Journal Pre-proof

Local segmentation of images using an improved fuzzy C-means

clustering algorithm based on self-adaptive dictionary learning

Jiaqing Miao, Xiaobing Zhou, Ting-Zhu Huang

PII: S1568-4946(20)30140-X
DOI: https://doi.org/10.1016/j.asoc.2020.106200
Reference: ASOC 106200

To appear in: Applied Soft Computing Journal

Received date : 20 February 2018

Revised date : 21 February 2020
Accepted date : 26 February 2020

Please cite this article as: J. Miao, X. Zhou and T.-Z. Huang, Local segmentation of images using
an improved fuzzy C-means clustering algorithm based on self-adaptive dictionary learning,
Applied Soft Computing Journal (2020), doi: https://doi.org/10.1016/j.asoc.2020.106200.

This is a PDF file of an article that has undergone enhancements after acceptance, such as the
addition of a cover page and metadata, and formatting for readability, but it is not yet the definitive
version of record. This version will undergo additional copyediting, typesetting and review before it
is published in its final form, but we are providing this version to give early visibility of the article.
Please note that, during the production process, errors may be discovered which could affect the
content, and all legal disclaimers that apply to the journal pertain.

© 2020 Published by Elsevier B.V.

 Journal Pre-proof

Local Segmentation of Images Using an Improved Fuzzy

C-Means Clustering Algorithm Based on Self-adaptive
Dictionary Learning

of
Jiaqing Miaoa,b,c∗, Xiaobing Zhouc†, Ting-Zhu Huanga‡
a. School of Mathematical Sciences, University of Electronic Science and Technology of China,

pro
Chengdu, Sichuan 611731, PR China
b. The Engineering Technical College of Chengdu University of Technology,
Leshan, Sichuan 614000, PR China
c. Geophysical Engineering Department, Montana Technological University,
Butte, Montana, 59701, USA
re- Abstract
Image segmentation is an active research topic in image processing. The Fuzzy C-means
(FCM) clustering analysis has been widely used in image segmentation. As there is a large
amount of delicate tissues such as blood vessels and nerves in medical images, noise generated
lP
during imaging process can easily affect successful segmentation of these tissues. The tradi-
tional FCM algorithm is not ideal for segmentation of images containing strong noise. In this
study, we proposed an improved FCM algorithm with anti-noise capability. We first discussed
the algorithm of dictionary learning for noise reduction. Then we developed a new image
segmentation algorithm as a combination of the dictionary learning for noise reduction and
the improved fuzzy C-means clustering. Lastly we used the algorithm of the improved FCM
rna

to segment images, during which we removed the non-target areas making use of the grayscale
features of images and extracted accurately the areas of interests. The algorithm was tested
using synthetic Shepp-Logan images and real medical magnetic resonance imaging (MRI) and
computed tomography (CT) images. Compared to the synthetic data and real medical im-
ages segmented by the fuzzy C-means (FCM) clustering algorithm, the Kernel Fuzzy C-mean
(KFCM) clustering algorithm, spectral clustering algorithm, the sparse learning based fuzzy
C-means (SL FCM) clustering algorithm, and the modified spatial KFCM (MSFCM) algo-
rithm, the images segmented by the dictionary learning Fuzzy C-mean clustering (DLFCM)
Jou

algorithm have higher partition coefficient, lower partition entropy, better visual perception,
better clustering accuracy, and clustering purity.
Key words: dictionary learning; fuzzy C-means clustering; algorithm of the noise reduction;
image segmentation; MRI and CT.

1 Introduction
Out of the computer aided diagnosis (CAD) technologies for accurately clinical diagnosis, magnetic
resonance imaging (MRI) and computed tomography (CT) for disease diagnosis are two important
∗
E-mail: mjq 011114117@163.com.
†
Corresponding author. E-mail: xzhou@mtech.edu.
‡
Corresponding author. E-mail: tingzhuhuang@126.com.

1
 Journal Pre-proof

ones. Accurate image segmentation of the medical images is important for the correct diagnosis
of diseases. Segmented medical images are not only used to estimate the area of lesions but also
important basis for a doctor to make informative diagnoses. From the clinical diagnosis point of
view, segmented images are important resources for reconstruction and three-dimensional (3D) vi-
sualization. Due to the complexity of the medical imaging system and operations, noise can be
generated during the formation of medical images and data communication and transfer [1]. There-
fore, it is necessary to suppress artifacts and reduce random noise in medical images before they are

of
segmented. For medical image segmentation, some specific features such as the difference between
any tumor and its background tissues, location of brain blood vessels, vascular overlap, and vascular
thrombosis need also to be considered. Among many algorithms for medical image segmentation,
the fuzzy C-means (FCM) algorithm is one of the most popular [2]. The salient advantages of the

pro
FCM algorithm are that the segmentation process is unsupervised and the algorithm can be applied
to any noisy images. However, the accuracy of traditional FCM algorithms is limited. Research
has been carried out to improve the segmentation accuracy of the FCM algorithms. Chen et al.
(2016) proposed an improved anisotropic hierarchical fuzzy C-means method based on a multivari-
ate student t-distribution [3]. The method is very effective in segmenting brain MRI images. Other
clustering algorithms include the kernelization framework based on kernel PCA [4], the constrained
Laplacian rank algorithm [5], the spectral embedded clustering [6], and the initialization indepen-
re-
dent clustering method [7]. The development in fuzzy theory and dictionary learning in recent
years [8, 9] provides an effective solution for unsupervised segmentation algorithms. A sparse dic-
tionary can extract the main features of an image more effectively than conventional algorithms.
Subsequently, the main information in an image can be enhanced while the non-primary informa-
tion can be suppressed. Dictionary learning has become one of the hotspots in the application of
computer vision [10, 11, 12]. It has been widely applied to pattern recognition and super-resolution
lP
reconstruction of images. The objective of this study is to develop a new algorithm to improve
segmentation accuracy of medical images. The algorithm will integrate the noise reduction based
on adaptive dictionary learning into the improved FCM clustering algorithm to segment complex
medical images accurately and effectively [13, 14]. The proposed algorithm uses a sparse-coding
representation that is based on overcomplete dictionaries. Image segmentation is performed on a
pixel-by-pixel basis. The segmentation process considers not only the membership degree between
rna

pixels and cluster centers, but also the intrinsic characteristics of pixels in each class. The new
algorithm was then applied to segment brain MRI images and abdominal CT images, the results
were compared to those by other algorithms such as the traditional FCM clustering algorithm, the
Kernel Fuzzy C-mean (KFCM) algorithm [15, 16], spectral clustering (SC) [17], the sparse learn-
ing based fuzzy C-means (SL FCM) algorithm [18], and the modified spatial KFCM (MSFCM)
algorithm [19, 20, 21].
Jou

2 Dictionary-based noise reduction

Single-scale non-parametric class in the sparse representation (SR) includes Independent Component
Analysis (ICA), diffusion maps, Principal Component Analysis (PCA), and dictionary learning
[22, 23]. ICA is in general an over-complete method [24]. The diffusion maps algorithm is a non-
linear method that focuses on discovering the underlying manifold [25]. PCA is looking for a new
basis in data, which is a linear combination of the original bases that best re-express the data-set
[26]. Such learning can be based on different algorithms like Method of Optimal Directions (MOD)
[27] or K-SVD (Table 1) [12]. The K-SVD algorithm chosen in this study is accordingly a single-scale
method.
The K-SVD algorithm based on the dictionary-learning technique for noise reduction is a sparse
decomposition algorithm. There are two main types of sparse decomposition algorithms: the match-

2
 Journal Pre-proof

Table 1. Summary of the atomic SR algorithms

Transform name Dimension Introduced Reference
Independent Compo- 1D/2D/3D 2000 A. Hyvarinen [24]
nent Analysis (ICA)
Diffusion maps 1D/2D/3D 2006 R. R. Coifman [25]
Single-scale
Principal Component 1D/2D/3D 2014 J. Shlens [26]

of
Analysis (PCA)
Method of Optimal Di- 1D/2D/3D 1999 K. Engan [27]
rections (MOD)
Conventional KSVD 1D/2D/3D 2006 M Aharon [12]

pro
ing pursuit (MP) algorithms [13] and orthogonal matching pursuit (OMP) algorithms [28]. Dic-
tionary learning for noise reduction involves two steps: calculating the coefficients of the sparse
representation for each image block of an image followed by generating an adaptive dictionary
[8, 28].

2.1 Adaptive dictionary learning based on a noisy image itself

re-
In order to reduce the complexity of the OMP algorithm, we used the steepest descent algorithm
to solve the least square problem during an iterative process. Assuming Z = {yj }M j=1 , where
√ √ 2
M = ( N − n + 1) , N is the number of pixels in an image X, M is the √ number
√ of patches in
the image X, yj is an image block of a noisy image Y , and the size of yj is n × n. The problem
can be redefined as Eq.(1) [11]:
lP
X X
{D, aij , X} = arg min λ kX − Y k22 + µij kaij k0 + kDaij − Rij Xk22 , (1)
D,aij ,X
i,j i,j

where D is a dictionary, aij is the coefficient of sparse representation, X is the corresponding

denoised image of noisy image Y , λ and µij are normalization parameters, i and j are the number
rna

of rows and columns, respectively, and Rij is an n × N matrix for extracting an image block at
position (i, j).
Implementation of the noise reduction algorithm based on the dictionary derived from the image
itself can be described as follows:
(1) If both the dictionary D and the image X are given, aij can be obtained using the OMP
algorithm;
(2) With the coefficients of sparse representation aij obtained, the dictionary D is updated using
Jou

the kernel secular value decomposition (K-SVD) algorithm [12];

(3) When the updated dictionary D and the coefficients of sparse representation aij meet a
preset accuracy, the iterative procedure terminates and the denoised image X is obtained by Eq.(2);
otherwise, the above steps are repeated:

X −1 X
T T
X = (λI + Rij Rij ) (λY + Rij Daij ), (2)
i,j i,j

2.2 A case study using a synthetic medical image

To evaluate the noise-reduction algorithm, we used a synthetic Shepp-Logan image of 512 × 512
pixels in size as a case study with peak signal-to-noise ratio (PSNR) as the objective function. A

3
 Journal Pre-proof

higher P SN R value indicates that more noise is reduced. Results are shown in Fig.1. The first
row in Fig. 1 is three copies of a synthetic Shepp-Logan image. The second row from left to right
is the corresponding images to the first row but with added normally distributed random noise
of zero mean and standard deviation (SD) of 15, 20 and 30, respectively. The resultant P SN R
values are 24.60, 22.12, and 18.59, respectively. The third row is the dictionaries that are numbered
1, 2, · · · , and 6 from left to right. The 1, 3, 5 dictionaries were trained using the redundant discrete
cosine transform (DCT) dictionary algorithm, and the 2, 4, and 6 dictionaries are trained using

of
the adaptive dictionary method described in Section 2.2 and the size of the image blocks is 5 × 5.
The fourth row shows the noise-reduced images using the corresponding dictionary in the third row
of the same column; the corresponding P SN R values are 34.60, 42.29, 33.67, 39.84, 28.72, and
30.22 for images from left to right, respectively. For the image of SD = 15, the P SN R value using

pro
the redundant DCT dictionary is 34.60, lower than the P SN R value of 42.29 using the adaptive
dictionary.

re-
lP
rna

Fig. 1. Noise reduction procedure and results using a synthetic image, images with added noise of different
Jou

levels and dictionaries generated by the redundant DCT dictionary algorithm and the adaptive dictionary
training.

3 Images segmentation based on improved fuzzy C-means

clustering
Fuzzy C-means (FCM) clustering algorithm are widely used in image segmentation that includes
FCM, KFCM, MSFCM, PFCM, GFCM, and HFCM, and SL FCM etc. Bezdek et al. (1984) pro-
posed that the clustering criterion used to aggregate subsets is a generalized least-squares objective
function [2]. Zhang et al. (2003) proposed a clustering algorithm using the ’kernel method’ based

4
 Journal Pre-proof

on the classical fuzzy clustering algorithm (FCM), and called it the kernel fuzzy C-means algorithm
(KFCM) [16]. Zanaty (2012) proposed the kernelized fuzzy C-means algorithm with modified spatial
constraints (MSFCM) [19]. Pal et al. (2005) proposed a new model called possibilistic-fuzzy C-
means (PFCM) model. PFCM produces memberships and possibilities simultaneously, along with
the usual point prototypes or cluster centers for each cluster [29]. Zheng et al. (2015) proposed
two algorithms, generalized FCM (GFCM) and hierarchical FCM (HFCM) [30].Gu et al. (2017)
propose a new clustering algorithm, called sparse learning based fuzzy C-means (SL FCM). By this

of
way, some redundant information (i.e. the correlation among samples of different classes) in the
discriminant feature can be also removed, which can improve the clustering quality [18].

Table 2. Summary of the FCM algorithms

pro
Algorithm name Introduced Reference
Fuzzy clustering algorithm (FCM) 1984 J.C. Bezdek [2]
Kernel fuzzy C-means algorithm 2003 D.Q. Zhang [16]
(KFCM)
FCM
Fuzzy C-means with modified spatial 2012 E.A. Zanaty [19]
(MSFCM)
Possibilistic-fuzzy C-means (PFCM)
re- 2005 N. R. Pal [29]
Generalized FCM (GFCM) and hierar- 2015 Y. Zheng [30]
chical FCM (HFCM)
Sparse learning based fuzzy C-means 2017 J. Gu [18]
(SL FCM) algorithm
lP
The image segmentation algorithm based on the FCM clustering analysis is an algorithm of
fuzzy optimization [2, 16]. In this study, the objective function and spatial information were both
modified and they improve the traditional FCM clustering algorithms in image segmentation.

3.1 The traditional FCM algorithm

rna

Bezdek et al. (1984) generalized the clustering algorithm of objective function to fuzzy clustering
according to the idea of partition of fuzzy sets [2]. In order to improve the algorithm, the following
objective function is used [2, 9, 15]:
c X
X N
J2 (u, v) = (uik )2 (dik )2 . (3)
i=1 k=1
Jou

where c is the number of clusters, dik is the distance, N is the number of pixels of any gray
image,
Pc uik is the membership degree of the k-th pixel in the i-th cluster that satisfies the condition
i=1 ik = 1, ∀uik ∈ [0, 1].
u
The widely used traditional FCM algorithm is a clustering algorithm [31, 32]. The objective
function is minimized through iteration in order to achieve optimal image segmentation. The
objective function of the traditional FCM algorithm is defined as:
c X
X N
2
Jm = um
ik kxk − vi k , (4)
i=1 k=1

where {xk , k = 1, 2, · · · , N } is a data set of pixels of any gray image, {vi , i = 1, 2, ..., c} is the set of
the cluster centers, m is the index of the fuzzy weight [33]. The new values of membership degree

5
 Journal Pre-proof

uik and the new cluster centers vi are calculated using the following equation [34]:

P
N
2
um
ik xk
−1 /
(m−1)
kxk − vi k k=1
uik = P , vi = . (5)
2

c −1 /
(m−1) PN
kxk − vj k um
ik
j=1 k=1

of
However, the algorithm is essentially a local-search algorithm based on gradient descent and
thus has a greater dependence on the initial conditions.

3.2 Kernel fuzzy C-means (KFCM) algorithm

pro
Methods that are based on kernel functions have been successfully applied to pattern recognition,
function optimization, etc. [15, 16, 35, 36]. The feature space of a kernel function is expressed as
follows [37]:

K (x, y) = hΦ (x) , Φ (y)i , (6)

where h, i indicates the inner product. Any function that meets the Mercer theorem can be selected
re-
as a kernel function [38]. In this study, the Gaussian kernel function was used [39]:


K (x, y) = exp −kx − yk2 /σ 2 (7)

where σ is the characteristic parameter of the kernel function. The objective function for the KFCM
segmentation algorithm is defined as follows [32, 16]:
lP
c X
X N
2
Jm = um
ij kΦ (xk ) − Φ (vi )k , (8)
i=1 j=1

where Φ is a non-linear mapping function from a low-dimensional to high-dimensional feature space

[40]. kΦ (xk ) − Φ (vi )k2 in Eq.(8) is given by
rna

kΦ (xk ) − Φ (vi )k2 = K (xk , xk ) + K (vi , vi ) − 2K (xk , vi ) . (9)

According to Eq.(9), the objective function Eq.(8) can be expressed as follows

c X
X N
Jm = 2 um
ik (1 − K (xk , vi )) . (10)
Jou

i=1 k=1

The membership-degree value uik and the cluster centers vi can be obtained as follows

P
N
−1 / um
ik K (xk , vi ) xk
(1 − K (xk , vi )) (m−1) k=1
uik = P
c , vi = . (11)
−1 P
N
(1 − K (xk , vj )) /(m−1) um
ik K (xk , vi )
j=1 k=1

The robustness of the KFCM algorithm for medical image segmentation is better than the
traditional FCM algorithm [16].

6
 Journal Pre-proof

3.3 A new image segmentation algorithm of the fuzzy C-means based

on the kernel function and information of the spatial neighborhood
(1) Spatial feature of the neighboring pixels
The KFCM algorithm based on a kernel function has some shortcomings such as high sensitivity
to noise [41, 42].To overcome this feature in image segmentation, the neighborhood information of
a pixel is introduced to the KFCM clustering algorithm using a new space function that is defined

of
as follows [43]:
P
uit αt
t∈Nk
hik = P c P , (12)

pro
ujt αt
j=1 t∈Nk

√ √
where Nk represents a block of n × n in size (in this study, n = 25). When the block is
represented by a vector, it centers at the pixel xk that is assumed to belong to the i-th cluster.
If all of the neighboring pixels of the point xk in Nk do not belong to the i-th cluster, then the
minimum value of the spatial functions is obtained. uit is the membership degree of the pixel xt ’s
neighborhood and the pixel xt belongs to the i-th cluster. The smaller the difference of gray values
re-
between xk and xt is, the closer the relationship between xk and xt will be and the larger the value
of αt in Eq.(13) will be [14, 44]. The specific definition of αt is as follows:


αt = 1/exp β|xk − xt |2 (13)

where β determines the degree of decrease. After some experiments, we found that β = 0.9 is
lP
desirable.
(2) Modified spatial KFCM (MSFCM) algorithm
The fuzzy membership degree and cluster centers can be defined as follows using the space
function and kernel functions [19, 20]:

P
N
0
m
uik K (xk , vi ) xk
rna

0 upik hqik 0 k=1

uik = Pc , vi = . (14)
P
N
m
upik hqjk uik
0
K (xk , vi )
j=1 k=1

The relative importance between the original membership degree and function space is controlled
by the parameters p, q [21].
Jou

3.4 The image segmentation algorithm based on dictionary learning

fuzzy C-mean clustering (DLFCM)
The FCM image segmentation algorithm is a typical unsupervised algorithm, which does not re-
quire human intervention. Implementation of the algorithm is through calculating the minimum
Euclidean distance between the target pixels and the clustering center [45]
n
X
ui = arg min kxi − V ui k22 + λkU k0 . (15)
V, ui
i=1

where V = [v1 , v2 , · · · , vc ] is the matrix of clustering centers that can be obtained by the dictionary
learning. U = [u1 , u2 , · · · , un ] is the matrix of membership degrees, where ui is the membership
degree of the sample xi , that is, ui can be determined by the sample xi . Because Eq.(15) is a typical

7
 Journal Pre-proof

nondeterministic polynomial (NP)-hard problem, sparse representation can be used to improve the
objective function of the clustering algorithm [46, 47].
n
X
ui = arg min kxi − V ui k22 + λkU k1 , (16)
V, ui
i=1

where k·k1 is the l1 norm. Since Eq.(16) reflects the similarity only between the internal structures

of
of the image, that is, it only makes use of the sparse coefficient and the correlation of clustering
centers, it does not fully reflect the Euclidean distance between each pixel and the cluster center in
the FCM algorithm. In this study we proposed a new fuzzy clustering algorithm that is based on
dictionary learning for a non-convex optimization problem. A new regularization term was added

pro
to the FCM algorithm to build a non-convex model. The newly developed algorithm is as effective
as other methods of non-convex matrix rank approximations. The logDet non-convex function was
chosen because the convergence is relatively stable in the process of minimization of the function.
Therefore, Eq.(16) can be optimized through transforming it to the following form [18]
n
X c X
X n
ui = arg min kxi − V ui k22 + λkU k1 +γ m
wik kxi − vk k22 , (17)
V, ui , W
i=1 k=1 i=1

where wik is a weight coefficient,

P
c

k=1
re-
wik = 1, wik ∈ [0, 1]. It is well known that using the l1 -norm
to approximate the l0 -norm is not perfect. In this study we considered another regularization term
to replace the non-convex l0 -norm. Reference [48] has shown the good performance of log det as
a non-convex substitution model. We will show that the log det model is a better substitution for
lP
the l0 -norm than the l1 -norm in solving the minimization problem. Firstly, l1 -norm assigns the
same weight to all singular values of each image. In fact, for many medical images, the singular
values have clear physical meanings with different importance and should be treated differently. For
instance, larger singular values convey low-frequency
 information. Secondly, the log det () function,
P1/2
defined as Eq.(18): L (U, ε) = log det + εI , is a better approximation to the l0 -norm. Thus,
the sum of the log det () function of singular values is a better surrogate than l1 -norm. In addition,
rna

the log det () function has the advantage of automatically assigning a weight to each singular value.
The matrix U ∈ Cc×n , c ≤ n is neither a square nor a semi-positive matrix in general. According
to reference [48], we have:

1/2 


.
L (U, ε) = log det U U T + εI = log det QΣ1/2 Q−1 + εI = log det Σ1/2 + εI , (18)

where Σ is a diagonal matrix. The diagonal elements of Σ are eigenvalues of the matrix U U T .
Jou

Therefore the matrix can be decomposed orthogonally to U U T = QΣ1/2 Q−1 . Because Σ1/2 is a
diagonal matrix, its diagonal elements are the singular values of the matrix U . Therefore, the log det
model can be used to replace the U ’s rank minimization model. Making use of Eq.(18), we can
obtain a low-rank approximation model as shown in Eq.(19):
n
X c X
X n
ui = arg min kxi − V ui k22 + λL (U, ε) +γ m
wik kxi − vk k22 . (19)
V, U, L
i=1 k=1 i=1

Using Alternating Direction Method of Multipliers (ADMM) [49, 50] we were able to develop
a better method to solve Eq.(19). The advantage of the ADMM algorithm is that we can divide
Eq.(19) into two sub-problems and solve them separately. The two sub-problems can be expressed

8
 Journal Pre-proof

as follows:
n
X c X
X n
V = arg min kxi − V ui k22 +γ m
wik kxi − vk k22 (a)
V
i=1 k=1 i=1
n (20)
X
ui = arg min kxi − V ui k22 + λL (U, ε) (b)
U
i=1

of
(1) Solving the V sub-problem
The dictionary V is solved and updated using Eq.(20a), and the specific form for m = 2 is given
by
n
X n
X

pro
V = arg min kxi − V ui k22 +γ kXi wi − V wi k22 , (21)
V
i=1 i=1

where X = [x1 , x2 , · · · , xn ], wi = [wi1 , wi2 , · · · , wic ]T , Xi = [xi , xi , · · · , xi ], and wik have the following
form:
−1
kxk − vi k2
wik = P c
−1 . (22)
kxk − vj k2

X n n
X
!−1
Xn
re- j=1

Using the gradient descent method to solve Eq.(20a), we have:

n
X
!
T T T T
V = ui ui + r wi wi xi ui + rXi wi wi . (23)
i=1 i=1 i=1 i=1

(2) Solving the U sub-problems

lP
In order to solve Eq.(20b) for U we need to convert it to the following minimization problem:
U = arg min kX − V U k2F + λL (U, ε) . (24)
U

Since L (U, ε) is the logarithm of the product of singular values, Eq.(24) can be rewritten as:
rna

n
λX
min kX − V U k2F + log (σj (U ) + ε) , (25)
U η j=1

where σj (U ) is the j-th singular value of U . For simplicity, we will use σj instead of σj (U ) to
Pn
represent the j-th singular value of U . Although log (σj + ε) is nonconvex, using the local
j=1
minimization method we can effectively solve the nonconvex problem. According to reference [51,
Jou

52], it is easy to get

n
(k+1) 2 λ X σj
U = arg min X − V U (k) F + . (26)
U η j=1 σj(k) + ε
n  
P   
(k) (k)
Let 1/ σj + ε σj be denoted by ϕ σj , σj , Eq.(26) can be reduced to
j=1

2
 
(k)
U (k+1) = arg min X − V U (k) F
+ τ ϕ σj , σj , (27)
U
 
(k)
where τ = λ/η. Since the weight 1/ σj + ε of σj is increasing, Eq.(27) is not a convex optimiza-
tion problem. So the global minimum point cannot be found. However, the local minimum point
can be found [48]. A stepwise procedure of the algorithm is summarized in Algorithm 1.

9
 Journal Pre-proof

Algorithm 1 Image segmentation algorithm based on dictionary learning fuzzy C-means clustering
(DLFCM) method
1. Initialization:
(1) firstly, the initial cluster centers V = [v1 , v2 , · · · , vc ] are obtained by dictionary learning;
(2) set the initial parameters λ, η, and τ = λη ;
(3) initialize the weights wi = [wi1 , wi2 , · · · , wic ]T with given wik by Eq.(22).
2. Iteration:

of
(1) calculate U (k) by solving Eq.(27);
(2) calculate V (k) by solving Eq.(23);
(3) update weights wi = [wi1 , wi2 , · · · , wic ]T by solving Eq.(22);
(4) k = k + 1;

pro
(5) going back to step (1) until stop condition is met.

3.5 Evaluation metrics for the effectiveness of algorithm

We used two methods to evaluate the effectiveness of the new algorithm through analyzing the
image segmentation results. One method uses the partition coefficient and partition entropy; the
other one uses the Xie-Beni function. For further evaluation, we also used F1-measure (F1) [53],
re-
clustering accuracy (CA) [54], and clustering purity (CP) [55]. Partition coefficient and partition
entropy are defined, respectively, as follows:
c N
1 XX 2
vpc = u , (28)
n i=1 k=1 ik
lP
c N
−1 X X
vpe = (uik lg uik ). (29)
n i=1 k=1

Clustering can achieve the best results when the partition coefficient reaches a maximum and
rna

the partition entropy reaches a minimum. Another evaluation model is the Xie-Beni function that
should reach a minimum if the segmentation algorithm performs the best [56]
c X
X N   
2 2

vxb = u2ik kxk − vi k / N min kvj − vi k . (30)
k6=j
i=1 k=1

An indicator of a good segmentation is that the distribution of the pixels within each cluster is
compact and the fuzzy correlation between clusters is as small as possible. That is, vxb should be
Jou

minimized.

4 Results and analysis of the experiments

4.1 A Case study with a synthetic Shepp-Logan image
We used a synthetic noise-free Shepp-Logan image and images of different levels of noise to inves-
tigate the sensitivity of the new algorithm to noise level [57]. The noisy images were generated by
adding normally distributed random noise of zero mean and standard deviations of 15, 20 and 30,
respectively, to the original noise-free Shepp-Logan image. We compared the segmentation results
by the FMC algorithm, the KFCM algorithm, the spectral clustering algorithm, the SL FCM algo-
rithm, the MSFCM clustering algorithm, and the new DLFCM algorithm developed in this study.

10
 Journal Pre-proof

Parameters used for the segmentation using all the algorithms were specified as follows: the pixel
neighborhood was a 5 × 5 block, the number of clusters was 4, the iteration termination criterion
was 0.00001, which is the absolute difference of cluster center between two consecutive iterations,
and the maximum number of iterations was 300, the fuzzy weighting index m = 2, p = 1, q = 1,
and the width of the Gaussian kernel function was σ = 0.5.
Fig. 2 shows the fuzzy clustering results of a synthetic Shepp-Logan image with no noise by
different algorithms. Column (a) is five copies of the image. Columns (b) to (e) are the four clusters

of
segmented by five different algorithms. From the top to the bottom, the first to the bottom rows
show the segmented results using the traditional FCM algorithm, the KFCM algorithm, the spectral
clustering algorithm, the SL FCM algorithm, the MSFCM algorithm, and the newly developed
DLFCM algorithm, respectively.

pro
Fig. 3 shows the fuzzy clustering results of a synthetic Shepp-Logan image with added noise of
standard deviation of 15. Column (a) is five copies of the noisy image. From the top to the bottom,
the first to the fifth rows show the segmented results using the traditional FCM algorithm, the
KFCM algorithm, the spectral clustering algorithm, the SL FCM algorithm, the MSFCM algorithm,
and the newly developed DLFCM algorithm, respectively.
Fig. 4 shows the fuzzy clustering results of a synthetic Shepp-Logan image with added noise of
standard deviation of 20. Column (a) is five copies of the noisy images. From the top to the bottom,
re-
the first to the bottom rows show the segmented results using the traditional FCM algorithm, the
KFCM algorithm, the spectral clustering algorithm, the SL FCM algorithm, the MSFCM algorithm,
and the newly developed DLFCM algorithm, respectively.
Fig. 5 shows the fuzzy clustering results of a synthetic Shepp-Logan image with added noise of
zero mean and standard deviation of 30. Column (a) is five copies of the noisy image. From the top
to the bottom, the first row to the bottom rows show the segmented results using the traditional
lP
FCM algorithm, the KFCM algorithm, the spectral clustering algorithm, the SL FCM algorithm,
the MSFCM algorithm, and the newly developed DLFCM algorithm, respectively.
The partition coefficient, partition entropy, and the Xie-Beni function were calculated for the
cases shown in Figs. 2-5. The results are shown in Table 3. A detailed analysis of the numerical
results shown in Table 3 shows that (1) for any of the algorithms, as the noise level increases, the
partition coefficient vpc decreases, the partition entropy vpe and the Xie-Beni function vxb increases;
rna

(2) for images of various levels of noise, the DLFCM algorithm resulted in the highest vpc and lowest
vpe and vxb in comparing to the other algorithms. These results show that the DLFCM algorithm
has the best performance in noisy image segmentation among the algorithms studied, even for the
cases that the noise level is high with standard deviation of 30. For images with no noise, the
SL FCM algorithm performs similarly well as or a slightly better than the DLFCM algorithm (see
Table 3). However, because of a lack of efficient noise suppression capacity, the algorithm did not
perform as well as the DLFCM algorithm in the case of noisy images.
Jou

To further evaluate the newly developed DLFCM algorithm, we calculated the F1-measure
[53], clustering accuracy [54], and clustering purity [55] for the cases shown in Figs. 2-5. The
results are shown in Table 4. A detailed analysis of the numerical results shown in Table 4 show
that (1) the sparse representation technique has favorable classification ability, and it has been
introduced into the fuzzy clustering component in the SL FCM algorithm. The weighted distance
method in SL FCM can enhance similarity of samples of the same class and difference of samples
of different classes, thus improve the clustering accuracy. Hence, the SL FCM algorithm performs
as well as or slightly better than the DLFCM algorithm in segmentation of noise-free images with
similar structures; (2) for any of the algorithms, as the noise level increases, the values of the
F1, CA, and CP decrease; (3) for all images of various levels of noise, the DLFCM algorithm
resulted in the highest values of F1, CA, and CP in comparing to the other algorithms. For the
noise-free image, the segmentation results of the spectral clustering algorithm are the best, but the
segmented results by the DLFCM algorithm are very similar. The spectral clustering is based on

11
 Journal Pre-proof

of
pro
re-
lP
rna
Jou

Fig. 2. Segmented results of a synthetic Shepp-Logan image with no noise. The number of clusters was
set to 4. Column (a) (from left to right): copies of the original synthetic Shepp-Logan image. Columns
(b) to (e) are the four clusters segmented. From top down, Row 1 to Row 6 are the segmented results
using the traditional FCM algorithm, the KFCM algorithm, the spectral clustering algorithm, the SL FCM
algorithm, the MSFCM algorithm, and the new DLFCM algorithm, respectively.

12
 Journal Pre-proof

of
pro
re-
lP
rna
Jou

Fig. 3. Same as Fig.2, segmentation results of the synthetic Shepp-Logan image that is contaminated by
random noise of zero mean and standard deviation of 15.

13
 Journal Pre-proof

of
pro
re-
lP
rna
Jou

Fig. 4. Same as Fig.2, segmented results of the synthetic Shepp-Logan image with added random noise of
zero mean and standard deviation of 20.

14
 Journal Pre-proof

of
pro
re-
lP
rna
Jou

Fig. 5. Same as Fig.2, segmented results of the synthetic Shepp-Logan image with added random noise of
zero mean and standard deviation of 30.

15
 Journal Pre-proof

Table 3. Segmentation performance of four algorithms on synthetic Shepp-Logan Images with

different levels of noise
Image Algorithm vpc vpe vxb
FCM 0.865 0.177 0.061
The original image of KFCM 0.831 0.155 0.056
Shepp-Logan with no SL FCM 0.999 0.000 0.000
noise MSFCM 0.934 0.082 0.032

of
DLFCM 0.998 0.001 0.009
FCM 0.833 0.192 0.084
Shepp-Logan image KFCM 0.812 0.186 0.074
with added noise of SL FCM 0.866 0.113 0.061

pro
SD = 15 MSFCM 0.892 0.094 0.054
DLFCM 0.958 0.005 0.013
FCM 0.786 0.212 0.095
Shepp-Logan image KFCM 0.803 0.205 0.091
with added noise of SL FCM 0.827 0.130 0.083
SD = 20 MSFCM 0.852 0.121 0.079
DLFCM 0.923 0.011 0.015

Shepp-Logan image
with added noise of
re-
FCM
KFCM
SL FCM
0.703
0.751
0.770
0.294
0.265
0.214
0.121
0.107
0.099
SD = 30 MSFCM 0.803 0.162 0.094
DLFCM 0.906 0.048 0.068
lP
the spectra method in the graph theory that works very well for the noise-free image segmentation.
However, the segmented results by the spectral clustering algorithm deteriorate quickly as the noise
increases, while those by the DLFCM algorithm deteriorate much more slowly than those by the
spectral clustering algorithm. Since medical images generally contain more or less noise during the
imaging process, the newly developed algorithm has the best noise resistance among the algorithms
rna

evaluated.
Higher partition coefficient, lower partition entropy, lower values of the Xie-Beni function, higher
F1-measure, clustering accuracy, and clustering purity of the new algorithm show that it can segment
objects from noisy images more accurately than the other algorithms tested (FCM, KFCM, SC,
and MSFCM).

4.2 A case study with a brain MRI image

Jou

We used a real brain MRI image for further algorithm evaluation. The brain MRI medical image was
produced by a 0.35T permanent-magnet MRI spectrometer and was provided by the Siemens. The
new algorithm was tested against a traditional FCM clustering algorithm, the Kernel Fuzzy C-mean
(KFCM) algorithm, spectral clustering (SC), the sparse learning based fuzzy C-means (SL FCM)
clustering algorithm, and the modified spatial KFCM (MSFCM) algorithm. Parameters used for
segmentation were specified as follows: the pixel neighborhood was a 5 × 5 block, the number of
clusters was set to 3, the iteration accuracy was 0.00001, which is the absolute difference of cluster
centers between two consecutive iterations, the maximum number of iterations was 300, the fuzzy
weighting index m = 2, p = 1, q = 1, and the width of the Gaussian kernel function was σ = 0.5.
Fig. 6 shows the fuzzy clustering results of a brain MRI image. Column (a) shows five copies of
the image, column (b) the cluster of cerebral white matter, column (c) the cluster of encephalocoele,
and column (d) the cluster of gyri tissue and cerebral gray matter. From the top to the bottom, the

16
 Journal Pre-proof

of
pro
Table 4. F1-measure (F1) [53], clustering accuracy (CA) [54], and clustering purity (CP) [55] for
five algorithms on synthetic Shepp-Logan images with different levels of noise.
Image Algorithm F1- Clustering Clustering
measure accuracy purity
FCM 0.994 0.987 0.988
KFCM 0.996 0.989 0.991
Shepp-Logan image Spectral 1.000 1.000 1.000
with no noise re-
SL FCM 0.999 0.999 0.999
MSFCM 0.997 0.993 0.994
DLFCM 0.999 0.999 0.998
FCM 0.761 0.842 0.846
Shepp-Logan image KFCM 0.889 0.941 0.924
with added noise of Spectral 0.809 0.889 0.880
lP
SD = 15 SL FCM 0.893 0.927 0.939
MSFCM 0.922 0.955 0.961
DLFCM 0.996 0.998 0.998
FCM 0.721 0.803 0.796
Shepp-Logan image KFCM 0.779 0.875 0.867
with added noise of Spectral 0.771 0.856 0.848
rna

SD = 20 SL FCM 0.816 0.874 0.861

MSFCM 0.843 0.896 0.895
DLFCM 0.988 0.984 0.978
FCM 0.640 0.725 0.736
Shepp-Logan image KFCM 0.679 0.760 0.775
with added noise of Spectral 0.530 0.693 0.705
SD = 30 SL FCM 0.559 0.762 0.795
Jou

MSFCM 0.698 0.846 0.868

DLFCM 0.983 0.936 0.937

17
 Journal Pre-proof

of
pro
re-
lP
rna
Jou

Fig. 6. Segmented result of clustering of a brain MRI medical image. The number of clusters was set to
3. Column (a) (from left to right) is five copies of a brain MRI medical image. Columns (b) to (d) are
the three clusters segmented. From top down, the first to the bottom rows show the segmented results
using the traditional FCM algorithm, the KFCM algorithm, the spectral clustering algorithm, the SL FCM
algorithm, the MSFCM algorithm, and the new DLFCM algorithm, respectively.

18
 Journal Pre-proof

first to the last rows show the segmented results using the traditional FCM algorithm, the KFCM
algorithm, the spectral clustering algorithm, the SL FCM algorithm, the MSFCM algorithm, and
the newly developed DLFCM algorithm, respectively. Comparing the segmented results with the
original image for each algorithm and the segmented results among the algorithms, we can see from
Fig. 6 that the detailed anatomical information such as the edge of the brain, blood vessels of the
brain, nerve system in the segmented images by the MSFCM algorithm are better preserved than
the other algorithms.

of
4.3 A Case study with an abdominal CT image
As another case study, we used an actual abdominal CT image for algorithm evaluation. The

pro
abdominal CT image was produced by a 16-slice Emotion CT spectrometer and was provided by the
Siemens. Parameters used for image segmentation were specified as follows: the pixel neighborhood
was a 5 × 5 block, the number of clusters was 4, the iteration accuracy was 0.00001, the maximum
number of iterations was 300, the fuzzy weighting index m = 2, p = 1, q = 1, and the width of the
Gaussian kernel function was σ = 0.5.
Fig. 7 shows the fuzzy clustering results of an abdominal CT image. Column (a) shows five
copies of the image, column (b) the cluster of skin and subcutaneous adipose tissue, column (c)
the cluster of bone tissue, column (d) the cluster of abdominal visceral tissue, and column (e) the
re-
cluster of tissue of abdominal cavity. From the top to the bottom, the first to the last rows show
the segmented results using the traditional FCM algorithm, the KFCM algorithm, the spectral
clustering algorithm, the SL FCM algorithm, the MSFCM algorithm, and the newly developed
DLFCM algorithm, respectively. Comparing the segmented results with the original image for
each algorithm and the segmented results among the algorithms, we can see from Fig. 7 that the
lP
anatomical information details such as the edge of the abdomen, blood vessels of the abdomen, etc.
in the segmented images by the DLFCM algorithm are better preserved than the other algorithms.
For real MRI and CT images, the actual noise and the standard segmented images as a reference
are not known. Thus, accuracy evaluation of algorithms using real MRI and CT images is not
possible since no standard segmented images are available. Therefore, we can only use the partition
coefficient, the partition entropy, and the Xie-Beni function as the matrices for evaluation and
rna

inter-comparison among the algorithms when real medical images are used. The results of these
matrices for Figs. 6 and 7 are shown in Table 5. A detailed analysis of the numerical results shown
in Table 5 shows that for either the real MRI or CT image, the DLFCM algorithm resulted in the
highest vpc and lowest vpe and vxb in comparing to the other algorithms, demonstrating that the
new algorithm has strong promotional value in practical application.
From the results shown in Figs. 6 and 7 and Table 5, we can see that the DLFCM algorithm
outperformed the other algorithms in preserving detailed structures of objects in actual medical
Jou

image segmentation. The better performance of the DLFCM algorithm than the other algorithms
for the brain MRI and abdominal CT medical images is consistent with the case studies in Section
4.1 using the synthetic images with or without noise. The new algorithm has the advantages of
high computation speed and improved accuracy in segmenting selected objects.

4.4 A comparison in central processing unit (CPU) time

The computational cost of the newly developed DLFCM algorithm for processing a specific image
mainly includes the costs in solving the V and U sub-problems. In each iteration, the main com-
putational cost in solving the V sub-problem is in calculating matrix product and inverse matrix,
and the computational complexity is O (2c3 ), where c is the size of the matrix. For solving the U
sub-problem, the computational complexity (sparse and transformation updates) of the proposed
algorithm is O (nc2 ), where c and n are the number of rows and columns of U , respectively.

19
 Journal Pre-proof

of
pro
re-
lP
rna
Jou

Fig. 7. Segmented results of an abdominal CT image by various algorithms. The number of clusters
was set to 4. Column (a) (from left to right) is five copies of the abdominal CT image. Columns (b) to
(e) are the four clusters segmented. From top down, the first to the last rows show the segmented results
using the traditional FCM algorithm, the KFCM algorithm, the spectral clustering algorithm, the SL FCM
algorithm, the MSFCM algorithm, and the new DLFCM algorithm, respectively.

20
 Journal Pre-proof

Table 5. Segmentation performance of five algorithms on real MRI and CT images

Image Algorithm vpc vpe vxb
FCM 0.594 0.421 0.437
KFCM 0.625 0.402 0.416
MRI brain image Spectral 0.744 0.294 0.310
SL FCM 0.752 0.226 0.261
MSFCM 0.782 0.255 0.269

of
DLFCM 0.825 0.191 0.130
FCM 0.657 0.361 0.379
KFCM 0.673 0.356 0.368
Abdominal CT image Spectral 0.767 0.251 0.268

pro
SL FCM 0.813 0.199 0.234
MSFCM 0.854 0.176 0.188
DLFCM 0.912 0.102 0.113

A Dell server with Intel(R) Xeon(R) CPU E5-2620 2*2.00GHz (2 processors) and 32GB RAM
was used for this study. The FCM, KFCM, spectral, SL FCM, and MSFCM algorithms were com-
re-
pared with the newly developed DLFCM algorithm in terms of the CPU time in the segmentation
of the five images (Figs. 3-7). Table 6 shows the CPU time that is the average of 10 runs in
segmentation of each image using each of the algorithms.

Table 6. CPU time (units: second) comparison among the FCM, KFCM, spectral, SL FCM, MS-
FCM, and DLFCM algorithms
lP
Images Algorithm FCM KFCM Spectral SL FCM MSFCM DLFCM
Shepp-Logan image 5.085 10.167 0.475 5.216 3.134 2.484
with noise of SD = 15
Shepp-Logan image 5.832 9.662 0.590 5.352 3.391 2.625
with noise of SD = 20
rna

Shepp-Logan image 6.126 10.074 0.624 6.025 5.442 3.435

with noise of SD = 30
MRI brain image 10.742 13.276 0.587 9.339 13.054 9.008
Abdominal CT image 8.612 13.041 0.558 7.138 10.605 7.977

The CPU times shown in Table 6 indicate that for the five images tested, the newly developed
DLFCM algorithm costed similar or less CPU time than the other three algorithms. However, for
the spectral clustering algorithm, the CPU time is much less than the DLFCM algorithm, but the
Jou

accuracy of the DLFCM algorithm is much higher than the spectral clustering algorithm.

4.5 Segmentation of brain MRI images with 3D visualization

We investigated the repeatability of the new algorithm in segmenting a time series of images and
algorithm stability. We segmented 94 brain MRI images that were produced by a 0.35T permanent-
magnet MRI spectrometer and were provided by the Siemens. The segmented images were then
used for a three-dimensional display or 3D visualization [58].In Fig. 8, the parameter settings are
the same as Fig. 6: the pixel neighborhood was a 5 × 5 block, the number of clusters was set to
3, the iteration accuracy was 0.00001, which is the absolute difference of cluster centers between
two consecutive iterations, the maximum number of iterations was 300, the fuzzy weighting index
m = 2, p = 1, q = 1, and the width of the Gaussian kernel function was σ = 0.5. Fig. 8 shows the

21
 Journal Pre-proof

results, in which the top row is the 3D display of the original MRI brain images. The second row
shows the skull stripping results of the DLFCM clustering algorithm. The bottom row is the 3D
display of the segmented brain MRI images in which brain tissues were accurately segmented and
visible. The cerebral ventricle tissues were obtained when skull stripping of the MRI brain images
was performed using the DLFCM clustering algorithm. As the gray values of the cerebral ventricle
tissues are relatively low, image enhancement was performed to clearly display the 3D image of
the cerebral ventricle. The accurately segmented brain MRI medical images have the advantage

of
that the brain tissue can be helpful not only for the doctor’s accurate diagnosis but also in laying
out the foundation for developing a procedure of intelligent operations. Fig. 8 shows that the new
algorithm can segment successfully the time series of MRI images and it was very stable during
the whole procedure. At the same time, the skull stripping results in the 3D visualization gave us

pro
confidence that the segmented images from the algorithm can form a holistic image of the whole
brain, which is beneficial to diagnosis since it gives a doctor a whole picture of the brain, reducing
the confusion caused by different scales in individual segmented images.

re-
lP
rna
Jou

Fig. 8. Segmentation of 94 brain MRI images using the DLFCM algorithm. The top row is the 3D display
of the original MRI medical images. The second row shows the skull stripping results of the DLFCM
clustering algorithm. The bottom row shows the cerebral ventricle segmented using the DLFCM clustering
algorithm, which is the best representation of the segmented brain MRI images.

From the results of the above case studies, we can see that the KFCM clustering segmenta-
tion algorithm is better than the traditional clustering algorithms in terms of higher vpc and lower
vpe and vxb , higher F1-measure, clustering accuracy, and clustering purity comparing to the other
algorithms when noisy images were segmented. While the KFCM algorithm can preserve more
details than the traditional FCM clustering algorithms, the edge of the segmented objects by the

22
 Journal Pre-proof

KFCM clustering algorithm is relatively poor, that is, the segmented objects have jagged edges.
The spectral clustering algorithm is not ideal for segmentation of noisy images. For the SL FCM
algorithm, segmentation of noisy images is ideal. Since the MSFCM algorithm introduces spatial
information during segmentation, the results with both the synthetic Shepp-Logan images of var-
ious levels of noise and brain MRI and abdominal CT medical images indicate that the MSFCM
algorithm performs better than the other clustering algorithms. For noisy images, results of the
DLFCM algorithm are the best among the algorithms studied. The DLFCM algorithm can reduce

of
running time. But for medical images with intensity inhomogeneity, the DLFCM algorithm is not
ideal.

5 Discussion and conclusions

pro
The Gaussian noise or amplifier noise such as sensor noise and electronic circuit noise caused by
low light, high temperature, transmission can be added to a MR image during its acquisition [59].
Generally, in the low intensity regions of an MRI magnitude image, the noise distribution tends
to be the Rayleigh distribution, while in the regions of high intensity, the noise tends to be the
Gaussian distribution [60, 57]. This study focused on an FCM segmentation algorithm that applies
to MRI images contaminated with Gaussian noise.
re-
In this study, we first described a noise-reduction algorithm based on the adaptive redundant
dictionary learning. Then we developed an image segmentation algorithm based on the FCM
clustering and its improved algorithm. This newly developed algorithm is referred to as the DLFCM
algorithm. The traditional FCM algorithm does not consider the relationship between adjacent
pixels and thus large errors are often generated when it is applied to segmentation of medical
images. To overcome the shortcomings of the traditional FCM clustering algorithm, we improved
lP
it in two aspects: improved membership degree and amendments of distance metrics used in the
objective function. The newly developed DLFCM algorithm was evaluated against the traditional
FCM algorithm, the KFCM algorithm, the spectral clustering algorithm, the SL FCM algorithm,
and the MSFCM algorithm in segmenting synthetic Shepp-Logan images of various levels of added
noise, actual brain MRI images and an abdominal CT image.
rna

The main advantages of the new algorithm compared to the other algorithms are that it has
better accuracy in segmentation of images of either noise or noise-free and less running time (Table
3). But the proposed new algorithm also has disadvantages, that is, the algorithm cannot produce
ideal segmentation results for medical image with intensity inhomogeneity. This may be because
the new algorithm does not consider the characteristics of intensity deviation of medical images
in segmentation, so the segmentation results of the algorithm for medical image with intensity
inhomogeneity are poor. In the future, this situation should be improved. From the images and
charts in Section 4 we can see that the DLFCM algorithm has better segmentation results and
Jou

certain advantages in both quantitative and qualitative evaluations.

In summary, we developed a new algorithm for image segmentation based on dictionary learning
for noise reduction and improved fuzzy C-means clustering that is referred to as dictionary learn-
ing Fuzzy C-mean clustering (DLFCM) algorithm. We used a suite of matrices such as partition
coefficient, partition entropy, Xie-Beni function, F1-measure, clustering accuracy, and clustering
purity for algorithm assessment and inter-comparison. Results show that: (1) the newly developed
DLFCM algorithm can segment successfully both synthetic images of various noise levels and real
MRI and CT images; it not only effectively reduces the random noise, but also enhances the seg-
mentation quality by preserving more detailed structure of objects to be segmented; (2) it performs
similarly well as the spectral clustering algorithm for noise-free images, but outperforms the spectral
clustering algorithm for noisy images; and (3) it outperforms the traditional FCM algorithm, the
KFCM algorithm, the SL FCM algorithm, and the MSFCM algorithm for either noise-free or noisy

23
 Journal Pre-proof

images.

Acknowledgements. The authors are very grateful to the anonymous referees for their very
helpful comments and suggestions that improved largely the quality of manuscript. This research is
supported by NSFC (61772003) and NSFC (61702083).

References

of
[1] J.Q. Miao, T.-Z. Huang, X. Zhou, Y. Wang, J. Liu, Image segmentation based on an active contour
model of partial image restoration with local cosine fitting energy, Information Sciences 447 (2018)
52-71.

pro
[2] J.C. Bezdek, R. Ehrlich, W. Full, FCM: The fuzzy c-means clustering algorithm. Computers & Geo-
sciences 10(2-3) (1984) 191-203.

[3] Y.J. Chen, H. Zhang, Y.H. Zheng, B. Jeon, Q.M.J. Wu. An improved anisotropic hierarchical fuzzy
c-means method based on multivariate student t-distribution for brain MRI segmentation. Pattern
Recognition 60 (2016) 778-792.

[4] C.S. Zhang, F.P. Nie, S.M. Xiang, A general kernelization framework for learning algorithms based
re-
on kernel PCA. Neurocomputing 73(4-6) (2010) 959-967.

[5] F.P. Nie, X.Q. Wang, M. I. Jordan, et al., The Constrained Laplacian Rank algorithm for graph-based
clustering, Thirtieth AAAI Conference on Artificial Intelligence (2016) 1969-1976.

[6] F.P. Nie, Z.M. Zeng, I.W. Tsang, et al., Spectral embedded clustering: A framework for in-sample and
out-of-sample spectral clustering, IEEE Transactions on Neural Networks 22(11) (2011) 1796-1808.
lP
[7] F.P. Nie, D. Xu, X.L. Li, Initialization independent clustering with actively self-training method.
IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics) 42(1) (2012) 17-27.

[8] J.Q. Miao. Algorithm of Noise Reduction for Adaptive Dictionary Learning Research Based on Brain
MRI Images, Applied Mechanics and Materials. Trans Tech Publications 687 (2014) 4123-4127.
rna

[9] G. Liu, Z. Zhou, Xie S. Regularized fuzzy clustering for fast image segmentation, Information Science
and Technology (ICIST), 2013 International Conference on. IEEE (2013)1164-1167.

[10] M. Elad, M. Aharon, Image denoising via sparse and redundant representations over learned dictio-
naries, IEEE Transactions on Image processing 15(12) (2006) 3736-3745.

[11] D.Y. Zhang, C.H. Li, X.Z. Li, et al., An Image De-noising Algorithm Based on Redundance Removed
Dictionary. Journal of Xiamen University (Natural Science) 51(4) (2012) 691-695.
Jou

[12] M. Aharon, M. Elad, A. Bruckstein, K-SVD: An algorithm for designing overcomplete dictionaries for
sparse representation, IEEE Transactions on signal processing 54(11) (2006) 4311-4322.

[13] S.G. Mallat, Z.F. Zhang, Matching pursuits with time-frequency dictionaries. IEEE Transactions on
signal processing, 41(12)(1993) 3397-3415.

[14] Y. Li, Y. Shen. An automatic fuzzy c-means algorithm for image segmentation, Soft Computing 14(2)
(2010) 123-128.

[15] Y. Xie, Research on image segmentation algorithm based on fuzzy clustering, Proceedings of SPIE-The
International Society for Optical Engineering 8878(4) (2013).

[16] D.Q. Zhang, S.C. Chen, Clustering incomplete data using kernel-based fuzzy c-means algorithm,
Neural Processing Letters 18(3) (2003) 155-162.

24
 Journal Pre-proof

[17] W.Y. Chen, Y. Song, H. Bai, et al., Parallel spectral clustering in distributed systems. IEEE transac-
tions on pattern analysis and machine intelligence 33(3) (2011) 568-586.

[18] J. Gu, L. Jiao, S. Yang, et al., Sparse learning based fuzzy c-means clustering, Knowledge-Based
Systems 119(C) (2017) 113-125.

[19] E.A. Zanaty. Determining the number of clusters for kernelized fuzzy C-means algorithms for auto-
matic medical image segmentation, Egyptian Informatics Journal 13(1) (2012) 39-58.

of
[20] V. Vapnik, The nature of statistical learning theory, Springer science & business media 2013.

[21] Y. Yang, S.X. Guo, R.L. Tian, et al., Fuzzy C-Means Clustering Algorithm for Image Segmentation
Based on Improved Particle Swarm Optimization. Advanced Materials Research. Trans Tech Publi-

pro
cations 532 (2012) 1553-1557.

[22] R. Kafieh, H. Rabbani, I. Selesnick. Three Dimensional Data-Driven Multi Scale Atomic Represen-
tation of Optical Coherence Tomography, IEEE Transactions on Medical Imaging 34(5) (2015) 1042-
1062.

[23] H. Rabbani, Z. Amini. Classification of Medical Image Modeling Methods: A Review, Current Medical
Imaging Reviews 12(2) (2016) 130-148.
re-
[24] A. Hyvarinen and E. Oja. Independent component analysis: algorithms and applications, Neural
networks, 13 (2000) 411-430.

[25] R. R. Coifman and S. Lafon. Diffusion maps, Applied and computational harmonic analysis 21(1)
(2006) 5-30.

[26] J. Shlens. A Tutorial on Principal Component Analysis, International Journal of Remote Sensing 51(2)
lP
(2014).

[27] K. Engan, S. O. Aase, and J. Husoy. Frame based signal compression using method of optimal direc-
tions (MOD), IEEE International Symposium on Circuits and Systems 4 (1999) 1-4.

[28] Y.C. Pati, R. Rezaiifar, P.S. Krishnaprasad, Orthogonal matching pursuit: Recursive function ap-
proximation with applications to wavelet decomposition, Signals, Systems and Computers, 1993. 1993
rna

Conference Record of The Twenty-Seventh Asilomar Conference on. IEEE (1993) 40-44.

[29] N. R. Pal, K. Pal, J. M. Keller, et al., A possibilistic fuzzy c-means clustering algorithm, IEEE
transactions on fuzzy systems 13(4) (2005) 517-530.

[30] Y. Zheng, B. Jeon, D. Xu, et al., Image segmentation by generalized hierarchical fuzzy C-means
algorithm, Journal of Intelligent & Fuzzy Systems 28(2) (2015) 961-973.
Jou

[31] X.J. Dong, X.Y. Zhang. Fuzzy Clustering Segmentation Algorithm Research on Sports Image. Applied
Mechanics and Materials 339(2013) 297-300.

[32] H. Shen, J. Yang, S. Wang, et al., Attribute weighted mercer kernel based fuzzy clustering algo-
rithm for general non-spherical datasets, Soft Computing-A Fusion of Foundations, Methodologies
and Applications 10(11) (2006) 1061-1073.

[33] Y. He, M.Y. Hussaini, J. Ma, et al., A new fuzzy c-means method with total variation regularization for
segmentation of images with noisy and incomplete data. Pattern Recognition 45(9) (2012) 3463-3471.

[34] B.N. Li, C.K. Chui, S. Chang, et al., Integrating spatial fuzzy clustering with level set methods for
automated medical image segmentation, Computers in biology and medicine 41(1) (2011) 1-10.

[35] C.X. Song, K. Ma, C. Qin, P. Xiao, Infrared image segmentation based on clustering combined with
sparse coding and spatial constraints. Acta Physica Sinica 62(4) (2013) 1-10.

25
 Journal Pre-proof

[36] K.R. Muller, S. Mika, G. Ratsch, et al., An introduction to kernel-based learning algorithms, IEEE
transactions on neural networks 12(2) (2001) 181-201.

[37] K.S. Tan, W.H. Lim, N.A.M. Isa, Novel initialization scheme for Fuzzy C-Means algorithm on color
image segmentation, Applied Soft Computing 13(4) (2013) 1832-1852.

[38] Y. Yang, S. Guo, R. Tian, et al., A Novel Kernel-based Fuzzy c-means Algorithm with Spatial Infor-
mation for Image Segmentation, Energy Procedia 13 (2011) 485-491.

of
[39] H. Fritz, L.A. GarcA-Escudero, A. Mayo-Iscar, Robust constrained fuzzy clustering, Information
Sciences 245 (2013) 38-52.

[40] J. Wang, S. Wang, F. Chung, et al., Fuzzy partition based soft subspace clustering and its applications

pro
in high dimensional data, Information Sciences 246 (2013) 133-154.

[41] L. Liao, T. Lin, B. Li, MRI brain image segmentation and bias field correction based on fast spatially
constrained kernel clustering approach, Pattern Recognition Letters 29(10) (2008) 1580-1588.

[42] K.S. Chuang, H.L. Tzeng, S. Chen, et al., Fuzzy c-means clustering with spatial information for image
segmentation, computerized medical imaging and graphics 30(1) (2006) 9-15.

[43] I.A. Yusoff, N.A.M. Isa, K. Hasikin, Automated two-dimensional K-means clustering algorithm for
re-
unsupervised image segmentation, Computers & Electrical Engineering 39(3) (2013) 907-917.

[44] H. Liu, C.M. Zhang, et al., Research on a pulmonary nodule segmentation method combining fast self-
adaptive FCM and classification, Computational and mathematical methods in medicine 4(1) (2015)
1-14.

[45] J. Wang, J. Yang, K. Yu, et al., Locality-constrained linear coding for image classification. Computer
lP
Vision and Pattern Recognition 119 (5) (2010) 3360-3367.

[46] J. Yang, K. Yu, Y. Gong, et al., Linear spatial pyramid matching using sparse coding for image
classification. Computer Vision and Pattern Recognition, 2009. IEEE Conference on (2009) 1794-
1801.

[47] S. Gao, I.W.H. Tsang, L.T. Chia. Kernel sparse representation for image classification and face recog-
rna

nition, European Conference on Computer Vision (2010) 1-14.

[48] W. Dong, G. Shi, X. Li, et al., Compressive sensing via nonlocal low-rank regularization. IEEE
Transactions on Image Processing 23(8) (2014) 3618-3632.

[49] J. Xu, Z.G. Chang, Self-adaptive image sparse representation algorithm based on clustering and its
application. Chin Acta Photonica Sinica 28 (2011) 40-45.
Jou

[50] M. Hong, Z.Q. Luo. On the linear convergence of the alternating direction method of multipliers.
Mathematical Programming 162(1-2) (2017) 165-199.

[51] S. Ravishankar, Y. Bresler, Learning sparsifying transforms, IEEE Transactions on Signal Processing
61(5) (2013) 1072-1086.

[52] S. Ravishankar, Y. Bresler, Learning sparsifying transforms for image processing, Image Processing
(ICIP), 2012 19th IEEE International Conference (2012) 681-684.

[53] Z.C. Lipton, C. Elkan, B. Naryanaswamy, Optimal Thresholding of Classifiers to Maximize F1 Mea-
sure, Machine Learning and Knowledge Discovery in Databases (2014) 225-239.

[54] M. Brun, S. Chao, J. Hua, et al., Model-based evaluation of clustering validation measures, Pattern
Recognition 40(3) (2007) 807-824.

26
 Journal Pre-proof

[55] S.C. Sripada, D.M.S. Rao, Comparison of purity and entropy of k-means clustering and fuzzy c means
clustering, Indian Journal of Computer Science and Engineering (IJCSE) 2(3) (2011) 343-346.

[56] Z. Yu, O.C. Au, R. Zou, et al., An adaptive unsupervised approach toward pixel clustering and color
image segmentation, Pattern Recognition 43(5) (2010) 1889-1906.

[57] H. Rabbani, R. Nezafat, and S. Gazor, Wavelet-domain medical image denoising using bivariate
laplacian mixture model, IEEE transactions on biomedical engineering 56(12) (2009): 2826-2837.

of
[58] A.W.C. Liew, H. Yan, An adaptive spatial fuzzy clustering algorithm for 3-D MR image segmentation,
IEEE transactions on medical imaging 22(9) (2003) 1063-1075.

[59] H. M. Ali, MRI medical image denoising by fundamental filters, High-Resolution Neuroimaging-Basic

pro
Physical Principles and Clinical Applications IntechOpen (2018) 111-124.

[60] R. D. Nowak, Wavelet-based Rician noise removal for magnetic resonance imaging, IEEE Transactions
on Image Processing 8(10) (1999) 1408-1419.

re-
lP
rna
Jou

27
Highlights:
Journal Pre-proof

• A new image segmentation algorithm was developed based on dictionary learning and
improved fuzzy C-Means Clustering.
• The algorithm removes non-target areas and extracts areas of interests.
• Segmented images by the algorithm have better visualization and accuracy than conventional
algorithms.

of
pro
re-
lP
rna
Jou
 Journal Pre-proof
Declaration of Interest Statement
We declare that we have no financial and personal relationships with other people or
organizations that can inappropriately influence our work, there is no professional or
other personal interest of any nature or kind in any product, service and/or company
that could be construed as influencing the position presented in, or the review of, the
manuscript entitled.

of
pro
re-
lP
rna
Jou