Computerized Medical Imaging and Graphics 48 (2016) 21–29

Contents lists available at ScienceDirect

Computerized Medical Imaging and Graphics

journal homepage: www.elsevier.com/locate/compmedimag

A novel level set model with automated initialization and controlling

parameters for medical image segmentation
Qingyi Liu a , Mingyan Jiang a,∗ , Peirui Bai b , Guang Yang a
a
School of Information Science and Engineering, Shandong University, Jinan 250100, China
b
College of Electronic, Communication and Physics, Shandong University of Science and Technology, Qingdao 266590, China

a r t i c l e i n f o a b s t r a c t

Article history: In this paper, a level set model without the need of generating initial contour and setting controlling
Received 25 December 2014 parameters manually is proposed for medical image segmentation. The contribution of this paper is
Received in revised form 6 November 2015 mainly manifested in three points. First, we propose a novel adaptive mean shift clustering method
Accepted 9 December 2015
based on global image information to guide the evolution of level set. By simple threshold processing,
the results of mean shift clustering can automatically and speedily generate an initial contour of level set
Keywords:
evolution. Second, we devise several new functions to estimate the controlling parameters of the level set
Medical image segmentation
evolution based on the clustering results and image characteristics. Third, the reaction diffusion method
Adaptive mean shift clustering
Level set method
is adopted to supersede the distance regularization term of RSF-level set model, which can improve
Automated initialization the accuracy and speed of segmentation effectively with less manual intervention. Experimental results
Reaction diffusion method demonstrate the performance and efficiency of the proposed model for medical image segmentation.
© 2015 Elsevier Ltd. All rights reserved.

1. Introduction
The purpose of medical image segmentation is to separate the
components of the interested region from their background. The
result can be used to assist doctors to treat their patients by making
a diagnosis. Since the real medical image is mostly limited by poor
resolution, weak contrast and intensity inhomogeneity [1], it is dif-
ficult to segment accurately by traditional segmentation method,
such as the threshold or region growing method. In recent years,
level set methods have become increasingly popular for medical
image segmentation [2,3]. There are several advantages of the level
set method [4]. For example, its numerical computations involv-
ing curves and surfaces can be performed on a fixed Cartesian grid.
Moreover, it is able to represent contours with complex topological
changes [5].
The level set methods can be categorized into edge-based mod-
els and region-based models [6,7]. Edge-based models use the local
edge information to drive the active contour toward the object
boundaries. However, the edge-based models are usually quite
sensitive to the initial conditions and easily suffer from bound-
ary leakage problems in images with weak object boundaries [8].
Region-based models aim to identify each interested region by
∗ Corresponding author. Tel.: +86 13361073798.
E-mail address: jiangmingyan@sdu.edu.cn (M. Jiang).
using a certain region descriptor to guide the motion of the active
contour [9]. Nevertheless, defining a certain region descriptor is
difficult for images with intensity inhomogeneity or weak con-
trast. Being different from classical methods based on assumption
of intensity homogeneity, a level set model based on local region
information, which is called region-scalable fitting (RSF) level set
model, was proposed by Li et al. [13]. It uses RSF energy term to
guide the motion of the contour. This method is able to segment
the images with intensity inhomogeneity. However, all these meth-
ods are sensitive to the initialization of the contour and need to set
controlling parameters manually.
In traditional level set models, the periodical re-initialization is
necessary to maintain stable evolution of the zero level set curves
[10]. However, it may cause undesirable results for image seg-
mentation and add unnecessary computation cost. Then, Li et al.
proposed a classical method which is called distance regularized
level set evolution (DRLSE) [11] which regularizes the level set
function during evolution. Therefore, the re-initialization proce-
dure can be eliminated. Subsequently, this method is used in many
new or classical level set models [12,13]. However, DRLSE method
still has the drawbacks such as limited anti-leakage capability for
weak boundaries and sensitivity to noise. Recently, a new method
was proposed for remit the re-initialization of the level set evolu-
tion, which is called reaction-diffusion (RD) method. This method
has much better performance on weak boundary anti-leakage, and
the implementation of the RD equation is very simple and it is more

http://dx.doi.org/10.1016/j.compmedimag.2015.12.005
0895-6111/© 2015 Elsevier Ltd. All rights reserved.
22 Q. Liu et al. / Computerized Medical Imaging and Graphics 48 (2016) 21–29

n
  
robust to noise. These advantages of RD method have been veri-
where g (X − Xi ) /h2 is assumed to be a positive num-
i=1
fied by applying it to the classical level set model, geodesic active
contours (GAC) model and Chan–Vese (C–V) model. ber. This condition is easy to satisfy for all the profiles met in
In this paper, we propose a novel level set model for medi- practice. From Eq. (3), the first term is proportional to the density
cal image segmentation. Compared to some previous models, our estimate at X computed with the kernel:
model has significant contributions in the following aspects. First,   
cg,d 
n
 X − Xi 2
we improve the mean shift clustering method by a new adaptive f̂h,G (X) = g  h  (4)
bandwidth algorithm which makes the clustering more accurate nhd
i=1
and faster. Furthermore, this improvement also solves the problem
that the previous mean shift algorithm is quite sensitive to the set- and the second term is the popular mean shift vector:
ting of bandwidth. Second, the novel mean shift clustering method n
  
(X − Xi ) /h2
Xi g
is used to obtain the initial contour of RSF-level set evolution close i=1
to the genuine boundaries of objects and estimate the controlling
 2  − X (5)
parameters of the level set model. Third, we apply the reaction dif-
fusion method instead of the previous distance regularized method
to avoid the re-initialization of RSF-level set evolution. It makes this
algorithm’s iteration faster and more robust to noise.
This paper is structured into the following sections. Section 2
describes the background and related works. Section 3 presents
the novel level set model. Experimental results and the relevant
discussions are shown in Section 4. The final section of the paper
offers some conclusions.
mh,G(x) = 
g (X − Xi ) /h
i=1
n
Eq. (5) shows that the mean shift vector computed with kernel
G (x) is proportional to the normalized density gradient obtained
with kernel K (x) at location X. The mean shift vector always points
to the direction of maximum increase in the density. The first step
of the mean shift procedures is to compute the mean shift vector
mh,G(x) . Then, translate the kernel G (x) by mh,G(x) . At the end of the
procedure, it is guaranteed to converge at a nearby point until some
convergence criterions are satisfied.

2. Background and related works 2.2. Reaction diffusion method

2.1. Mean shift clustering The evolution of the classical level set function  (t, x, y) can be
expressed as follows:
Mean shift clustering is a simple iterative procedure that shifts ⎧  
⎨ ∂
each data point to the mean of data points in its assigned neigh- + F ∇   = 0
borhood [14], and it is widely used in computer vision and image ∂t (6)
processing field [15,16]. ⎩
 (t = 0, x, y) = 0 (x, y)
The image segmentation with mean shift clustering is a straight-
forward extension of the discontinuity preserving smoothing   the speed function that controls the motion of
where F indicates
algorithm. Each pixel associates with a significant mode of the joint the contour. ∇  denotes the normal direction, and 0 (x, y) is the
domain density located in its neighborhood [17]. initial contour.
Given n data points Xi (i = 1, 2, . . ., n) in a d-dimensional Euclid- During evolution, the traditional level set function may become
ian space Rd and a symmetric positive definite d × d bandwidth too flat or too steep near the zero level set, causing serious
matrix H, which is proportional to the identity matrix, i.e., H = h2 I, numerical errors. Therefore, a procedure called re-initialization is
where h > 0 is a bandwidth parameter. The multivariate kernel den- periodically employed to keep it to be a signed distance function
sity estimator with bandwidth matrix H is given by the following (SDF). Unfortunately, this re-initialization method is very time-
equation: consuming and may fail when the LSF deviates much from SDF.
  
To eliminate the re-initialization procedure, some formulations
ck,d 
n
 X − Xi 2 have been proposed in recent years. One of the most common meth-
f̂h,K (X) =   h  (1)
ods is distance regularized level set evolution [10,11]. The authors
nhd
i=1 propose a signed distance penalizing energy term as follows:

where ck,d denotes a positive normalization constant,  (x) indicates 1   
the function called the profile of the kernel, only for x ≥ 0. P () = ∇  (x) − 1 2 dx (7)
2
We define a function g (x) = − (x). Then, we assume the deriva- ˝

tive of the kernel profile  (x) exists for all x ∈ [0, ∞) , except for a Recently, a novel means has been put forward to avoid re-
finite set of points, the kernel G (x) can be redefined as: initialization by reaction diffusion function [18].
  By adding a diffusion term ε∅ into the level set equation in Eq.
G (x) = cg,d g x2 (2) (6), RD function for level set model can be defined as:
⎧  
where cg,d denotes the corresponding normalization constant. ⎨ t = ε∅ + 1 F ∇ 
According to the linearity of Eq. (1) and introducing g (x) into ε (8)
it, the density gradient estimator is obtained as the gradient of the ⎩
 (0, x, ε) = 0 (x)
density estimator as follows:
  where ε represents a small positive constant, and  indicates the

n  
2ck,d  X − Xi 2 Laplacian operator. The diffusion term ε∅ gradually regularizes
∇ˆ fh,G (X) = g  
nhd+2 h   constant
the LSF to be a piecewise  in each segment domain, and
i=1 the reaction term 1/ε F ∇  forces the final stable solution of
  
⎡ i=1   ⎤ Eq. (8) to F ∇  = 0.
X i g (X − Xi ) /h2
n The equilibrium solution of Eq. (8) is a piecewise constant as
×⎣   2  − X⎦ (3)
ε → 0+ , which is the characteristic of phase transition. However, RD
g (X − Xi ) /h
i=1
n equation has the intrinsic problem of phase transition, which is the
 Q. Liu et al. / Computerized Medical Imaging and Graphics 48 (2016) 21–29 23

stiff parameter ε−1 that makes Eq. (8) difficult to implement. There- where t0 is a constant threshold which is suggested to be 1 in our
fore, a splitting method has been presented to reduce the side effect model.
of the stiff parameter, which is called two-step splitting method To match with the initial contours defined in Eq. (10), a regula-
(TSSM). tion process should be installed as follows:
 In the first step of TSSM, we solve the reaction term t = 0 (x, y) = −6 (0.5 − D (x)) (12)
1/ε F ∇  with  (0, x, ε) = ∅n to obtain the intermediate solu-
tion. The diffusion equation t = ε∅ with  (t = 0, x) = ∅(n+1)/2 is By using Eq. (11), we obtain a two-value matrix 0 (x) ranged
solved to ensure the smoothness of the level set function and keep from −3 to 3. It satisfies the initial contour demands in the RSF-level
zero level set not to be moved. set model [13].
In order to improve the robust to noise and computational
 
By choosing small time steps t1 and t2 , the above operation
will be discretely approximated as ∅(n+1)/2 = ∅n + t1 F ∇  and efficiency of our model, we apply the reaction diffusion method
  instead of the previous distance regularized method to avoid the re-
∅n+1 = ∅(n+1)/2 + t2 ∅(n+1)/2 . Obviously, the parameter ε can
  initialization of RSF-level set evolution. The novel level set function
be integrated into the time steps t1 and t2 as t1 ← t1 −ε−1 is defined as follows:
and t2 ← t2 ε. Therefore, we only need to consider the two time ⎧  
steps to keep numerical stability. ⎪
⎨  = ı () div ∇  − ı () ˇ e (x) − ˇ e (x)
The RD method has been applied to the classical level set model, t
∇   1 1 2 2
(13)
GAC model and C–V model [18]. The segmentation experiments ⎪
⎩
prove the property superiority of RD method such as robust to noise  (t = 0, x) = −6 (0.5 − D (x))
and high computational efficiency. Through the comparison with previous RSF-level set function,
it is obvious that the signed distance penalizing energy term is no
3. The proposed model longer required. Therefore, the calculation cost will be reduced to
some extent, and the parameter  can be removed. It will reduce
The RSF-level set model has several desirable advantages such the manual operation of level set evolution.
as flexible initialization and effective solving scheme. In addition, With this proposed model, the parameters ˇ1 and ˇ2 can be set
it has a unique advantage in coping with intensity inhomogeneity. as 1 without the loss of segmentation accuracy because the regular-
Nevertheless, the initial contour and evolution controlling param- ized clustering results approximate the genuine boundaries well.
eters have to be set manually. To reduce the unnecessary manual However, the evolution controlling parameter  and scale param-
intervention and improve efficiency of the algorithm, a novel auto- eter of Gaussian kernel used to compute e1 (x) e2 (x) in Eq. (13)
mated level set model is thereby proposed for medical image still have significant influence on the segmentation result. Hence,
segmentation. It automates the initialization and parameter config- we propose following formulas to estimate the parameters  and
uration of the level set method which includes integrating adaptive in our model.
mean shift clustering and RD method with the RSF-Level set model. The evolution controlling parameter  can be estimated from the
The mean shift clustering is a general nonparametric method, result of mean shift clustering and characteristics of the segmented
which is based on the global image information without any prior image as follows:
knowledge. However, the accuracy and efficiency of this method hmax + hf
are sensitive to the manual setting of bandwidth parameter h. To = × max (I) (14)
2n2
solve this problem, we propose an adaptive formulation for band-
where n represents the subclass number of the adaptive mean shift
width parameter selection as follows:
clustering, which is obtained automatically during the adaptive
2
 mean shift clustering procedure. hf denotes the bandwidth which
hn = hmax 1 − ch e−((mn −mn−1 )/mn ) (9)
is calculated in the last iteration of adaptive mean shift clustering,
and hmax is the maximum bandwidth set manually.
where mn represents the mean of total pixel values in current The scale parameter in the definition of K is difficult to be
search region, ch is weighted coefficients, and hmax is the predefined set in the RSF-level set method. When a small is chosen, the seg-
maximum bandwidth. Obviously, a larger bandwidth parameter mentation will be more sensitive to the initial contour. However, if
will be chosen when the pixel value distributes more uniform. we choose a large , more robust initial contour can be obtained,
Conversely, a smaller bandwidth parameter will be chosen. but the computational cost and segmentation accuracy will have
Benefitting from the flexible initialization of RSF-level set model, a negative effect [15]. We propose an automated parameter esti-
the initial contour 0 (x) can be defined as the following equation: mation formula to find an appropriate scale parameter for our
 model. First, we defined the length l and the area s by
−C, x ∈ ˝0 
0 (x) = (10)
C, x ∈ ˝ − ˝0 l= ı (0 ) dx (15)

where ˝0 denotes a subset in the image domain ˝, and C is a con-
s= H (0 ) dx (16)
stant. Eq. (10) shows that the evolution of the level set function
can start from an arbitrary binary region. In traditional level set
where 0 is the initial contours, ı (0 ) indicates Dirac function
model, the initial contour with a regular shape, such as a circle or
slightly smoothed, and H (0 ) is Heaviside function as:
a rectangle, must be manually obtained.

In our model, we implement the adaptive mean shift clustering 1, 0 ≥ 0
first, B (x) is the results of clustering. Then a two-value matrix D (x) H (0 ) = (17)
is obtained from the clustering result by a simple thresholding: 0, 0 < 0
 Then, the scale parameter is estimated by
1 B (x) > t0
D (x) = (11) l
0 other = c (18)
ns
24 Q. Liu et al. / Computerized Medical Imaging and Graphics 48 (2016) 21–29

Fig. 1. RSF-level set segmentation results of two test images. Panels (a)–(f) show the segmentation procedures applied to the T-shape image with different initial contour
and the same parameter setting. Panels (g)–(l) show the segmentation procedures applied to the ultrasound image of vessel transverse section with different initial contour
and the same parameter setting. The parameters for the T-shape image are ˇ1 = ˇ1 = 1,  = 195.075 and = 3.0. The parameters for the ultrasound image of vessel transverse
section are ˇ1 = 1, ˇ2 = 2,  = 1300.5, and = 15. The first column images show the manually initial contours denoted with rectangles. The second column images show the
segmentation results after 50 iterations. The third column images show the final segmentation results.

where n is the subclass number of the adaptive mean shift cluster-
ing. c is the regulating parameter, and we use it with c = 500 for
all the experiments in this paper.
In the proposed model, only one parameter, the maximum band-
width in the mean shift clustering needs to be set manually, and it
is not necessary to be selected very accurately. In addition, a more
efficient method to avoid re-initialization has been introduced into
our model. Undoubtedly, those improvements have several prac-
tical benefits, such as making the segmentation more efficient and
higher accuracy.
4. Experimental results and discussion
The proposed model was implemented with Matlab R2010a
on PC with Intel Core i3-3240 CPU 3.4 GHz and 8 GB RAM. The
operation system is Windows 7 service pack 1. The experiments
and performance evaluation were carried on synthetic images and
real medical images from different modalities. Unless otherwise
specified, we used the parameters ch = 0.5, c = 500, ˇ1 = ˇ2 = 1, and
time steps t1 = 0.1 and t2 = 0.05 for all the experiments in this
paper.
The first experiment is designed to evaluate the usefulness of
adaptive mean shift clustering for level set segmentation. We used
previous RSF-level set model with manual initialization for a real
image of T-shaped object and an ultrasound image of vessel trans-
verse section, respectively. Meanwhile our novel method was used
to segmentation of the same images. The segmentation results,
the number of level set iterations and the computation time were
recorded as the indexes of measuring the algorithm efficiency.
Fig. 1 shows the segmentation procedures of the two test images
using the RSF-level set model with different initial contours, respec-
tively. The first column images (Fig. 1a, d, g, and j) show the
manual initial contours denoted with rectangles. The second col-
umn images show the segmentation results after 50 iterations, and
 Q. Liu et al. / Computerized Medical Imaging and Graphics 48 (2016) 21–29 25

Fig. 2. Our model segmentation results of the same images. The first and second rows show separately the segmentation procedures applied to the T-shape image (Fig. 2a–d)
and the ultrasound image of vessel transverse section (Fig. 2e–h) with hmax = 20 and 10, respectively. The first column shows the original images. The second column shows
the results of adaptive the mean shift clustering. The third column shows the initial contours obtained by regulating the clustering results. The fourth column shows the
segmentation results of our model.

the third column images show the final segmentation results. The
RSF-level set method achieves satisfied segmentation result for the
T-shaped object image with 320 iterations (Fig. 1c), but fails to
achieve satisfied segmentation result for the ultrasound image of
vessel transverse section over 500 iterations (Fig. 1i). With another
initial contour, the RSF-level set method cannot correctly segment
the objects (Fig. 1f and l). The computation time of four segmenta-
tion procedures are 9.48 s, 9.33 s, 19.30 s, 18.63 s, respectively.
Fig. 2 illustrates the segmentation procedures of our novel
model applied to the same images as Fig. 1. It is not difficult to
see that satisfactory segmentation results (Fig. 2d and h) of both
test images were obtained. Benefiting from the appropriate ini-
tialization, our novel method successfully extracts the boundaries
of interesting objects. The computation times needed to segment
the T-shape object and ultrasound image are 1.65 s and 4.52 s.
Apparently, the computational cost of our model is significantly
reduced.
According to the above results comparison between segmen-
tation results of RSF-level set model and the model proposed in
this paper, we find that the new algorithm in the aspect of the
computing efficiency and the accuracy of segmentation is superior
to the traditional RSF-level set model. The primary reasons have
the following aspects. First, RSF-level set model is very sensitive
to the initialization of the contour, so that the final result is largely
influenced by the initial contour. However, initial contour of level
set evolution is an automatic contour which is automatically
generated by mean shift clustering results in our model. Second,
the equation evolution is just driven by employing the local region
information of image, which is easily affected by noise and weak
edges. In comparison, since mean shift algorithm in our model is
based on the global region information, more comprehensive image
information is utilized for image segmentation. Third, the compu-
tation speed of adaptive mean shift algorithm is much faster than
RSF-level set model. Level set evolution in our algorithm is starting
from the mean shift clustering results which are approached object
contour, rather than the manual initial contour, so that the overall
computation time of the algorithm is reduced significantly.
The target of the second experiment was to evaluate the effec-
tiveness of RD method to avoid the re-initialization in level set
evolution. A synthetic image and an X-ray image of vessel were

Table 1
The comparison of computation time (s) and iteration numbers of different models for suprasellar section image, sellar section image and metacarpal image.

Model Suprasellar section image Sellar section image Metacarpal image

t0 t n t0 t n t0 t n

GAC model – 25.86 1500 – 19.36 1500 – 8.1 1500

C–V model – 25.10 500 – 11.67 500 – 9.92 500
Fuzzy level set model 0.66 27.85 100 0.67 6.36 100 0.73 15.19 300
MS-RSF model 0.16 23.74 100 0.14 5.27 100 0.12 5.12 100
Our model 0.11 21.26 100 0.11 2.65 100 0.11 3.45 100

Table 2
The comparison of segmentation accuracy of different models for suprasellar section image, sellar section image and metacarpal image.

Model Suprasellar section image Sellar section image Metacarpal image

Dice Hausdorff (pixel) Dice Hausdorff (pixel) Dice Hausdorff (pixel)

GAC model 0.3010 58.600 0.4844 60.374 0.3877 46.819

C–V model 0.8652 28.461 0.8829 28.018 0.9175 20.223
Fuzzy level set model 0.8544 58.181 0.9129 12.207 0.9708 8.623
MS-RSF model 0.8876 41.182 0.9670 8.062 0.9607 20.099
Our model 0.8911 5.831 0.9687 6.083 0.9834 3.000
26 Q. Liu et al. / Computerized Medical Imaging and Graphics 48 (2016) 21–29

Fig. 3. Our model segmentation results with different methods to avoid the re-initialization. The first and third rows show our model with DRLSE procedures applied to
a synthetic image and X-ray image with hmax = 25 and 15, respectively. The second and fourth rows show our model with RD method procedures applied to a synthetic
image and X-ray image with the hmax = 25 and 15, respectively. The first column shows the original images. The second column shows the results of the adaptive mean shift
clustering. The third column shows the initial contours obtained by mean shift clustering results. The fourth column shows the final segmentation results.

segmented by our model with two different avoid re-initialization
methods which are traditional DRLSE method and RD method,
respectively. The comparing of the segmentation results is shown
in Fig. 3.
Fig. 3 illustrates the performance of different methods to avoid
the re-initialization in our model. The synthetic image and X-ray
image segmentation procedures of our model with DRLSE method
were shown in the first and third row, and the same images seg-
mentation procedures of our model with RD method were shown
in the second and fourth row. The fourth column images are the
final results, and their computation times are 5.90 s, 5.49 s, 1.48 s
and 1.34 s, respectively.
The validity of the improvement in the way to avoid the
re-initialization for our model is proved by the second experi-
ment. Although DRLSE method can effectively avoid the tedious
processing of re-initialization for the RSF-level set evolution.
However, as we can see from Fig. 3d and l, DRLSE method causes
the segmentation results of RSF-level set model to be more
sensitive to noise. Hence, a new way to avoid the re-initialization
is introduced to our model, which improves noise-resistibility
and further enhances the computing efficiency. This improvement
brings about a satisfactory segmentation effect, as shown in Fig. 3h
and p.
In the third experiment, we compare our model with four dif-
ferent level set methods, including two classical level set models
and two kinds of automatic initialization level set models which
are GAC model, C–V model, fuzzy level set model [19] and MS-RSF
model [20], respectively. GAC model is a classical edge-based level
set model. The convergence of active contour curve in GAV model
is driven by the image gradient. C–V model is also called active
contour without edges, which is driven by region information. GAC
model and C–V model need to set initial contour and controlling
parameters manually. Fuzzy level set model applies spatial fuzzy
clustering to extract global region information, and the cluster-
ing results are utilized for initial level set function. The clustering
number needs to set in advance. MS-RSF model is another level set
 Q. Liu et al. / Computerized Medical Imaging and Graphics 48 (2016) 21–29 27

Fig. 4. Segmentation results of suprasellar section CT image with five level set methods. The first row shows the initial contours; the second row shows the final segmentation
results.

method without manual initial contour. However in this method,
most controlling parameters still need to set manually, and it uses
DRLSE method to avoid the re-initialization.
Fig. 4 shows the segmentation results for suprasellar section
CT image; Fig. 5 shows the segmentation results for sellar section
CT image and Fig. 6 shows the segmentation results for transverse
metacarpal bones CT image. The above figures from the column (a)
to column (e) are GAC model, C–V model, fuzzy level set model,
MS-RSF model and our model, respectively. The brain images come
from our cooperative hospital. They are two CT slice images from
two different patients. The metacarpal image is in an open dataset
which comes from the Institute of Sport and Exercise Medicine,
North University of China. It is a left hand CT image from a volleyball
athlete. Table 1 shows the computation times and iteration num-
bers of these segmentation results, including initialization times
(t0 ) and overall segmentation times (t). Because the GAC model
and C–V model need to initialize manually, the initialization times
cannot be recorded.
By observing Figs. 4–6, we can discover that the segmentation
results of other four methods are imprecise. The GAC model only
can segment one interested boundary contour which is not suffi-
ciently precise. A high precision could be obtained with C–V model
and MS-RSF model, but they are easy to generate excess bound-
ary contour. The final contour of fuzzy level set method cannot
penetrate the deeper concave part of object boundary. Neverthe-
less, segmentation results of our model have greatly improved for
the accuracy of segmentation, and are not apt to generate pseudo
boundary. As we can see from Table 1, for the above tested three
images, our model is the fastest. On the other side, by comparing
the generating speed of the initial contour between our model and
MS-RSF model, the effectiveness of adaptive bandwidth algorithm
we proposed is demonstrated.
In order to quantitative measuring the segmentation accuracy
of our model, we use the Dice coefficient and the Hausdorff dis-
tance as standard. The segmentation results by human are used as
the golden standard, which are shown in Fig. 7, and then the Dice

Fig. 5. Segmentation results of sellar section CT image with five level set methods. The first row shows the initial contours; the second row shows final segmentation results.
28 Q. Liu et al. / Computerized Medical Imaging and Graphics 48 (2016) 21–29

Fig. 6. Segmentation results for transverse section of metacarpal CT image with five level set methods. The first column shows the initial contours; the second column shows
final segmentation results.

Fig. 7. Ground truth segmentation for three images. The black contours are manual segmentation results.
 Q. Liu et al. / Computerized Medical Imaging and Graphics 48 (2016) 21–29
coefficient and the Hausdorff distance of five different models’ seg-
mentation results are computed, and the Hausdorff distance is in
pixel. For larger dice coefficients and smaller Hausdorff distance,
the higher segmentation accuracy is obtained. As we can see from
the comparing results of Table 2, our algorithm has the highest
accuracy.
Medical images make it difficult to be segmented precisely, since
they have their own inherent deficiencies, such as poor resolution,
serious noise and intensity inhomogeneities. The images which
were used in our experiments have their own different characteris-
tics. The ultrasound image of vessel transverse section has serious
noise and obvious weak boundary, the X-ray images are typical
images with intensity inhomogeneities, and the CT images have
more complex object boundaries. However, the above experiments
verify that our model yields significant performance advantages
in computational efficiency and segmentation results with both
manual initialization level set model and automatic initialization
level set model. It is worth mentioning that our model is a level set
model with automated initialization and controlling parameters,
and reduces the influence of manual intervention on the algorithm
distinctly.
To integrate a novel adaptive mean shift with the RSF-level set
model, the key issues are matching the clustering results with the
initial contours of the level set evolution and ensuring the posi-
tive effects of our model. Above experimental results demonstrate
the effectiveness of the binarization processing, regulation opera-
tion, and parameter estimation equation in our model. Moreover,
by using the RD method that we used to avoid re-initialization in
our model, the level set evolution becomes faster and more robust
to noise.
5. Conclusion
A novel level set model was proposed for automated medi-
cal image segmentation in this paper. Our model began with the
mean shift clustering which was modified by adaptive bandwidth
algorithm, and utilized the results of this clustering to initiate the
level set formulation. The clustering results could be the initial
contour which approximated the boundaries of interest region by
binarization processing and regulation operation. The controlling
parameters of the level set evolution also could be automatically
estimated from the result of mean shift clustering and charac-
teristics of the segmented image by the function we proposed.
Moreover, in order to enhance the robust to noise and computing
efficiency, RD method was used to be the new way to eliminate re-
initialization in our model. The experimental results presented in
this study validate the accuracy and the effectiveness of our model
with all these improvements for segmenting medical images auto-
matically. However, our method does exist some limitations. This
method still cannot completely eliminate the manual intervention
in the processing of segmentation and hmax still needs be chosen
according to experience. In future studies, we will investigate better
29
method of the automated initialization for level set model. We will
also explore more effective methods to drive the evolution of the
level set curves according to the characteristics of medical images.
Acknowledgements
We would like to thank the helpful comments and sugges-
tions from the anonymous reviewers. This work was supported by
the National Natural Science Foundation of China (No. 61471225),
the National Natural Science Foundation of China (No. 61571275);
the National Science Foundation of Shandong province, China (No.
ZR2014FM039).
References
[1] Pham DL, Xu CY, Prince JL. A survey of current methods in medical image
segmentation. Annu Rev Biomed Eng 2000;2:315–37.
[2] Wu Pengfei, Liu Yiguang, Li Yongzhong. Robust prostate segmentation using
intrinsic properties of TRUS images. IEEE Trans Med Imaging 2015;34:1321–35.
[3] Yang Xi, Gao Xinbo, Tao Dacheng. An efficient MRF embedded level set method
for image segmentation. IEEE Trans Image Process 2015;24:9–21.
[4] Singh Amarjot, Au Jason, Saeedi Parvaneh. Automatic segmentation of trophec-
todermin microscopic images of human blastocysts. IEEE Trans Biomed Eng
2015;62:382–93.
[5] Paragios N. A level set approach for shape-driven segmentation and tracking
of left ventricle. IEEE Trans Med Imaging 2003;22:773–6.
[6] Vasilevskiy A, Siddiqi K. Flux-maximizing geometric flows. IEEE Trans Pattern
Anal Mach Intell 2002;24(12):1565–78.
[7] Vese L, Chan T. A multiphase level set framework for image segmentation using
the Mumford and Shah model. Int J Comput Vis 2002;50:271–93.
[8] Caselles V, Kimmel R, Sapiro G. Geodesic active contours. Int J Comput Vis
1997;22:61–79.
[9] Li CM, Huang R, Ding ZH. A level set method for image segmentation in the pres-
ence of intensity inhomogeneities with application to MRI. IEEE Trans Image
Process 2011;20:2007–15.
[10] Li CM, Xu CY, Gui CF, Fox MD. Level set evolution without re-initialization: a new
variational formulation. IEEE Comput Soc Conf Comput Vis Pattern Recognit
2005;1:430–6.
[11] Li CM, Xu CY, Gui CF, Fox MD. Distance regularized level set evolu-
tion and its application to image segmentation. IEEE Trans Image Process
2010;19:3243–54.
[12] Li CM, Kao CY, Gore JC, Ding ZH. Implicit active contours driven by local binary
fitting energy. IEEE Int Conf Comput Vis Pattern Recognit 2007;1:1–7.
[13] Li CM, Kao CY, Gore JC, Ding ZH. Minimization of region-scalable fitting energy
for image segmentation. IEEE Trans Image Process 2008;17:1940–9.
[14] Comaniciu D, Meer P. Mean Shift: a robust approach toward feature space
analysis. IEEE Trans Pattern Anal Mach Intell 2002;24:603–19.
[15] Cho Hanjoo, Kang Suk-Ju, Cho Sung In. Image segmentation using linked mean-
shift vectors and its implementation on GPU. IEEE Trans Consum Electron
2014;60:719–27.
[16] Banerjee Biplab, Surender Varma G, Buddhiraju Krishna Mohan. Satellite image
segmentation: a novel adaptive mean-shift clustering based approach. In: Geo-
sciences and Remote Sensing Symposium, IEEE International. 2012. p. 4319–22.
[17] Bai PR, Fu CH, Cao MY, Han Y. Improved mean shift segmentation scheme for
medical ultrasound images. In: 4th International conference on bioinformatics
and biomedical engineering vol. 1. 2010. p. 1–4.
[18] Zhang KH, Zhang L. Reinitialization-free level set evolution via reaction diffu-
sion. IEEE Trans Image Process 2013;22:258–71.
[19] Li BN, Chui CK, Chang S, Ong SH. Integrating spatial fuzzy clustering with level
set methods for automated medical image segmentation. Comput Biol Med
2011;41:1–10.
[20] Bai PR, Liu QY, Li L. A novel region-based level set method initialized with mean
shift clustering for automated medical image segmentation. Comput Biol Med
2013;43:1827–32.