Available online at www.sciencedirect.

com

Magnetic Resonance Imaging 31 (2013) 262 271

3D cerebrovascular segmentation combining fuzzy vessel enhancement

and level-sets with anisotropic energy weights
Nils Daniel Forkert a,, Alexander Schmidt-Richberg b , Jens Fiehler c , Till Illies c ,
Dietmar Mller d , Dennis Sring a , Heinz Handels b , Jan Ehrhardt b
a

Department of Computational Neuroscience, University Medical Center Hamburg-Eppendorf, 20246 Hamburg, Germany
b
Institute of Medical Informatics, University of Lbeck, 23538 Lbeck, Germany
c
Department of Diagnostic and Interventional Neuroradiology, University Medical Center Hamburg-Eppendorf, 20246 Hamburg, Germany
d
Department Computer Engineering, Faculty of Mathematics, Informatics and Natural Sciences, University of Hamburg,
22527 Hamburg, Germany
Received 2 January 2012; revised 16 July 2012; accepted 17 July 2012

Abstract
The aim of this work is to present and evaluate a level-set segmentation approach with vesselness-dependent anisotropic energy weights,
which focuses on the exact segmentation of malformed as well as small vessels from time-of-flight (TOF) magnetic resonance angiography
(MRA) datasets.
In a first step, a vesselness filter is used to calculate the vesselness dataset, which quantifies the likeliness of each voxel to belong to a
bright tubular-shaped structure and estimate the corresponding vessel directions from a given TOF dataset. The vesselness and TOF datasets
are then combined using fuzzy-logic and used for initialization of a variational level-set method. The proposed level-set model has been
extended in a way that the weight of the internal energy is locally adapted based on the vessel direction information. Here, the main idea is to
weight the internal energy lower if the gradient direction of the level-set is similar to the direction of the eigenvector extracted by the
vesselness filter. Furthermore, an additional vesselness force has been integrated in the level-set formulation.
The proposed method was evaluated based on ten TOF MRA datasets from patients with an arteriovenous malformation. Manual
segmentations from two observers were available for each dataset and used for quantitative comparison. The evaluation revealed that the
proposed method yields significantly better segmentation results than four other state-of-the-art segmentation methods tested. Furthermore,
the segmentation results are within the range of the inter-observer variation.
In conclusion, the proposed method allows an improved delineation of small vessels, especially of those represented by low intensities
and high surface curvatures.
2013 Elsevier Inc. All rights reserved.
Keywords: Magnetic resonance angiography; Cerebrovascular disorders; Fuzzy logic; Level-set segmentation

1. Introduction
Vascular diseases are a major cause of death and
disability worldwide and a large number of people suffer
from a cerebrovascular incidence each year. Besides
ischemic strokes, hemorrhages are the second major cause
for cerebral strokes and account for approximately 1020%
of all cases [1]. Most hemorrhages are thought to result from
the rupture of small vessels in the presence of long standing
Corresponding author. Tel.: +49 40 7410 59828; fax: +49 40 7410
54882.
E-mail address: n.forkert@uke.uni-hamburg.de (N.D. Forkert).
0730-725X/$ see front matter 2013 Elsevier Inc. All rights reserved.
http://dx.doi.org/10.1016/j.mri.2012.07.008

hypertonus or rupture of a pathological vessel wall of larger

arteries or veins. Examples for such vessel pathologies are
aneurysms [2] and arteriovenous malformations [3]. Especially if diagnosed previous to a hemorrhagic event, an
interventional procedure may signicantly reduce the
bleeding risk of such vessel pathologies.
Segmentations of the cerebral vessels are often required in
the clinical routine for improved diagnosis assistance,
surgery planning, postoperative monitoring and research
purposes [4]. In addition to this, cerebrovascular segmentations come along with further benets such as an improved
registration of angiographic datasets [5], possible blood ow
simulations [6], improved vessel visualization using surface

N.D. Forkert et al. / Magnetic Resonance Imaging 31 (2013) 262271

models [7], quantication of pathologies [8] and preoperative denition a vessel-free path needed for brain tumor
biopsies and deep brain stimulations in case of epilepsy and
Parkinson disease [9].
Due to the fact that the manual delineation of vessel
structures is time-consuming and error-prone, the automatic
segmentation of vascular structures has been in the focus of
research for several years. Numerous factors have to be taken
into consideration for the development of an automatic
segmentation approach. Among others, these factors include
the high complexity of vessel shapes, including diameter and
curvature, especially in case of pathological structures, as well
as noise and other imaging artifacts, different vessel contrasts,
the representation of surrounding organs and general aspects
like image dimension or resolution [10]. Therefore, a general
method that can extract the vascular system from any kind of
angiographic dataset is not available.
Due to the high blood-to-background contrast [11], the 3D
time-of-ight (TOF) magnetic resonance angiography
(MRA) is a commonly used MR imaging technique in
clinical practice for diagnosis of the cerebral vascular system
[12] and numerous vessel segmentation approaches have
been proposed in the past that can be used for delineation of
vessel structures from this image sequence. A rough overview
of the state-of-the-art of vessel segmentation techniques is
given in the following.

2. State-of-the-art
In general, typical state-of-the-art vessel segmentation
schemes can, for example, be distinguished into threshold-,
scale-space- and deformable-model-based methods as well
as hybrid approaches that combine two or more of these
basic techniques into one framework.
Threshold-based approaches aim to extract one global
threshold or locally adaptive thresholds for segmenting vessel
structures. One comparably simple and fast threshold-based
approach is the Z-buffer segmentation (ZBS) method as
described by Chapman et al. [13]. In this approach, a
maximum intensity projection of the 3D TOF dataset is
computed rst while the corresponding voxel coordinates are
saved in a Z-buffer image. After a consistency check of the Zbuffer, a set of seed points is extracted, which are used to
obtain the nal segmentation via volume growing. Another
volume-growing segmentation method was presented by Yi
and Ra [14]. Here, the 3D dataset is separated into several
cube-shaped sub-volumes, which are used to calculate locallyadaptive thresholds that are then applied in a volume-growing
process. Stochastic models are another common approach for
the extraction of a threshold. The basic idea of these models is
to extract the histogram of a given image and t a mixture of
statistical distributions that represent different tissues using the
Expectation Maximization (EM) algorithm [15]. Wilson and
Noble [16] proposed a stochastic model based on two
Gaussian distributions, one for modeling the cerospinal

263

uid, bone and background and the second for the brain
tissue and eyes, and one uniform distribution for arteries and
subcutaneous fat. Hassouna et al. [17] found that an improved
tting accuracy can be achieved if using four distributions. In
this work, a Rayleigh distribution was used to model the
background and three Gaussian distributions were used to
design the remaining tissue distributions, whereas the
Gaussian distribution modeling the high intensity region
corresponds to vascular structures. After thresholding, a
Markov random eld is employed in a post-processing step.
Intensity-based segmentation techniques are usually simple to
implement, fast and are able to delineate malformed vessels,
which are represented by high intensities. However, these
approaches tend to be sensitive for noise artifacts leading to
holes within the segmentation and leakages to non-vascular
tissues. Moreover, intensity-based approaches often fail to
segment small vessels that are represented by low intensities.
Scale-space-based methods are another common general
technique for cerebrovascular segmentation. Here, especially
line-lter methods [1820] have been used extensively in the
past. The main idea of these approaches is to convolve the
angiographic image with Gaussian derivatives of different
standard deviations to obtain information about the local
image geometry. The different scales are used to deal with
different vessel sizes. Commonly, the Hessian matrix is used
to obtain the required second-order derivative information
but the use of the Weingarten matrix has also been proposed
as an alternative [21]. Certain relations of the eigenvalues of
the matrix are used to discriminate tubular structures in the
image, while the corresponding eigenvectors can be used to
estimate the vessel orientation in space. The enhanced socalled vesselness images have been used for direct visualization [19], thresholding [18] or active contour segmentation
[20]. Scale-space approaches allow an improved small-vessel
extraction but usually fail to delineate malformed vessels,
which do not exhibit a typical tubular vessel shape.
Deformable-model-based approaches deform an initial
contour or surface by internal and external forces. Generally,
the internal forces keep the evolving contour smooth, while
the external energy drives the contour towards dened image
features. Due to the fact that the parameterization of active
contours is very challenging for vascular segmentation in
3D, implicit active contours such as level-sets are more
frequently used for this purpose. Several level-set vessel
segmentation schemes have been proposed in the past that
differ regarding the integration and extent of shape priors and
usage of gradient features or global intensity statistics. One
well known level-set segmentation approach that utilizes a
strong shape prior is the curve evolution method proposed by
Lorigo et al. [22]. The main idea of this approach is to evolve
line structures in the 3D image domain. To account for the
special characteristics of vessels, the width-limited surface is
evolved by constraining its lowest curvature. In contrast to
this, Vasilevskiy and Siddiqi proposed a level-set framework
for vascular segmentation that is based on ux-maximizing
ow [23]. Here, the main idea is to align the surface normal

264

N.D. Forkert et al. / Magnetic Resonance Imaging 31 (2013) 262271

Fig. 1. Illustration of the single processing steps of the proposed method.

to the intensity gradient direction. Gooya et al. [24] extended

this approach by adding region statistical measures. Levelset segmentation schemes are topologically exible but may
lead to insufcient small vessel delineation as edge and
intensity information may be too weak such that the internal
energy is stronger than the external energy.
Several more approaches for the cerebrovascular segmentation have been proposed in the past. A good overview
of recent segmentation techniques is for example given by
Suri et al. [25] or Lesage et al. [21].
The aim of this work is to present and evaluate a
cerebrovascular segmentation scheme fusing the benets of
intensity-, scale-space- and deformable-model-based approaches. Fig. 1 shows the single steps of the proposed
method that will be described in more detail in the following.
3. Methods and materials
3.1. Material and preprocessing
Overall, 11 datasets of patients with an arteriovenous
malformation of various volumes were available for this

work. All MR imaging measurements were performed on a

3T Trio scanner (Siemens, Erlangen, Germany) using an
eight-channel-phased array-head-coil. Among others, 3D
TOF MRA image sequences were acquired for each patient.
The TOF image sequences used in this study were acquired
using a TR 36 ms, TE 6 ms, ip angle 25, bandwidth 178 Hz/
Px, 6/8 slice partial Fourier, ow compensation, ve slabs
sequence each consisting of 40 partitions with an image inplane resolution of 0.47 mm and 0.5 mm slice thickness.
Informed consent was obtained from all patients. The study
was approved by the local ethics committee (No. 2706/2005).
Prior to segmentation, each TOF image sequence was
preprocessed using the histogram-based slab boundary
artifact reduction method proposed by Kholmovski et al.
[26] to reduce slice-related intensity variations caused by
the multi-slab acquisition. After this, in-slice intensity nonuniformities, for example caused by poor radio frequency
coil uniformity, were corrected using the N3 algorithm [27].
Finally, a skull-stripping algorithm especially designed for
TOF image sequences [28] was applied in the last step of
the preprocessing to exclude non-cerebral tissues from the
TOF-images.

N.D. Forkert et al. / Magnetic Resonance Imaging 31 (2013) 262271

3.2. Vesselness parameter extraction

In this work, the vesselness lter proposed by Sato et al.
[18] is used to enhance tubular structures in the TOF image
sequence and obtain information about the vessel direction in
3D space. Briey described, this lter analyzes the
eigenvalues of the Hessian matrix:
2
3
Txx Txy Txz
H = 4 Tyx Tyy Tyz 5
1
Tzx Tzy Tzz
where Txx, Txy,Tzz represent the second-order partial derivatives of the TOF image T(x) with x=(x,y,z) and the
image domain 3. Let the eigenvalues of H be dened
by 1, 2 and 3 and the corresponding eigenvectors by e1, e2
and e3. In this case, e1 represents the direction where the
second derivative achieves its maximum (see Fig. 2), which
gives an estimation of the course of a vessel in 3D space. This
direction information will be included in the level-set
approach in the following. The fact that tubular structures
should exhibit a 1 close to 0 and large negative values for 2
and 3 is used to calculate the vesselness measure for each
voxel. Practically, the Hessian matrix is calculated using the
second derivatives of a Gaussian of different standard
deviations, which enables an enhancement of tubular
structures of different diameters. The vesselness images
obtained for the different scales can be combined to a nal
vesselness image using a voxel-wise maximum operator.
3.3. Initial cerebrovascular segmentation
To fuse the benets of the pre-processed TOF image with the
corresponding vesselness image, both datasets are combined
voxel-wisely using fuzzy logic based on an analytically
designed rule base as described in [29]. The main idea of the
rule base used for fuzzy value calculation is to weight the

265

response of the vesselness lter stronger if the TOF intensity is

low whereas the TOF intensity is weighted stronger in case of
low vesselness values. The resulting fuzzy image exhibits high
values for all vessels, especially for small and malformed
vessels, while low values are assigned to the brain tissue. To
obtain the initial vessel segmentation, which is required for the
level-set segmentation process, global intensity thresholding
using a threshold of init is applied to the fuzzy image.
Despite the improved display of vessels in the fuzzy
parameter image, the optimal threshold selection remains
challenging. Lower thresholds lead to an improved small
vessel detection but also come along with an increasing
amount of false-positives caused by noise. In contrast to this,
higher thresholds decrease the number of false-positives but
also lead to a decreased detection of small vessels. The levelset procedure described in the following was especially
designed to overcome this trade-off and enable satisfying
small vessel delineations from the calculated fuzzy parameter images while reducing false-positive segmentations.
3.4. Level-sets with vesselness-dependent anisotropic
energy weights
A variational level-set based segmentation approach in
the style of [30] was used in this work for the extraction of
the nal cerebrovascular segmentation from 3D TOF data
sets. Here, the surface of an object is expressed implicitly as
the zero level-set of the level-set function : with b0
in the object. The optimal level-set was computed in this
work by minimizing the energy functional


J  := E F;  + | ; + V ;
2
where F(x) is the value calculated by fuzzy-based combination of intensity and vesselness information. The functional
consists of two terms representing the internal | and external

Fig. 2. Slice from a TOF image sequence and 3D visualization of the rst eigenvectors calculated by the vesselness lter.

266

N.D. Forkert et al. / Magnetic Resonance Imaging 31 (2013) 262271

energy E as well as an additional term representing the

vesselness force V.
The region-based external energy term, which drives the
contour towards dened image properties, is dened as:
E F;  : = 1H xlogpV F x

+ H xlogpBG F xdx;
where H denotes the Heaviside function, which is used to
describe inside and outside of the object. Using this
formulation an integration of a priori knowledge about
intensity distributions inside (pV) and outside (pBG) the
vessels becomes possible. These intensity distributions can
be estimated by sampling the fuzzy values inside and outside
an initial cerebrovascular segmentation using a ParzenWindow strategy [31] with Gaussian kernels:
!
1
1
ggi 2
pj g =
p exp
4
jGj j gi Gj 2
22
where j {V", BG"} and G denotes the set of sampled
fuzzy values.
More precisely, two thresholds were used in this work:
one for initialization of the level-set segmentation (init) and
one for estimating the fuzzy value probabilities (prob) with
the Parzen-Window technique. Higher values should be
preferred for init such that the initial segmentation contains
mostly true vessels. In contrast to this, lower thresholds are
recommended for prob, which is supposed to enable an
improved evolution into non-segmented vessels.
The internal energy is typically used to keep the object
boundaries smooth. However, small vessels are often
represented by low intensities and edge values. In such a
case, the internal energy is typically stronger than the
external energy, which may lead to an insufcient evolution
into small vessels, even in case of a low prob. One solution
to avoid this problem is to lower the weight for the internal
energy, which enables improved small vessel delineations
but also results in more leakages into non-vascular tissues.
Thus, a local adaption of the internal energy weight may lead
to an improved level-set evolution into small vessels while
preventing false-positive leakages into non-vascular tissues.
For this reason, the internal energy is dened in this work by:
 
| ; : = jj H xjjdx:
5

Here, the parameter c controls the principle smoothing

applied, while its actual inuence is weighted by the second
term. Fig. 3 illustrates the basic idea of this locally adaptive
internal energy weight. If is close to 0 or , the weight
converges against zero and no smoothing is applied, which
allows the level-set to evolve into small vessels. On the other
hand, if the vectors are orthogonal, is equal to c and the
segmentation is smoothed as usual.
Finally, the energy functional includes a vesselness force
term:
p
V  : = V H xcos2 x V xdx:
8
This energy term is used to actively drive the contour
along with the vessels. More precisely, the principle
inuence of this term is weighted by the parameter V,
which is locally adapted by the angle and the vesselness
value V(x) at the actual location. The weight of this term is
highest in case is close to 0 or and high vesselness
values such that a level-set evolution into structures with
high vesselness values are favorable.
The described energy functional was optimized in this
work using an iterative update scheme.
3.5. Implementation details, experiments and evaluation
Ten of the 11 available TOF datasets were independently
segmented by two observers using volume growing and
manual correction in the orthogonal slices. These manual
segmentations were used for interobserver comparison as well
as for evaluation of the automatic segmentations. The
remaining TOF dataset was segmented by only one observer
and used for parameter optimization but not for the evaluation.
The basic vesselness images used in this work for fuzzy
image generation and vessel direction estimation were
calculated by computing the Hessian operator over ve
sigma log-scales ([0.25,1.5]). All parameters required for

The function :[0,c] in this formulation is used to

locally adapt the weight of the internal energy depending on
the angle between e 1, the rst eigenvector calculated by
the vesselness lter, and . It is dened by:


x : = c 1cos2 x
6
with
cosx =

e1 x
1 jjxjj

Fig. 3. Illustration of the vesselness-dependent anisotropic smoothing

procedure. No smoothing is applied in Case (a), normal smoothing is applied
in Case (b).

N.D. Forkert et al. / Magnetic Resonance Imaging 31 (2013) 262271

distance from each centerline voxel to the closest vessel

boundary was calculated using the distance transform
described by Danielsson [34]. Finally, the closest 3D
centerline voxel was determined for each voxel part of the
vascular segmentation and used for denition of the vessel
radius at this location.

the proposed level-set model were optimized using the TOF

dataset that was only segmented by one observer. The
following parameters have been used for segmentation of the
remaining ten TOF datasets: init=92, prob=80, V=5, c=0.5
and 100 iterations for the level-set evolution.
Aside from the automatic segmentation using the
described fuzzy-based level-set segmentation (FLS) approach with anisotropic energy weights, each preprocessed
dataset was also segmented using four other automatic
segmentation approaches for comparison of the proposed
method: the intensity-based ZBS approach [13], the more
sophisticated intensity-based stochastic model segmentation
(SMS) method [17], the fuzzy image segmentation (FIS)
used for initialization of the presented level-set approach and
the level-set curve evolution segmentation (CES) approach
[22]. The calculated fuzzy image and the corresponding
fuzzy image segmentation were also used for initialization of
the curve evolution segmentation as implemented in the
vascular modeling toolkit [32].
The Dice coefcient:
D A; B =

2jABj
jAj + jBj

267

4. Results
Table 1 shows the results of the quantitative evaluation of
the segmentation results using the Dice-coefcient. Here, it
can be seen that the two observers agree with a mean Dicecoefcient D of 0.791 (standard deviation =0.039).
No signicant differences were found regarding the Dicecoefcients of the automatic segmentation methods compared to the manual segmentations of the two observers
(0.36bPb.87). Therefore, the mean Dice-coefcients D
stated in the following have been averaged over both
observers and all ten datasets.
The worst automatic segmentation results were achieved
by the ZBS method (D=0.55, =0.11). Compared to this, the
intensity-based SMS approach lead to signicantly better
cerebrovascular segmentation results (D=0.725, =0.081,
Pb.0001). However, on average all three remaining
segmentation methods using the fuzzy-image as input lead
to better segmentation results regarding the Dice-coefcient.
Comparing these three segmentation techniques, it becomes
apparent that both level-set segmentation techniques lead to
a quantitative improvement of the FIS (D=0.762, =0.038)
used for initialization of both techniques. A comparison of
the two level-set methods reveals that the CES algorithm
performed worse (D=0.783, =0.037) than the proposed
FLS with anisotropic energy weights (D=0.806, =0.028).
The FLS even lead to segmentation results, which are
slightly better than the determined inter-observer agreement.
Overall, the FLS performed signicantly better than any
other automatic segmentation approach tested in this study
(ZBS: Pb.0001, SMS: P=.0004, FIS: P=.048, CES: P=
.0004), while no signicant difference to the results of the
inter-observer comparison was found (P=.305).
Overall, the quantitative results show that the FLS lead to
the best segmentation results in 9 of the 10 datasets

was used for quantitative evaluation of the available

segmentations, whereas A and B denote two segmentations.
Dice-values close to 1 indicate a good consensus. The Dicecoefcient was used for inter-observer comparison as well as
for quantitative comparison of the automatic segmentation
results to the manual gold standard segmentations.
Two-sided t-tests were used to test for signicant
differences between the quantitative Dice-values of the interobserver comparison and those of the automatic segmentation
evaluation. A P value less than .05 was assumed to indicate
statistical signicance. Statistical analysis was performed
using SPSS (version 18.0, SPSS, Chicago, IL, USA).
Furthermore, the number of voxels belonging to small
vessel structures with a radius of less than 0.5 mm was
calculated for all manual and automatic segmentations for
evaluation purposes. For extraction of this parameter in a
voxel-wise manor, the available segmentations were rst
used to calculate the corresponding 3D centerlines using the
method proposed by Lee et al. [33]. In a second step, the

Table 1
Quantitative results of the inter-observer comparison (IOC) and automatic vessel segmentations using the Dice-coefcient: ZBS, SMS, FIS, CES and FLS
Dataset

IOC

ZBS

SMS

FIS

CES

FLS

1
2
3
4
5
6
7
8
9
10

0.732
0.824
0.808
0.767
0.748
0.838
0.785
0.753
0.827
0.825

0.492
0.604
0.598
0.662
0.467
0.473
0.746
0.378
0.483
0.600

0.747
0.701
0.802
0.771
0.669
0.817
0.683
0.761
0.542
0.756

0.720
0.816
0.761
0.790
0.767
0.711
0.790
0.756
0.799
0.711

0.729
0.835
0.787
0.792
0.773
0.794
0.816
0.764
0.814
0.722

0.763
0.838
0.805
0.815
0.785
0.790
0.832
0.785
0.852
0.790

268

N.D. Forkert et al. / Magnetic Resonance Imaging 31 (2013) 262271

Fig. 4. 3D visualizations from the axial and sagittal view using volume rendering of one selected TOF dataset masked with the two manual segmentations (rst
row), Z-buffer (second row, left) and stochastic model segmentation (second row right), fuzzy image (third row, left) and Curve-Evolution segmentation (third
row, right) and fuzzy-level-set segmentation (last row).

evaluated, while the SMS performs best in the remaining

case (Dataset 6).
These results can also be conrmed visually (see Fig. 4).
Here, it becomes apparent that the ZBS is only capable of
extracting vessels with rather large diameters. Overall, the
SMS approach segments more vessels than the ZBS but also
leads to more false-positives caused by noise (see central
region). In contrast to this, the three segmentation schemes
using the fuzzy-images as basis are able to segment many
small vessels while no obvious noise artifacts are present.
However, the result of the proposed FLS exhibits most vessel

structures, which is especially obvious in the top left region

in the axial view.
These visual results can be further conrmed by the
second quantitative evaluation counting the number of voxels
associated to small vessel structures with a radius less than
0.5 mm (see Table 2). The results show that the two
observers delineated an average of 90928 voxels per dataset,
which were determined to belong to small vessel structures.
Comparing the results of the automatic segmentations, it
becomes apparent that the ZBS lead to the worst small vessel
extraction (=28482 voxels). The results reveal that the

N.D. Forkert et al. / Magnetic Resonance Imaging 31 (2013) 262271

269

Table 2
Number of voxels segmented belonging to small vessel structures with a maximal diameter of 0.5 mm for the manual and automatic segmentations
Dataset

IOC

ZBS

SMS

FIS

CES

FLS

1
2
3
4
5
6
7
8
9
10

106482
116061
90638
99450
113376
87046
61541
69327
73461
91901

32858
36825
32517
33416
28258
22765
27681
12023
24989
33497

81284
113204
60317
73353
96776
38300
49467
46615
102882
69935

101022
114445
92085
90595
105975
47430
51939
74430
65835
79106

46065
52349
46207
43218
52163
39963
41040
34372
28186
38282

104703
122169
92071
100586
112706
44518
56885
83222
66430
83306

SMS extracted considerably more small vessels (=73213

voxels) while the FIS (=82286 voxels) and FLS performed
even better (=86660 voxels). In contrast to this, a loss of
small vessels was found for the CES, which can be ascribed
to a thickening of the vessel segmentation in some parts
while other thin vessel structures of the initial segmentation
were eliminated by the CES (=42184 voxels). Nevertheless, the CES generally lead to a better adaption of the
segmentation to the vessel boundaries such that the Dice
similarity measure is still improving compared to the initial
fuzzy-image segmentation.

5. Discussion
The concurrent and exact delineation of small and
malformed vessels is a very challenging task for automatic
methods as well as human observers. It was previously
reported by Nowinski et al. [35] that the sensitivity for small
vessel extraction using their method is as low as 16.5% and it
was estimated that a manual renement would take 8 weeks.
Therefore, perfect ground-truth segmentations are usually not
available for typical clinical datasets. To the authors'
knowledge, an evaluation of inter-observer agreement for
cerebrovascular segmentations has not been performed to date.
The inter-observer agreement evaluation based on ten TOF
datasets performed in this study revealed a mean Dicecoefcient of 0.791, which seems like a good result regarding
the fact that overlap measures are known to be not optimal for
thin structures like vessels. Unfortunately, better suited
validation metrics for this certain purpose are not available yet.
Nevertheless, the evaluation revealed that the proposed
level-set segmentation technique leads to results comparable
to those achieved by human observers. The presented
segmentation method did not perform superior compared
to the other segmentation approaches tested in this work in
only one case. This one dataset exhibits a large apical located
arteriovenous malformation and many large dilated draining
veins, which lead to a suboptimal intensity non-uniformity
correction and also to an insufcient vessel enhancement.
Therefore, the intensity-based stochastic model segmentation, which does not rely on the vesselness image, performed
best in this case.

The second quantitative evaluation counting the number

of voxels belonging to vessel structures with a maximal
radius of 0.5 mm revealed that the integration of the locally
adaptive internal energy weight leads to an improved small
vessel delineation compared to the results of the curve
evolution level-set method, which does not include this
property. The mean number of voxels belonging to such
small vessel structures is also comparable to that achieved by
the human observers. The mean number of voxels
corresponding to small vessel is even higher than that of
the manual segmentations if excluding the problematic
dataset 6 from this analysis. However, this quantitative
measure is also not quite optimal for comparing cerebrovascular segmentations as a general underestimation of the
vessels diameters would also lead to a misleadingly increase
of segmented voxels belonging to small vessels. Furthermore, noise-dependent over-segmentations may also increase this number. This problem was for example observed
for the fuzzy image segmentation method, which generally
leads to smaller estimates of the vessel diameters compared
to the other methods. However, comparing these quantitative
results with visual impression from 3D visualizations
suggest that this parameter still gives a good indication for
the ability of a method to segment small vessels.
Fuzzy techniques have been integrated into various
segmentation methods for different problems, e.g., Refs.
[3638]. The presented method is not the rst work, which
combines fuzzy control techniques with level-set methods.
For example, Ciofolo and Barillot [39] presented a fuzzy
control driven level-set method for the segmentation of brain
tissue. In this approach, a priori atlas knowledge was
combined with image-based intensity information. RivestHnault and Cheriet [40] also proposed a brain segmentation, combining a level-set method and fuzzy techniques. In
their work, fuzzy C-means clustering was used for
initialization of a level-set approach. However, to the
authors' knowledge, a level-set segmentation method in
combination with fuzzy logic has not been proposed yet for
the problem of cerebrovascular segmentation.
It has to be emphasized that the proposed method still
exhibits some drawbacks. Turbulent and fast blood ow may
result in a local reduction of TOF intensities. This problem
can be observed in some of the large vessels and

270

N.D. Forkert et al. / Magnetic Resonance Imaging 31 (2013) 262271

arteriovenous malformation structures. Due to the fact that

the vesselness lter response in such vessels is also low, the
current method is not capable of segmenting them
successfully. However, this does only occur seldom and a
manual correction of this does not require a long time.
Even though the proposed method was especially developed
with the goal to achieve improved malformed as well as small
vessel delineations, gaps within the nal segmentations can still
be observed. Experienced human observers are capable of
delineating the course of a vessel even if it is interrupted by
parts of very low intensities, which is problematic for the
presented method. Therefore, depending on the required
segmentation quality, a post-processing may be necessary.
It has to be pointed out that so far only one TOF dataset
with a corresponding manual segmentation from one
observer has been used for parameter optimization. Therefore, further improvements may be possible by more
sophisticated parameter training methods.
However, most previous segmentation methods have been
evaluated using visual inspection or only a small number of
reference gold-standard segmentations. To the authors'
knowledge such an extensive evaluation with this number
of manual segmentations has not been performed yet.
Therefore, the current database may serve as the basis for
the evaluation of other previously published methods as well
as new automatic cerebrovascular segmentation approaches.
In conclusion, the proposed method allows a signicantly
better extraction of the cerebrovascular system compared to
the other state-of-the-art methods evaluated in this study,
which is even comparable to human observer segmentations.
The integration of locally adaptive energy weights together
with an additional vesselness force allows improved small
vessel delineations compared to level-set methods without
this additional properties.
References
[1] Thrift AG, Dewey HM, Macdonell RA, McNeil JJ, Donnan GA.
Incidence of the major stroke subtypes: initial findings from the North
East Melbourne stroke incidence study (NEMESIS). Stroke 2001;
32(8):17328.
[2] Wiebers DO, Whisnant JP, Huston III J, Meissner I, Brown Jr RD,
Piepgras DG, et al. Unruptured intracranial aneurysms: natural history,
clinical outcome, and risks of surgical and endovascular treatment.
Lancet 2003;362(9378):10310.
[3] Fiehler J, Illies T, Piening M, Saring D, Forkert N, Regelsberger J,
et al. Territorial and microvascular perfusion impairment in brain
arteriovenous malformations. Am J Neuroradiol 2009;30(2):35661.
[4] Handels H, Ehrhardt J. Medical image computing for computersupported diagnostics and therapy. Advances and perspectives.
Methods Inf Med 2009;48(1):117.
[5] Sring D, Fiehler J, Ries T, Forkert ND. Rigid 3D-3D registration of
TOF MRA integrating vessel segmentation for quantification of
recurrence volumes after coiling cerebral aneurysm. Neuroradiology
2012;54(2):1716.
[6] Groden C, Laudan J, Gatchell S, Zeumer H. Three-dimensional
pulsatile flow simulation before and after endovascular coil embolization of a terminal cerebral aneurysm. J Cereb Blood Flow Metab
2001;21(12):146471.

[7] Shapiro LB, Tien RD, Golding SJ, Totterman SM. Preliminary results of
a modified surface rendering technique in the display of magnetic
resonance angiography images. Magn Reson Imaging 1994;12(3):4618.
[8] Ries T, Wegscheider K, Wulff A, Radelfahr K, Sring D, Forkert ND,
et al. Quantification of recurrence volumes after endovascular
treatment of cerebral aneurysm as surrogate endpoint for treatment
stability. Neuroradiology 2011;53(8):5938.
[9] Brunenberg EJ, Vilanova A, Visser-Vandewalle V, Temel Y,
Ackermans L, Platel B, et al. Automatic trajectory planning for deep
brain stimulation: a feasibility study. Med Image Comput Comput
Assist Interv 2007;10(Pt 1):58492.
[10] Kirbas C, Quek F. A review of vessel extraction techniques and
algorithms. ACM Comput Surv 2004;36(2):81121.
[11] Al-Kwifi O, Emery DJ, Wilman AH. Vessel contrast at three Tesla in
time-of-flight magnetic resonance angiography of the intracranial and
carotid arteries. Magn Reson Imaging 2002;20(2):1817.
[12] Davis WL, Warnock SH, Harnsberger HR, Parker DL, Chen CX.
Intracranial MRA: single volume vs. multiple thin slab 3D time-offlight acquisition. J Comput Assist Tomogr 1993;17(1):1521.
[13] Chapman BE, Stapelton JO, Parker DL. Intracranial vessel segmentation from time-of-flight MRA using pre-processing of the MIP Zbuffer: accuracy of the ZBS algorithm. Med Image Anal 2004;8(2):
11326.
[14] Yi J, Ra JB. A locally adaptive region growing algorithm for vascular
segmentation. Int J Imaging Syst Technol 2003;13(4):20814.
[15] Dempster AP, Laird NM, Rubin DB. Maximum likelihood from
incomplete data via the EM algorithm. J R Stat Soc Ser B 1977;39(1):
138.
[16] Wilson DL, Noble JA. An adaptive segmentation algorithm for timeof-flight MRA data. IEEE Trans Med Imaging 1999;18(10):93845.
[17] Hassouna MS, Farag AA, Hushek S, Moriarty T. Cerebrovascular
segmentation from TOF using stochastic models. Med Image Anal
2006;10(1):218.
[18] Sato Y, Nakajima S, Shiraga N, Atsumi H, Yoshida S, Koller T, et al.
Three-dimensional multi-scale line filter for segmentation and
visualization of curvilinear structures in medical images. Med Image
Anal 1998;2(2):14368.
[19] Frangi AF, Niessen WJ, Vincken KL, Viergever MA. Multiscale vessel
enhancement filtering. Lect Notes Comput Sci 1998;1496:1307.
[20] Lorenz C, Carlsen IC, Buzug TM, Fassnacht C, Weese J. A multi-scale
line filter with automatic scale selection based on the Hessian matrix
for medical image segmentation. Lect Notes Comput Sci 1997;1252:
15263.
[21] Lesage D, Angelini ED, Bloch I, Funka-Lea G. A review of 3D vessel
lumen segmentation techniques: models, features and extraction
schemes. Med Image Anal 2009;13(6):81945.
[22] Lorigo LM, Faugeras OD, Grimson WE, Keriven R, Kikinis R, Nabavi
A, et al. CURVES: curve evolution for vessel segmentation. Med
Image Anal 2001;5(3):195206.
[23] Vasilevskiy A, Siddiqi K. Flux maximizing geometric flows. IEEE
Trans Pattern Anal Mach Intell 2002;24(12):156578.
[24] Gooya A, Liao H, Matsumiya K, Masamune K, Dohi T. Effective
statistical edge integration using a flux maximizing scheme for
volumetric vascular segmentation in MRA. Inf Process Med Imag
2007;20:8697.
[25] Suri JS, Liu K, Reden L, Laxminarayan S. A review on MR vascular
image processing: skeleton versus nonskeleton approaches: Part II.
IEEE Trans Inf Technol Biomed 2002;6(4):33850.
[26] Kholmovski EG, Alexander AL, Parker DL. Correction of slab
boundary artifact using histogram matching. J Magn Reson Imaging
2002;15(5):6107.
[27] Sled JG, Zijdenbos AP, Evans AC. A nonparametric method for
automatic correction of intensity nonuniformity in MRI data. IEEE
Trans Med Imaging 1998;17(1):8797.
[28] Forkert ND, Saring D, Fiehler J, Illies T, Moller D, Handels H.
Automatic brain segmentation in time-of-flight MRA images. Methods
Inf Med 2009;48(5):399407.

N.D. Forkert et al. / Magnetic Resonance Imaging 31 (2013) 262271

[29] Forkert ND, Schmidt-Richberg A, Fiehler J, Illies T, Moller D,
Handels H, et al. Fuzzy-based vascular structure enhancement in timeof-flight MRA images for improved segmentation. Methods Inf Med
2011;50(1):7483.
[30] Schmidt-Richberg A, Handels H, Ehrhardt J. Integrated segmentation
and non-linear registration for organ segmentation and motion field
estimation in 4D CT data. Methods Inf Med 2009;48(4):3449.
[31] Parzen E. On estimation of a probability density function and mode.
Ann Math Stat 1962;33:106576.
[32] Piccinelli M, Veneziani A, Steinman DA, Remuzzi A, Antiga L. A
framework for geometric analysis of vascular structures: application to
cerebral aneurysms. IEEE Trans Med Imaging 2009;28(8):114155.
[33] Lee TC, Kashyap RL, Chu CN. Building skeleton models via 3-D
medial surface/axis thinning algorithms. Graph Models Image Process
1994;56(6):46278.
[34] Danielsson P. Euclidean Distance Mapping. Comput Graph Image
Process 1980;14:22748.
[35] Nowinski WL, Volkau I, Marchenko Y, Thirunavuukarasuu A, Ng
TT, Runge VM. A 3D model of human cerebrovasculature derived

[36]

[37]

[38]

[39]

[40]

271

from 3T magnetic resonance angiography. Neuroinformatics 2009;

7(1):2336.
Lin GC, Wang CM, Wang WJ, Sun SY. Automated classification of
multispectral MR images using unsupervised constrained energy
minimization based on fuzzy logic. Magn Reson Imaging 2010;28(5):
72138.
Udupa JK, Wei L, Samarasekera S, Miki Y, van Buchem MA,
Grossman RI. Multiple sclerosis lesion quantification using fuzzyconnectedness principles. IEEE Trans Med Imaging 1997;16(5):
598609.
Shen S, Szameitat AJ, Sterr A. An improved lesion detection approach
based on similarity measurement between fuzzy intensity segmentation
and spatial probability maps. Magn Reson Imaging 2010;28(2):24554.
Ciofolo C, Barillot C. Atlas-based segmentation of 3D cerebral
structures with competitive level sets and fuzzy control. Med Image
Anal 2009;13(3):45670.
Rivest-Henault D, Cheriet M. Unsupervised MRI segmentation of
brain tissues using a local linear model and level set. Magn Reson
Imaging 2011;29(2):24359.