Segmentation

J Biomed Phys Eng 2014; 4(1)
www.jbpe.org
Quantitative Comparison of SPM, FSL,
and Brainsuite for Brain MR Image
Segmentation
Kazemi K.1*
, Noorizadeh N.1
1
Department of Elec-
trical and Electronics
Engineering, Shiraz
University of Technology,
Shiraz, Iran
*Corresponding author:
Kazemi K.
Department of Electrical
and Electronics Engi-
neering
Shiraz University of
Technology
Shiraz, Iran
E-Mail: kazemi@sutech.
ac.ir
ABSTRACT
Background: Accurate brain tissue segmentation from magnetic resonance (MR)
images is an important step in analysis of cerebral images. There are software packag-
es which are used for brain segmentation. These packages usually contain a set of skull
stripping, intensity non-uniformity (bias) correction and segmentation routines. Thus,
assessment of the quality of the segmented gray matter (GM), white matter (WM) and
cerebrospinal fluid (CSF) is needed for the neuroimaging applications.
Methods: In this paper, performance evaluation of three widely used brain segmen-
tation software packages SPM8, FSL and Brainsuite is presented. Segmentation with
SPM8 has been performed in three frameworks: i) default segmentation, ii) SPM8
New-segmentation and iii) modified version using hidden Markov random field as
implemented in SPM8-VBM toolbox.
Results: The accuracy of the segmented GM, WM and CSF and the robustness of
the tools against changes of image quality has been assessed using Brainweb simulated
MR images and IBSR real MR images. The calculated similarity between the seg-
mented tissues using different tools and corresponding ground truth shows variations
in segmentation results.
Conclusion: A few studies has investigated GM, WM and CSF segmentation. In
these studies, the skull stripping and bias correction are performed separately and they
just evaluated the segmentation. Thus, in this study, assessment of complete segmen-
tation framework consisting of pre-processing and segmentation of these packages is
performed. The obtained results can assist the users in choosing an appropriate seg-
mentation software package for the neuroimaging application of interest.
Keywords
MRI, Brain, Segmentation, SPM, FSL, Brainsuite
Introduction
W
ith progress in magnetic resonance (MR) imaging techniques,
structural and functional brain imaging are playing an impor-
tant role in neuroscience and experimental medicine. Increas-
ing anatomical scans of the human brain have emerged automated cere-
bral MR image analysis tools. These tools are applied to characterizing
differences in the shape and neuroanatomical configuration of different
brains. In particular, qualitative or quantitative information extraction
from cerebral MR images relies on accurate and reliable brain tissue
segmentation. The goal of medical image segmentation is to separate
an image into a number of different and disjoint sets of voxels where
Original
13

www.jbpe.orgKazemi K, Noorizadeh N
each set corresponds to the real anatomy of the
patient.
Segmentation of cerebrospinal fluid (CSF),
gray matter (GM) and white matter (WM)
from MR images is a challenging task. The
main difficulties are tissue intensities non-
uniformity (bias), noise artifacts and partial
volume effect (PVE). A number of techniques
have been proposed for automatic segmenta-
tion of GM, WM and CSF from cerebral MR
images: statistical-based segmentation [1-4]
geometrical-based segmentation [5-7], atlas-
based segmentation [8-12] and learning-based
segmentation methods [13]. Based on these
methods, several tools have been developed
by researchers to automate brain tissue seg-
mentation. However, assessment of the quality
of the segmented GM, WM and CSF is needed
to compare segmentation methods. The chal-
lenge in tissue segmentation now lies in hav-
ing a robust classification approach based on
image intensity values representing GM, WM
and CSF.
Actually, there are software packages which
are most widely used in neuroimaging anal-
ysis. These packages usually contain a set
of skull stripping, intensity non-uniformity
(bias) correction and automated segmentation
routines. There are three software packages
wildly used in the neuroimaging community
for structural and functional brain imaging
study. These packages are: SPM [8], written
by the Wellcome Department of Imaging Neu-
roscience at University College London, UK,
FMRIB Software Library (FSL) [14], written
by the Analysis Group, FMRIB, Oxford, UK,
and BrainSuite [15] written by the Laboratory
of Neuro Imaging at the University of Califor-
nia Los Angeles and the Biomedical Imaging
Research Group at the University of Southern
California.
There are different studies to assess the ac-
curacy of these software packages in skull
stripping. However, there is a few studies
investigate GM, WM and CSF segmentation
using complete package. Tsang et al. [16],
compared the segmentation algorithms that
are deployed in the two widely used software
packages SPM5 and FSL. Klauschen et al.
[17] conducted an experiment to evaluate tis-
sue segmentation using SPM, FSL and Free-
surfer. They assumed that skull stripping and
bias is corrected separately using FSL. Thus,
they evaluated the segmentation algorithm of
the packages.
In these studies, the assessment of the seg-
mentation packages was performed based on
the same skull-striping and intensity non-uni-
formity correction methods. Therefore, they
only compared the effects of the tissue seg-
mentation method. The neuroscientists usually
select and apply a package for skull striping,
intensity non-uniformity correction and seg-
mentation for brain volumetry or functional
analysis. However, the previous comparison
studies just evaluated part of these packages
such as segmentation [17] or registration [18].
On the other hand, the latest versions of these
packages perform the segmentation in differ-
ent methods which are not studied in former
studies. Therefore, comparison of complete
segmentation framework consisting of pre-
processing and segmentation of these pack-
ages is necessary.
In this study, we investigated the accuracy of
the latest version of software packages SPM,
FSL and Brainsuite for brain tissue segmen-
tation which are the most widely used brain
tissue segmentation software packages. Fur-
thermore, different brain tissue segmentation
methods developed in these packages, i.e.,
New-segmentation in SPM8, are studies and
compared. The comparison is performed using
a variety of independent datasets, and the most
widely used metrics in the literature. Accord-
ing to the authors knowledge such complete
comprehensive analysis has not been previ-
ously carried out and can be used by users of
14

www.jbpe.org Quantitative Comparison of SPM, FSL, and Brainsuite
these software packages.
A description of the datasets used, brief re-
view on the segmentation routines applied in
these three packages and details of the quan-
titative measures used in validation are de-
scribed in the next section. The results of ex-
periments are described in the Results section.
Finally, the discussion is given in the Discus-
sion section.
Materials and Methods
MRI Brain Data
In order to assess the performance of the
methods, validation was performed based on
simulated images using Brainweb simulator
and real MR images.
BrainWeb-Simulated Brain Data
The simulated 3D MR images (181×217×
181 voxels of 1mm3
isotropic resolution)
which are used as test data, are provided by
the BrainWeb simulated brain database from
the McGill University. This database provides
realistic simulations of MRI acquisition with
different levels of intensity non-uniformity
and noise. Simulated MR images are generat-
ed based on an anatomical model of normal
brain. Eighteen datasets with different noise
level (n) range between 0%, 1%, 3%, 5%, 7%
and 9% and intensity non-uniformity (rf) with
0%, 20% and 40% are used.
IBSR-Real Data
In order to evaluate the performance of the
methods on real MR data, the cerebral MR
datasets from 20 normal subjects provided by
Center for Morphometric Analysis at Massa-
chusetts General Hospital (IBSR) are used.
The real MR brain data and their hand-guided
expert segmentation results are available at
these datasets. IBSR provides performance re-
sults from five other automatic segmentation
methods making it convenient to compare the
results with those reported by others. These 20
datasets involve different levels of difficulty
such as low contrast scans, relatively smaller
brain volumes and considerable intensity non-
uniformity.
Segmentation Methods
There are number of tools available for brain
tissue segmentation with different usages.
In this study, the accuracy of three fully au-
tomated brain tissue segmentation packages
SPM, FSL and Brainsuite were compared.
These packages incorporate tools for intensity
non-uniformity correction and skull-stripping/
masking procedures. FSL and BrainSuite per-
form skull-stripping and intensity non-unifor-
mity correction as pre-processing for the seg-
mentation procedure, while SPM does these
steps concurrently during segmentation.
SPM
The SPM (Statistical Parametric Mapping)
software SPM8 is a MATLAB software pack-
age implementing statistical methods for anal-
ysis of functional and structural neuroimages.
Brain tissue segmentation with SPM8 can be
performed in three frameworks: i) default seg-
mentation, ii) SPM8 New-segmentation and
iii) modified version using hidden Markov
random field as an additional spatial constraint
as implemented in SPM-VBM toolbox.
SPM8 default Segmentation
The default segmentation routine imple-
mented in SPM [8] is based on a unified seg-
mentation model that performs tissue segmen-
tation, registration and intensity
non-uniformity (bias) correction all in the
same model. The principal idea of this method
is to model image intensities as a mixture of k
Gaussians, where each Gaussian cluster is
modelled by its mean ( kµ ), variance ( kσ ) and
a mixing proportion. In the unified model, the
tissue probability maps (TPMs) are used as a
priori information of the tissue classes. The
Bayes rule is employed to produce the poste-
rior probability of each tissue class. In SPM,
the segmentation routine automatically seg-
15

www.jbpe.org
ments the input MR image into GM, WM, and
CSF. However, the classification is probabilis-
tic in the sense that a probability value of be-
longing to each of the classes is assigned to
each voxel in the output images. When gener-
ating binary images, voxels corresponding to
grater tissue probability in the maps are count-
ed as members of that particular class.
The recent version of SPM software SPM8
also takes advantage of different improved
models and a more robust initial affine trans-
formation for registration, an extended set of
tissue probability maps, a different treatment
of the mixing proportions and the capability of
working with multispectral data. In this study,
all of the segmentations by SPM8 are per-
formed using the default atlas (a modified ver-
sion of the ICBM Tissue Probabilistic Atlas,
available at http://www.loni.ucla.edu/ICBM/
ICBM_Probabilistic.html) and parameters for
this version.
SPM8 New-segmentation
In SPM8 a New-segmentation routine is im-
plemented which is an extension of the default
segmentation routine. The algorithm is essen-
tially the same as default unified segmentation
with the following modifications:
• A different treatment of the mixing propor-
tions.
• The use of an improved registration model.
• The ability to use multispectral data.
• An extended set of tissue probability maps,
which allows a different treatment of non-
brain voxels.
• A more robust initial affine registration.
The New-segmentation routine can segment
the brain into six tissue classes: GM, WM,
CSF, bone, soft tissue, and air/background.
SPM8-VBM
VBM is a collection of extensions to the
segmentation algorithm of SPM which uti-
lizes hidden Makrov random field as post
processing [19]. This modification to the al-
gorithm helps in determining the probability
a given voxel belong to a tissue class which is
achieved by calculating the MRF energy for a
given voxel, based on its proximity to the sur-
rounding voxels.
FSL
FMRIB’s Software Library (FSL) is a soft-
ware package developed by members of the
Oxford Centre for Functional MRI of the Brain
(Oxford University) which is composed of im-
age analysis and statistical tools for neuroim-
age data study [14]. Although FSL has many
different modules for functional and structural
MRI data analysis, in this section we are only
going to focus in FMRIB Automated Seg-
mentation Tool (FAST) which developed for
segmentation of brain tissues. The FSL-FAST
segmentation routine is based on a Hidden
Markov Random Field (HMRF) model that is
optimized using the Expectation-Maximiza-
tion algorithm [20].
In this study, FSL version 4.1 is employed
for the whole process of registration, skull
stripping and brain tissue segmentation. Since
the FSL-FAST brain tissue segmentation re-
quires skull stripped version of input MRI
data, as the first step of tissue segmentation,
skull stripping is performed using the FSL’s
own brain extraction tool. Then, in the second
step, FAST tool using probability maps as its
default settings is used to segment the brain
into three tissue classes of GM, WM and CSF
and performing bias correction.
Skull stripping using BET
Intracranial segmentation commonly re-
ferred to as “skull-stripping” removes extra-
cerebral tissues such as skull, eyeballs, and
skin. Skull-stripping facilitates image pro-
cessing such as surface rendering, cortical
flattening, image registration, and tissue seg-
mentation. Thus, as the first step of the FSL
segmentation, the “Brain Extraction Tool ver-
sion 2.1 (BET)” integrated in FSL software is
Kazemi K, Noorizadeh N
16

used to perform skull stripping and remove
non-brain parts of the image.
Brain tissue segmentation using FAST
MR imaging suffer from non-homogeneity
in radio-frequency field which results in non-
biological intensity non-uniformities across
the imaged brain. Thus, in the second step
FAST toolbox version 4.1 (FMRIB’s Auto-
mated Segmentation Tool) integrated in FSL
software is used to segment GM, WM and
CSF, whilst also correcting for intensity non-
uniformity. Accurate intensity non-uniformity
correction requires segmentation knowledge
while perfect segmentation requires intensity
non-uniformity to be corrected. The segmen-
tation routine implemented in FAST toolbox is
based on HMRF model and associated expec-
tation-maximization algorithm. In this meth-
od, the histogram of input image is modeled
as a mixture of Gaussians with mean and vari-
ance for each class. The segmentation allows
a reconstruction of the image; subtracting this
from the real image gives an estimate of the
non-uniformity. This whole process is then
iterated between segmentation and intensity
non-uniformity correction until convergence
[14,20]. The resulting outputs are intensity
non-uniformity corrected version and seg-
mented GM, WM and CSF from input data.
BrainSuite
BrainSuite is a suite of image analysis tools
designed to process MRI data of the human
head. Brain tissue segmentation routine of
BrainSuite starts by skull-striping using brain
surface extractor (BSE) tool followed by bias
field corrector (BFC) tool for intensity non-
uniformities correction and partial volume
classifier (PVC) tool for tissue segmentation.
Skull stripping using BSE
The brain tissue segmentation in Brainsuite
is started by applying BSE tool to remove non-
brain tissues from input cerebral MRI data. It
operates by using an edge detector to find a
boundary between the brain and the skull by
using mathematical morphological operators
to enhance the result. The edge detection re-
sult is improved by the application of an iso-
tropic diffusion filter before the edge detection
Part Item BrainWeb IBSR
BSE
Diffusion Iterations 8 3
Diffusion Constant 55 25
Edge Constant 0.65 0.64
Erosion Size 4 1
BFC
Histogram Radius 12 12
Sample Spacing 16 16
Control Point Spacing 64 64
Spline Stiffness 0.0001 0.0001
ROI Shape Cubiod Cubiod
Lower Limit 0.9 0.9
Upper Limit 1.1 1.1
PVC Spatial Prior 0.1 0.1
Table 1: The parameters that have been chosen for each simulated and real dataset.
17

www.jbpe.org
is performed. The parameters that have been
chosen are presented in table 1 for applied data
sets.
Intensity non-uniformities correction
using BFC
After skull stripping, BFC tool is used to
correct the image intensity non-uniformities
that are due to magnetic field variations. It es-
timates a correction field for the brain region
based on a series of local estimates of the tis-
sue gain variation.
Brain tissue segmentation using PVC
As the last step of the brain tissue segmenta-
tion, the PVC tool is used for tissue classifica-
tion. It is performed by partial volume classi-
fier. It assigns an integer tissue label to each
voxel in the image. These labels correspond to
the type of tissue that is estimated to be in that
voxel. It can generate labeling with three tis-
sue classes output that are GM, WM and CSF
and also with six tissue classes that are com-
posed of combination of these voxels.
Evaluation method
Similarity metrics
Two similarity metrics are used for quantita-
tive evaluation of the proposed method: Dice
coefficient [5] or similarity index and Jaccard
coefficient [8]. These metrics represent spatial
overlap between two binary images and their
values range between 0 (no overlap) and 1
(perfect agreement) as they are expressed as
percentage in the following:
2 A G
D
A G
∩
=
+
(1)
A G
J
A G
∩
=
∪
(2)
Success and error rate
The sensitivity (true positive fraction, TPF)
refers to the ability to correctly identify ap-
propriate tissue in the segmented mask. The
higher sensitivity shows the lower missed true
tissue voxels. It is defined as follows.
TP
TPF
TP FN
=
+
(3)
The specificity (true negative fraction, TNF)
refers to the ability of the proposed segmenta-
tion method to correctly remove non-desired
voxels. The higher Specificity shows the low-
er missed true non-desired voxels.
TN
TNF
TN FP
=
+
(4)
Sensitivity and specificity are the measures
that are computed based on true positive (TP),
false negative (FN) and false positive (FP),
true negative (TN).
Results
In order to assess the relative accuracy of
three brain tissue segmentation tools: SPM8,
FSL and Brainsuite, they are applied to seg-
ment GM, WM and CSF from simulated and
IBSR real MR images.
Comparison based on simulated data
The resulting GM,WM and CSF tissues
from segmentation of simulated MR image
with intensity non-uniformity rf=0% and noise
level n=0% using brain tissue segmentation
tools: SPM8 (including SPM8-Seg, SPM8-
VBM, SPM8-NewSeg), FSL and Brainsuite
are shown in figure1. Figure 2 illustrates the
quantitative assessment of the accuracy of the
segmented GM,WM and CSF from simulated
MR images with intensity non-uniformity lev-
els 0%, 20% and 40% in different noise levels
0%, 1%, 3%, 5%, 7% and 9% based on Dice
similarity metric.
Table 2 shows the mean similarity based on
Dice and Jaccard metrics for segmented GM,
WM and CSF from 18 simulated MR images.
Furthermore, the mean sensitivity and specific-
ity are shown. Here, SPM8 and FSL results are
relatively close when compared with Brain-
suite. However, SPM8-VBM performs bet-
18

Figure 1: The segmented GM, WM and CSF from simulated MR image with n=0% and rf=0%.
The first row shows the input MR image and the ground truth for GM, WM and CSF. The sec-
ond row from left to right shows the segmentation results using SPM8-Seg, SPM8-VBM, SPM8-
NewSeg, FSL and Brainsuite, respectively.
Tissue
Software
Packages
Dice Jaccarde Sensitivity Specificity
WM
SPM8-Seg 0.92±0.04 0.86±0.06 0.92±0.02 0.99±0.01
SPM8-VBM 0.95±0.02 0.90±0.03 0.96±0.01 0.99±0.01
SPM8-NewSeg 0.92±0.03 0.86±0.06 0.94±0.04 0.99±0.01
FSL 0.93±0.02 0.86±0.03 0.89±0.02 0.99±0.01
Brainsuite 0.91±0.06 0.84±0.09 0.90±0.10 0.99±0.01
GM
SPM8-Seg 0.90±0.03 0.83±0.06 0.90±0.04 0.99±0.01
SPM8-VBM 0.93±0.02 0.87±0.04 0.92±0.03 0.99±0.01
SPM8-NewSeg 0.91±0.03 0.83±0.06 0.91±0.03 0.99±0.01
FSL 0.92±0.02 0.85±0.04 0.95±0.02 0.98±0.01
Brainsuite 0.80±0.16 0.68±0.20 0.74±0.21 0.99±0.01
CSF
SPM8-Seg 0.62±0.05 0.46±0.05 0.96±0.04 0.94±0.01
SPM8-VBM 0.76±0.02 0.61±0.02 0.66±0.02 0.99±0.01
SPM8-NewSeg 0.74±0.07 0.59±0.09 0.79±0.04 0.98±0.01
FSL 0.74±0.06 0.6±0.08 0.66±0.09 0.99±0.01
Brainsuite 0.46±0.16 0.31±0.13 0.37±0.15 0.99±0.01
Table 2: Comparison of segmentation accuracy of SPM8, FSL and Brainsuite using Brainweb
simulated MR images.
19

www.jbpe.org
Figure 2: Quantitative evaluation of the segmented GM and WM from simulated MR images
based on Dice similarity metric using three packages: SPM, FSL and Brainsuite.
ter than two other packages. This can be seen
from figure 2 where SPM8-VBM segmented
accurately the WM and GM even in high level
of noise and intensity non-uniformity.
In order to investigate the significant dif-
ferences between segmentation methods of
SPM8, t-test was used to find significant differ-
ence between SPM8-Seg and SPM8-NewSeg.
However, we could not find a significance dif-
ference between them.
Comparison based on Real IBSR data
The extracted GM, WM and CSF tissues
from a selected real MR image using the ap-
plied brain tissue segmentation tools are
shown in figure 3. Figure 4 shows the accu-
racy of segmented GM and WM from IBSR
real MR images using Dice similarity metric.
Furthermore, since the segmentation accuracy
provided by the IBSR is measured by the Jac-
card similarity metric, the mean Dice and Jac-
carde similarity metrics are shown in table 3.
Similar to segmented WM and GM from
simulated images, here, SPM8 performed
better than FSL and Brainsuite for real MR
images. On the other words, comparing the
20

Figure 3: The segmented GM, WM and CSF from a selected subject from IBSR real MR images.
The first row shows the input MR image and the ground truth for GM, WM and CSF. The sec-
ond row from left to right shows the segmentation results using SPM8-Seg, SPM8-VBM, SPM8-
NewSeg, FSL and Brainsuite, respectively.
Figure 4: Dice similarity metric for segmented WM and GM from IBSR real MR images
results of IBSR dataset against the results of
the Brainweb dataset showed that the perfor-
mance of the SPM8 algorithm was consistent
for both datasets.
However, comparison of the results obtained
using SPM8 toolboxes showed that the SPM8-
NewSeg segmentation algorithm provided
higher similarity values when overlaid against
the ground truth of the expert segmented tissue
images. This result for SPM8 was confirmed
by t-test with p<0.05.
21

www.jbpe.org
Brain tissue segmentation
Brain (consisting of GM and WM) extrac-
tion is one of the most time-consuming steps
in neuroimage analysis. While numerous brain
extraction methods have been developed to
perform this step automatically, their output
varies and may affect the results of subsequent
image analysis. Thus, the extracted brain us-
ing three popular neuroimage analysis tools
SPM8, FSL and BrainSuite were compared.
The brain mask was created as the binarized
sum of the GM and WM after segmentation.
Figure 5 shows the accuracy of segmented
brain in term of Dice metric for simulated MR
image in different noise and intensity non-uni-
formity levels. Table 4 shows quantitative re-
sults in term of mean similarity for segmented
brain using different tools. It can be seen that
Table 3: Comparison of segmentation accuracy of SPM8, FSL and Brainsuite using IBSR real MR
images
Tissue
Software
Packages
WM
SPM8-Seg 0.82±0.03 0.70±0.04 0.81±0.02 0.99±0.01
SPM8-VBM 0.81±0.03 0.68±0.04 0.87±0.05 0.99±0.01
SPM8-NewSeg 0.82±0.03 0.70±0.04 0.85±0.02 0.99±0.01
FSL 0.68±0.18 0.54±0.18 0.86±0.19 0.97±0.01
Brainsuite 0.76±0.14 0.63±0.15 0.79±0.18 0.99±0.01
SPM8-Seg 0.79±0.03 0.65±0.04 0.73±0.05 0.99±0.01
GM
SPM8-VBM 0.76±0.04 0.62±0.05 0.67±0.05 0.99±0.01
SPM8-NewSeg 0.80±0.02 0.66±0.03 0.73±0.03 0.99±0.01
FSL 0.66±0.12 0.51±0.13 0.67±0.06 0.98±0.02
Brainsuite 0.72±0.11 0.57±0.13 0.64±0.15 0.99±0.01
Tissue
Software
Packages
Brainweb
SPM8-Seg 0.91±.03 0.86±.06 0.91±0.02 0.99±0.01
SPM8-VBM 0.94±.02 0.89±.03 0.94±0.02 0.99±0.01
SPM8-NewSeg 0.91±.03 0.84±.06 0.92±0.03 0.99±0.01
FSL 0.92±.02 0.86±.04 0.92±0.02 0.99±0.01
Brainsuite 0.85±.11 0.76±.14 0.82±0.16 0.99±0.01
IBSR
SPM8-Seg 0.80±.03 0.88±.03 0.77±0.02 0.99±0.01
SPM8-VBM 0.79±.04 0.88±.03 077±0.05 0.99±0.01
SPM8-NewSeg 0.81±.02 0.89±.04 0.72±0.02 0.99±0.01
FSL 0.67±.15 0.89±.03 0.76±0.08 0.97±0.02
Brainsuite 0.74±.12 0.89±.13 0.72±0.14 0.99±0.01
Table 4: Comparison of brain segmentation accuracy using SPM8, FSL and Brainsuite based on
BrainwebSimulated dataset and IBSR real dataset
22

www.jbpe.org
Figure 5: Dice similarity metric for segmented Brain from Brainwebsimulated MR images.
SPM8 performs better than other tools in term
of accuracy. Comparison between segmenta-
tion toolboxes of SPM8 shows that the SPM8-
VBM which utilizes MRF has better perfor-
mance. On the other hand, the t-test showed
no scientific difference between SPM8-Seg
and SPM8-NewSeg tools integrated in SPM8.
The same comparison was made to assess the
accuracy of segmented brain using real IBSR
MR images. The resulting Dice coefficients of
segmented brain (consisting of GM and WM)
were shown in figure 6. Table 4 shows the
mean similarity between the segmented brain
and corresponding ground truth based on Dice
and Jaccard similarity metrics. Similar to the
results obtained using simulated images, the
Quantitative Comparison of SPM, FSL, and Brainsuite
23

www.jbpe.org
Figure 6: Dice similarity metric for segmented Brain from IBSR real MR images.
SPM8 toolbox provide better results in com-
pared to the two other packages.
Discussion and Conclusion
Segmentation of brain tissues GM, WM and
CSF from MR images is challenging task be-
cause of some difficulties like tissue intensi-
ties non-uniformity, noise artifacts and partial
volume effect. Beside of different techniques
that have been proposed for automatic seg-
mentation of brain tissues from cerebral MR
images, there are software packages which
are most widely used in neuroimaging analy-
sis. Among them, SPM8, FSL and BrainSuite
software packages are usually applied by sci-
entists for structural and functional analysis of
brain.
In this paper, we quantitatively evaluated the
segmentation performance of these software
packages through a series of experiments.
Unlike the existing evaluations that perform
preprocessing such as skull striping and in-
tensity non-uniformity by the same tool and
just the segmentation routine is evaluated, in
this study, whole segmentation consisting of
pre-processing and tissue segmentation were
performed by each tool. Standard procedures
with parameters that give best results were
used.
Our analysis was performed based on Brain-
web simulated MR images and IBSR real MR
images. The results show variation in segmen-
tation performance of the tools for brain tis-
sue segmentation in term of accuracy. Differ-
ence in accuracy is the fact that these tools are
designed for slightly different purpose, even
though all can be used for m brain tissue seg-
mentation.
While SPM is designed to perform segmen-
tation of brain tissues consisting of GM, WM
and CSF, the FSL and Brainsuite can seg-
ment sub-cortical structures also. On the other
words, FSL and Brainsuite are general propose
tools in comparison with SPM.
Comparison of the tools based on simulated
and real MR images show that performance of
the Brainsuite was influenced by noise and in-
tensity non-uniformity. The FSL performance
was influenced by the noise of the image and
little by the image intensity non-homogeneity.
The accuracy of segmented GM and WM were
reduced about 5% by increasing the noise
24

www.jbpe.org
from 0% to 9% in the simulated MR images.
Finally, the obtained results show that similar
to two other tools, the segmentation was in-
fluenced by noise but little by image intensity
non-homogeneity. However, applying MRF as
post-processing in VBM toolbox of SPM re-
duce the influence of noise and improves the
segmentation accuracy of GM and WM. The
little influence of intensity non-homogeneity
on the segmentation results is results of inten-
sity non-homogeneity estimation during seg-
mentation procedure in SPM8 and FSL.
On the other hand, the accuracy of the tools
in brain segmentation was evaluated. The
qualitative evaluation using Dice and Jacca-
rd metrics shows SPM8 performs better than
others. Beside, based on t-test, no significant
difference between SPM8-Seg and SPM8-
NewSeg was found.
Conflict of Interest
None
References
1. Bezdek JC, Hall LO, Clarke LP. Review of MRI Seg-
mentation Techniques using Pattern Recognition.
Med Phys. 1993;20:1033-48.
2. Held K, Rota Kops E, Krause BJ, Wells WM, Kikinis
R, Muller-Gartner HW. Markov random field seg-
mentation of brain MR images. IEEE Trans Med Im-
aging. 1997;16:878-86. doi: 10.1109/42.650883.
PubMed PMID: 9533587.
3. Liew A, Yan H. An Adaptive Spatial Fuzzy Cluster-
ing Algorithm for MR Image Segmentation. IEEE
Trans Med Imag. 2003;22:1063-75.
4. Wells WM, Grimson WL, Kikinis R, Jolesz FA. Adap-
tive segmentation of MRI data. IEEE Trans Med Im-
aging. 1996;15:429-42. doi: 10.1109/42.511747.
5. Li C, Kao C, Gore JC, Ding Z. Minimization of re-
gion-scalable fitting energy for image segmenta-
tion. IEEE Trans Imag Process. 2008;17: 1940-9.
6. Shahvaran Z, Kazemi K, Helfroush MS, Jafarian
N, Noorizadeh N. Variational level set combined
with Markov random field modeling for simultane-
ous intensity non-uniformity correction and seg-
mentation of MR images. J Neurosci Methods.
2012;209:280-9.
7. Wang L, Chen Y, Pan X, Hong X, Xia D. Level set
segmentation of brain magnetic resonance images
based on local Gaussian distribution fitting en-
ergy. J Neurosci Methods. 2010;188:316-25. doi:
10.1016/j.jneumeth.2010.03.004. PubMed PMID:
20230858.
8. Ashburner J, Friston KJ. Unified segmentation.
NeuroImage. 2005;26:839-51. doi: 10.1016/j.neu-
roimage.2005.02.018. PubMed PMID: 15955494.
9. Collins DL, Holmes CJ, Peters TM, Evans AC. Au-
tomatic 3D Model-Based Neuroanatomical Seg-
mentation. Hum Brain Mapp. 1995;3:190-208.
10. Marroquin JL, Vemuri BC, Botello S, Calderon F,
Fernandez-Bouzas A. An accurate and efficient
bayesian method for automatic segmentation of
brain MRI. IEEE Trans Med Imaging. 2002;21:934-
45. doi: 10.1109/tmi.2002.803119. PubMed PMID:
12472266.
11. Van Leemput K, Maes F, Vandermeulen D, Suetens
P. Automated model-based tissue classification of
MR images of the brain. IEEE Trans Med Imag-
ing. 1999;18:897-908. doi: 10.1109/42.811270.
12. Zhou Y, Bai J. Atlas-based fuzzy connectedness
segmentation and intensity nonuniformity correc-
tion applied to brain MRI. IEEE Trans Biomed Eng.
2007;54:122-9. doi: 10.1109/tbme.2006.884645.
13. Hall LO, Bensaid AM, Clarke LP, Velthuizen RP,
Silbiger MS, Bezdek JC. A comparison of neural
network and fuzzy clustering techniques in seg-
menting magnetic resonance images of the brain.
IEEE Trans Neural Netw. 1992;3:672-82. doi:
10.1109/72.159057. PubMed PMID: 18276467.
14. Smith SM, Jenkinson M, Woolrich MW, Beck-
mann CF, Behrens TE, Johansen-Berg H, et al.
Advances in functional and structural MR image
analysis and implementation as FSL. NeuroIm-
age. 2004;23:S208-19. doi: 10.1016/j.neuroim-
age.2004.07.051. PubMed PMID: 15501092.
15. Shattuck DW, Leahy RM. BrainSuite: an automat-
ed cortical surface identification tool. Med Image
Anal. 2002;6:129-42. PubMed PMID: 12045000.
16. Tsang O, Gholipour A, Kehtarnavaz N, Gopi-
nath K, Briggs R, Panahi I. Comparison of tis-
sue segmentation algorithms in neuroimage
analysis software tools. Conf Proc IEEE Eng
Med Biol Soc. 2008;2008:3924-8. doi: 10.1109/
iembs.2008.4650068. PubMed PMID: 19163571.
Quantitative Comparison of SPM, FSL, and Brainsuite
25

www.jbpe.org
17. Klauschen F, Goldman A, Barra V, Meyer-Linden-
berg A, Lundervold A. Evaluation of automated
brain MR image segmentation and volumetry
methods. Hum Brain Mapp. 2009;30:1310-27. doi:
10.1002/hbm.20599. PubMed PMID: 18537111.
18. Wu M, Carmichael O, Lopez-Garcia P, Carter CS, Ai-
zenstein HJ. Quantitative comparison of AIR, SPM,
and the fully deformable model for atlas-based
segmentation of functional and structural MR
images. Hum Brain Mapp. 2006;27:747-54. doi:
10.1002/hbm.20216. PubMed PMID: 16463385;
PubMed Central PMCID: PMC2886594.
19. Cuadra MB, Cammoun L, Butz T, Cuisenaire O,
Thiran JP. Comparison and validation of tissue
modelization and statistical classification meth-
ods in T1-weighted MR brain images. IEEE Trans
Med Imaging. 2005;24:1548-65. doi: 10.1109/
tmi.2005.857652. PubMed PMID: 16350916.
20. Zhang Y, Brady M, Smith S. Segmentation of brain
MR images through a hidden Markov random field
model and the expectation-maximization algorithm.
IEEE Trans Med Imaging. 2001;20:45-57. doi:
10.1109/42.906424. PubMed PMID: 11293691.
21. Dice LR. Measures of the Amount of Ecologic As-
sociation between Species. Ecology. 1945;26:297-
302.
22. Jaccard P. The Distribution of Flora in the Alpine
Zone. New Phytol. 1912;11:37-50.
26

Segmentation

More Related Content

What's hot (17)

Similar to Segmentation (20)

Recently uploaded (20)

Segmentation