 errsm_03 0.974188 1.360000 169.131326 0.800000 0
4 errsm_04 0.973933 1.350000 157.414294 0.770000 0
5 errsm_05 0.963595 1.490000 148.234811 0.760000 0
6 errsm_09 0.942894 1.240000 161.720111 0.640000 0
7 errsm_10 0.970445 1.360000 178.990922 0.860000 0
8 errsm_11 0.965220 1.560000 162.602842 0.760000 0
9 errsm_12 0.971098 1.250000 154.384525 0.770000 0
10 errsm_13 0.973564 1.360000 179.241509 0.780000 0
11 errsm_14 0.967366 1.170000 145.726869 0.760000 0
12 errsm_16 0.969258 1.300000 167.700308 0.790000 0
13 errsm_24 0.976938 1.320000 162.213312 0.800000 0
14 errsm_25 0.977069 1.160000 160.847801 0.830000 0
15 errsm_31 0.965974 1.410000 180.545414 0.800000 0
16 errsm_35 0.940472 1.110000 154.124630 0.610000 0
17 errsm_36 0.954251 1.460000 175.945490 0.790000 0
18 errsm_37 0.975308 1.390000 147.528859 0.740000 0
19 errsm_44 0.957694 1.630000 160.090363 0.810000 0
20 GB 0.968203 1.300000 122.859090 0.830000 0
21 HB 0.977198 1.370000 135.946164 0.790000 0
22 JW 0.976327 1.280000 126.539476 0.680000 0
23 MT 0.958034 1.240000 121.892832 0.750000 0
24 PA 0.965978 1.360000 104.423937 0.760000 0
25 pain_pilot_1 0.962826 1.200000 166.212215 0.690000 0
26 pain_pilot_2 0.970687 1.410000 177.270257 0.760000 0
27 pain_pilot_3 0.967363 1.280000 147.872101 0.870000 0
28 pain_pilot_4 0.966756 1.460000 165.254728 0.790000 0
29 pain_pilot_7 0.969397 1.420000 154.295965 0.770000 0
30 sct_001 0.974191 1.200000 163.630688 0.740000 0
31 sct_002 0.970683 1.360000 141.603322 0.790000 0
32 T045 0.967967 1.340000 99.864195 0.690000 0
33 T047 0.978023 1.260000 47.066873 0.780000 0
34 VC 0.970585 1.290000 90.282064 0.800000 0
Mean 0.966633 1.324571 146.739414 0.761429
STD 0.009748 0.113716 28.525739 0.063392
Passed: 35/35
Total duration: 768s
Table 4. Results of validation for sct_straighten_spinalcord. Metrics (with pass/fail threshold
indicated in brackets) are: dice (>0.9): Dice coefficient between original cord segmentation
and cord segmentation following application of warping fields: curved-to-straight then straight-
to- curved; dist_max (<2mm): maximum distance error between straightened cord centerline
and straight line; mse (<1mm): Root mean square error between straightened cord centerline
and straight line.

9.3. Validation of template registration (sct_register_to_template)

Spinal Cord Toolbox (version dev-9af778d77ef6527478a1ad3f8f2179f6e277235c)
Running: /Users/benjamindeleener/code/spinalcordtoolbox/scripts/isct_test_function.py
-f sct_register_to_template -d /Users/benjamindeleener/data/sct_article/article_sct -p
"-i t2/[Link] -s t2/t2_seg_manual.[Link] -l t2/[Link]"
OS: osx (Darwin-15.0.0-x86_64-i386-64bit), Hostname: bendeleener, CPU available/used:
8/8, RAM: 16.00 gigabytes. Testing... (started on: 2016-05-30 [Link])

subject dice_anat2template dice_template2anat duration [s] status

0 ALT 0.936628 0.953276 1351.117136 0
1 AM 0.936680 0.944771 1178.249453 0
2 ED 0.928370 0.930995 1346.815756 0
3 errsm_03 0.938017 0.956620 1280.668386 0
4 errsm_04 0.938522 0.957343 1405.813964 0
5 errsm_05 0.940862 0.960945 1252.367538 0
6 errsm_09 0.935940 0.955477 1395.179839 0
7 errsm_10 0.943405 0.958073 1269.222949 0
8 errsm_11 0.938392 0.943425 1416.074127 0
9 errsm_12 0.938000 0.957287 1245.334603 0
10 errsm_13 0.936076 0.957255 1423.625025 0
11 errsm_14 0.941128 0.954045 1233.543628 0
12 errsm_16 0.934081 0.960331 1392.906353 0
13 errsm_24 0.940717 0.952959 1271.990924 0
14 errsm_25 0.937754 0.960685 1388.556740 0
15 errsm_31 0.939569 0.952918 1282.267960 0
16 errsm_35 0.938398 0.933919 1249.511212 0
17 errsm_36 0.932594 0.943117 1239.583563 0
18 errsm_37 0.931451 0.947375 1237.050915 0
19 errsm_44 0.931822 0.945038 1238.792484 0
20 GB 0.933848 0.949117 1195.073648 0
21 HB 0.939995 0.959203 1165.580928 0
22 JW 0.936659 0.955126 1194.341909 0
23 MT 0.935811 0.947558 1158.239632 0
24 PA 0.928328 0.946324 1179.257488 0
25 pain_pilot_1 0.941512 0.946667 1230.404822 0
26 pain_pilot_2 0.938875 0.948232 1248.543263 0
27 pain_pilot_3 0.941085 0.954422 1228.497101 0
28 pain_pilot_4 0.937990 0.952063 1237.206240 0
29 pain_pilot_7 0.938323 0.952721 1230.258578 0
30 sct_001 0.935758 0.951419 1234.360480 0
31 sct_002 0.935993 0.948645 1217.444211 0
32 T045 0.935448 0.960071 447.979074 0
33 T047 0.938260 0.957898 246.682566 0
34 VC 0.935728 0.954079 437.265764 0
Mean 0.936915 0.951697 1192.851665
STD 0.003478 0.007098 265.488879
Passed: 35/35
Total duration: 5773s
Table 5. Results of validation for sct_register_to_template. Metrics (with pass/fail threshold
indicated in brackets) are: dice_anat2template (>0.9): Dice coefficient between template cord
segmentation and subject cord segmentation following application of warping field anat-to-
template; dice_template2anat (>0.9): Dice coefficient between subject cord segmentation and
template cord segmentation following application of warping field template-to-anat.

9.4. Validation of vertebral labeling (sct_label_vertebrae)

Spinal Cord Toolbox (version dev-63ac386913fc2e047d7961cb1dd68b2f5cbbdc0b)
Running: /Users/jcohen/code/spinalcordtoolbox/scripts/isct_test_function.py -f
sct_label_vertebrae -d /Volumes/data_raid/data_shared/sct_testing/article_sct/ -p "-i
t2/[Link] -s t2/t2_seg_manual.[Link] -o t2_seg_labeled.[Link]"
OS: osx (Darwin-14.4.0-x86_64-i386-64bit), Hostname: django, CPU available/used: 8/8,
RAM: 16.00 gigabytes. Testing... (started on: 2016-05-31 [Link])

subject diff_man duration [s] max_dist rmse n_vert status

0 ALT 0 252.061417 3.000000 1.138550 11 0
1 AM 0 234.702592 3.316625 1.465656 11 0
2 ED 0 222.393800 2.000000 0.962250 11 0
3 errsm_03 0 247.364099 2.828427 1.170628 11 0
4 errsm_04 0 274.102957 3.162278 1.290994 11 0
5 errsm_05 0 240.327429 3.464102 1.539601 11 0
6 errsm_09 0 258.832062 2.828427 1.201850 11 0
7 errsm_10 0 271.705949 3.316625 1.360828 11 0
8 errsm_11 0 265.095368 3.741657 1.374369 11 0
9 errsm_12 0 245.545835 3.741657 1.742710 11 0
10 errsm_13 0 283.595454 2.449489 1.071517 11 0
11 errsm_14 0 227.338844 3.000000 1.232282 11 0
12 errsm_16 0 272.626050 3.316625 1.401058 11 0
13 errsm_24 0 259.355719 2.449490 1.122167 11 0
14 errsm_25 0 260.235856 3.000000 1.347151 11 0
15 errsm_31 0 274.250930 6.403124 2.177324 nan 99
16 errsm_35 0 254.338023 2.236068 1.036375 11 0
17 errsm_36 0 250.107030 3.316625 1.305260 11 0
18 errsm_37 0 242.177974 3.316625 1.088662 11 0
19 errsm_44 0 251.942417 3.605551 1.440165 11 0
20 GB 0 211.370936 2.449490 1.054093 11 0
21 HB 0 191.212892 2.449490 1.088662 11 0
22 JW 0 219.932286 2.449490 0.981307 11 0
23 MT 0 192.001864 2.449490 1.071517 11 0
24 PA 0 191.563530 2.449490 0.942809 11 0
25 pain_pilot_1 0 251.741721 3.162278 1.217161 11 0
26 pain_pilot_2 0 289.184444 2.449490 1.186342 11 0
27 pain_pilot_3 0 211.744835 3.000000 1.401058 11 0
28 pain_pilot_4 0 286.991825 3.000000 1.290994 11 0
29 pain_pilot_7 0 215.993892 3.000000 1.201850 11 0
30 sct_001 0 280.821177 3.316625 1.261980 11 0
31 sct_002 0 207.100989 2.449490 1.088662 11 0
32 T045 0 135.264882 3.000000 1.154701 11 0
33 T047 0 70.504924 2.449490 1.122167 11 0
34 VC 0 117.523521 2.236068 1.000000 11 0
Mean 0 233.172958 2.994408 1.243791
STD 0 48.345331 0.750945 0.240435
Passed: 34/35
Total duration: 1130s
Status legend: 0: Passed, 1: Crashed, 99: Failed, 200: File(s) missing
Table 6. Results of validation for sct_label_vertebrae. Metrics (with pass/fail threshold
indicated in brackets) are: diff_man (<3): Number of labels mismatch between results and
gold-standard (i.e., manual labeling of vertebral body center); max_dist (<4mm): Maximum
Frobenius label distance between results and gold-standard; rmse (<2mm): Root mean square
error of pairwise labels between results and gold-standard. One subject did not pass the test
(errsm_31) because of a lack of contrast between intervertebral discs and vertebral bodies. The
number of identified vertebral levels is indicated in the column n_vert.

9.5. Details and validation of gray matter segmentation

Methods: The method of Asman et al. (Asman et al., 2014) was implemented and new features
were added, including (i) the possibility to add prior information about the vertebral level in
order to gain accuracy on the target shape of the gray matter, (ii) the possibility to segment
images from any contrast by adding linear normalization between median intensity values of
white/gray matter based on the spinal cord internal structure pre-registration (combining our
multi-atlas-based model and vertebral level information), and (iii) the possibility to output
probabilistic segmentations of white/gray matter. The dictionary model was built from T2*-
weighted images of 37 healthy adult subjects (TIM Trio, Siemens Healthcare, axial orientation,
TR = 540 ms, TE =[5.41, 12.56, 19.16] ms, flip angle = 35°, bandwidth = 200 Hz/voxel,
resolution = 0.5×0.5×5mm3, coverage from C1 to T2 vertebral levels). Following gray matter
segmentation, SCT users can choose to improve the template-based registration as illustrated in
Figure 2 (purple color).
Validation: 10 subjects (different than the ones used to generate the model) were scanned at 3T
with the same acquisition parameters. Processing included the following steps: semi-automatic
cord segmentation, registration to template (to obtain vertebral level) and automatic gray matter
segmentation using multi-atlas framework. Manual segmentations of the spinal cord gray and
white matter were performed pixel-by-pixel using fslview using standardized procedure
(Yiannakas et al., 2012). The automatic probabilistic multi-atlas based segmentations were
thresholded at 0.5 and compared with manual segmentations using 2D Dice coefficient on white
and gray matter, along with Hausdorff distance and maximum median distance on skeletonized
segmentations.
Figure 8 shows examples of fully automatic gray matter segmentation results compared to
manual segmentations for T2*-w and MT images. Validation metrics are given hereafter, as
mean ± STD across ten subjects. Dice coefficient for the gray matter was 0.70 ± 0.06 for T2* and
0.66 ± 0.06 for MT. Dice coefficient for the white matter was 0.90 ± 0.02 for T 2* and 0.89 ± 0.02
for MT. Hausdorff distance was 1.20 ± 0.18 mm for T 2* and 1.64 ± 0.11 mm for MT. Median
distance was 0.34 ± 0.06 mm for T2* and 0.61 ± 0.05 mm for MT.
Figure 8: Automatic GM segmentation results for T2*-w (top) and MT (bottom) images. The
probabilistic automatic segmentations (right column) were thresholded at 0.5 and compared to
manual segmentations (middle column). Numerical results are displayed for each slice as well
as averaged across ten healthy subjects.

Thanks to the normalization feature, gray matter segmentation also works on images with
different contrast. As a proof-of-concept, diffusion-weighted images were acquired in 10
additional subjects. Acquisition parameters were: Siemens TIM Trio, 4ch neck coil, 2DRF
monopolar diffusion-weighted sequence (Finsterbusch et al., 2009), TR=820ms, TE=88ms,
matrix=208×42, resolution=0.8×0.8×5 mm3, 4 slices positioned orthogonal to the cord and
centered in the middle of the vertebral body (gap was adjusted per subject), b-value = 800 s/mm2,
24 directions, cardiac gated. Preprocessing consisted in motion correction as implemented in
SCT. The mean diffusion-weighted data, which exhibits inverted contrast compared to the T 2*-
weighted data (white matter bright, gray matter dark), was used for segmenting the gray matter.
Following segmentation, the warping field was updated, diffusion tensor were calculated, then
FA from all subjects was warped to the MNI-Poly-AMU template. Figure 9 (top panel) shows
mean FA across subjects, with an overlay of five spinal tracts. Figure 9 (bottom panel) shows the
mean ± STD values for FA, averaged across 10 subjects. Values are consistent with the literature
(Samson et al., 2016). Further validation is required to assess the specific benefits of accounting
for the gray matter for template-based analysis. It is likely that in the healthy adult population
which is well represented by the MNI-Poly-AMU atlas, there is a moderate benefit of using gray
matter segmentation to improve template registration. However, it is anticipated that in
pathological cases (e.g., where the shape of the gray matter does not follow the typical trends),
using subject-specific shape of the gray matter might improve the accuracy of template
registration.
Figure 9: Top panel shows FA maps registered to the MNI-Poly-AMU template (including gray
matter segmentation and registration), and averaged across ten healthy subjects. Five white
matter tracts are overlaid: gracilis (blue), cuneatus (yellow), corticospinal (green),
spinocerebellar (pink), and lemniscus (red). Bottom panel reports FA within specific spinal
tracts (mean ± STD across ten subjects).

9.6. Details and validation of SliceReg

Algorithm: Translations along x and y are estimated slice-by-slice using the convergence
framework implemented in ITK ([Link]). The cost function includes a regularization term
expressed as a polynomial function along z (assumed to be the spinal cord axis). Image metric
(cross-correlation, mutual information and mean squares), shrink factor (data subsampling),
smoothing and degree of polynomial function used for regularization can be specified. The
software outputs the registered image, forward and backward warping fields as well as a csv file
with x and y translations per slice.
Validation: To evaluate the performance of SliceReg in a large variety of sequences and
acquisition setups, data were acquired in 25 subjects in five different centers: Unité de
Neuroimagerie Fonctionnelle (Montreal, n=6), A.A. Martinos Center for Biomedical Imaging
(Boston, n=5), FMRIB (Oxford, n=4), Toronto Western Hospital (Toronto, n=5) and Pitié-
Salpêtrière Hospital (Paris, n=5). Sequences used were: diffusion-weighted EPI (n=7), spin echo
EPI (n=10), gradient echo EPI (n=6) and magnetization transfer gradient echo FLASH (n=2)
with variable amount of noise (different coils were used), spatial resolution (0.8 to 1.2mm in-
plane) and coverage (varied between 4 to 7 vertebral levels). For each sequence, two volumes
showing non-rigid deformations were selected as candidates for co-registration. SliceReg (3rd
order polynomial) was compared against ANTs (Avants et al., 2011) and FSL FLIRT (Jenkinson
et al., 2002), the two latter being constrained to two degrees of freedom (in-plane translations).
Mean square metric was used for all methods. Following within-pair registrations, the spinal
cord was segmented automatically using PropSeg (B. De Leener et al., 2014) and the estimated
transformations were applied to the segmentations using nearest neighbor interpolation. Overlap
between the two registered segmentations was assessed using Dice coefficients.
Results: Figure 10 compares the performance of SliceReg, ANTs 2D and FLIRT 2D for co-
registration of two volumes (within modality). Average Dice (N=25 subjects) was respectively
0.87±0.06, 0.83±0.10 and 0.81±0.10 for SliceReg, ANTs and FLIRT. Student’s t-test (paired for
dataset) showed significant difference between SliceReg and ANTs (p=0.004) and between
SliceReg and FLIRT (p=0.00007). Average computation times were 0.92ms, 0.65ms and 0.63ms
for SliceReg, ANTs and FLIRT, respectively. Despite the ~40% increase of computation time
compared to volume-based methods (because one transformation per slice is estimated as
opposed to a single one), the average processing time is below 1s, making it suitable for motion
correction pipelines.

Figure 10. Results of registration accuracy for SliceReg, ANTs 2D and FLIRT 2D methods
(N=25 subjects). Student’s t-test shows significantly higher Dice for SliceReg in comparison
with ANTs (p=0.004) and FLIRT (p=0.00007).

9.7. Validation of motion correction

Motion correction for diffusion-weighted MRI data was compared against FSL (mcflirt) and the
ACID Toolbox (Mohammadi et al., 2010) in 10 healthy subjects. Acquisition parameters are
listed in section 9.5. Details and validation of gray matter segmentation. Following motion
correction, diffusion tensors were computed and FA was estimated using ‘map’ method in the
spinal cord white matter across slices C2 to C5. Then, mean and standard deviation were
computed across subjects. For details the reader is referred to the validation script 18. Results are
presented in Table 7. sct_dmri_moco led to the highest mean FA across subjects, suggesting

12[Link]
that motion correction reduced blurriness in voxels at the periphery of the cord, which are more
prone to lower FA due to partial volume with CSF (where FA is close to zero). sct_dmri_moco
also led to the smallest standard-deviation across subjects. This comparison should be interpreted
with care, as more advanced approaches could have been with either of the software tested. For
example, mcflirt could be replaced with 2D slice-by-slice FLIRT, although this would have
required a dedicated script.
Mean SD
No motion correction 0.669 0.073
FSL mcflirt 0.633 0.083
ACID toolbox 0.680 0.078
sct_dmri_moco 0.703 0.056

Table 7. FA averaged across 10 subjects in the white matter from C2 to C5. Comparison of
motion correction.

9.8. Validation of metric quantification (sct_extract_metric)

We evaluated the effect of spatial resolution on the accuracy and precision of metric
quantification. The same simulation framework has been used as that one from Lévy et al., (Levy
et al., 2015), which investigated the effect of noise, across-tract variability and extraction method
on the accuracy and precision. In brief, a synthetic phantom with known values within each tract
was created from the WM atlas of the MNI-Poly-AMU template. The phantom consisted of 101
slices covering C7 to T1 vertebral levels at a resolution of 0.5 mm isotropic. Each tract was
assigned a random value between 35 and 45. A fixed value of ‘35’ was set in the gray matter,
and of ‘5’ in the CSF. After assigning a value for each tract, for the gray matter and for the CSF,
all labels were summed to yield a single 3D volume. Gaussian noise was then added to match an
SNR of 10 in the white matter. The absolute error between the true value and the estimated
metric was calculated for the following methods: binary (bin), weighted-average (wa), maximum
likelihood (ml), and maximum a posteriori (map). In order to remove the bias associated with the
choice of the value for a given tract, the experiment was repeated 200 times, each time with a
new randomly-generated phantom according to the bootstrapping method. The mean and STD of
absolute errors across bootstrap experiments were calculated. In order to investigate the effect of
spatial resolution, the bootstrapping experiment was repeated after downsampling the phantom at
the following resolutions: 0.8x0.8x5mm, 1x1x5mm and 1.5x1.5x5mm. The validation script and
data are accessible19.
Figure 11 shows results of the metric extraction for each resolution and each method. The key
aspects of these results are that: (i) methods ml and map perform best at all tested resolutions, (ii)
assuming an SNR of 10, to obtain less than 1% and 2% of quantification error the voxel fraction
should be at least 240 and 30, respectively (these numbers were upper-rounded from the graph
for the method map), (iii) the method ml provides high accuracy with at high voxel fraction (e.g.,
fasciculus cuneatus at 0.8mm in-plane over 20 slices), however it diverges dramatically when
voxel fraction decreases after a certain point, due to instabilities caused by noise (e.g., voxel
fraction <5). This motivates the use of the map method.

13[Link]

40
Figure 11. Simulation results comparing the accuracy and precision of metric extraction
methods at various resolutions. Metric extraction methods are: averaging within binary mask
(bin), weighted-averaging within probabilistic mask (wa), maximum likelihood estimation (ml)
and maximum a posteriori estimation (map) (Lévy et al., 2015). The abscissa shows different left
tracts, ordered by their volume fraction occupied within the white matter. The volume fraction of
a given tract (expressed here in percentage) represents the number of voxels of this tract divided
by the number of voxels in the white matter (each voxel being represented by its partial volume
information). An image of a randomly-generated synthetic phantom is shown at each resolution.
The ordinates show the mean and STD of the absolute error (in percentage of the true value).
Note that for very low volume fractions, the bin method is not represented because the tract is
too small to pass the 0.5 threshold.
Vi
ew
pu
bli
cat
ion
sta
ts

Segmentation Template and atlas

T2w
T1w
C1

C3

C5

T1

SCT Motion correction

Registration framework • 2D slice-by-slice
• Regularized across
slices & time
• Robust for
DWI (group-
wise)

Atlas-based analysis

and many more

features…