Вы находитесь на странице: 1из 10

Pattern Recognition Letters 34 (2013) 13291338

Contents lists available at SciVerse ScienceDirect

Pattern Recognition Letters


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

A modied interval type-2 fuzzy C-means algorithm with application in MR image segmentation
Cunyong Qiu , Jian Xiao, Long Yu, Lu Han, Muhammad Naveed Iqbal
School of Electrical Engineering, Southwest Jiaotong University, P.O. Box 606, Chengdu 610031, PR China

a r t i c l e

i n f o

a b s t r a c t
The fuzzy C-means (FCM) algorithm has signicant importance compared to other methods in Medical image segmentation. Conventional FCM algorithm is sensitive to noise especially in the presence of intensity inhomogeneity in MRI. Main reason is that a single fuzzier in FCM cannot properly represent pattern memberships for all clusters. In this paper, we present a novel algorithm for fuzzy segmentation of MRI data. The algorithm utilizes two fuzziers used in interval type-2 FCM and a spatial constraint on the membership functions. Also, in our investigation, validity functions are extended to generalized form for interval type-2 fuzzy clustering. The experimental results on both synthetic and MR images show that the proposed algorithm has better performance on image segmentation than conventional FCM based algorithms. 2013 Elsevier B.V. All rights reserved.

Article history: Received 7 April 2012 Available online 10 May 2013 Communicated by N. Sladoje Keywords: Image segmentation Magnetic resonance imaging Fuzzy C-means Interval type-2 fuzzy sets

1. Introduction In the applications of medical imaging, use of computers and digital techniques for analyses and processing has become necessary to deal with the increasing size and number of medical images. A variety of techniques are available for diagnostic imaging, such as magnetic resonance imaging (MRI), computed tomography (CT), digital mammography and so on. These technologies provide an effective means for noninvasively mapping the anatomy of a subject and are critical part in diagnosis and treatment planning. MRI has an edge over other diagnostic imaging methods (Wells et al., 1996) and the majority of researches and practitioners pertain to its use for MR images. In literature many algorithms for MR image segmentations have been presented (Balafar et al., 2010; Bezdek et al., 1993; Clark et al., 1994; Clarke et al., 1995; Kannan et al., 2010; Zhang and Chen, 2004). In general, fuzzy segmentation method, especially the fuzzy C-means algorithm (FCM) (Bezdek, 1981) is more effective with considerable benets. Unlike the hard clustering methods which force pixels to link exclusively to one cluster, FCM allows pixels to have relation with multiple clusters with varying degrees of membership. The introduction of fuzziness in FCM enables the clustering methods to retain more information from the original image, compared with other segmentation methods. Some modality specic image artifacts such as the phenomena of intensity inhomogeneity affect quantitative image analysis. Intensity inhomogeneity arises from the imperfections present in the image acquisition process. The effect is not usually noticeable
Corresponding author. Tel.: +86 156 0804 1011.
E-mail address: chaucer_qcy@my.swjtu.edu.cn (C.Y. Qiu). 0167-8655/$ - see front matter 2013 Elsevier B.V. All rights reserved. http://dx.doi.org/10.1016/j.patrec.2013.04.021

to a human observer but has great impact on image segmentation. To deal with this problem, many modied FCM-based algorithms have been proposed. In Ahmed et al. (2002), Chuang et al. (2006), Liew et al. (2000), Worth, Yang et al. (2009) the spatial information was incorporated into conventional FCM for better segmenting images. He and Greenshields, (2008), Tam et al., (2009) introduced Markov random eld (MRF) theory into FCM. Similarly, neural network (NN) approach was also incorporated in Birgani et al. (2008), Shen et al. (2005) to optimize the degree of spatial constraint attraction. Besides, a number of powerful kernel-based learning machines, e.g. support vector machine (SVM) (Cristianini and Shawe-Taylor, 2000) and kernel sher discriminant (KFD) (Mika et al., 1999) were proposed and have got successful applications in pattern recognition (Shawe-Taylor and Cristianini, 2004). Some of these kernel-based methods have been successfully used in FCM with application to image segmentation (Kannan et al. 2010; Zhang and Chen 2004; Zhang et al. 2007). An intrinsic shortcoming of conventional FCM algorithm is the existence of only one fuzzier m. Hence it is quite difcult to deal with patterns with different volumes. We will explore this problem in detail in next section. Secondly, FCM-based algorithms are all based on type-1 fuzzy set, which could not handle image pattern uncertainties well. The crisp membership functions (MF) used in type-1-set-based FCM has limitation on expressing the uncertainties in images and algorithms. The type-2 fuzzy set, as an extension of type-1 fuzzy set, has a secondary membership to dene the possibilities of uncertainties in primary membership. Thus the performance of handling uncertainties of type-2 fuzzy set is enhanced (Mendel, 2001, 2007; Mendel and John, 2002). In this paper, we present a modied interval type-2 fuzzy C-means algorithm

1330

C.Y. Qiu et al. / Pattern Recognition Letters 34 (2013) 13291338

(IT2MFCM) for MR image segmentation. Two fuzziers are used to overcome above mentioned discrepancies and handle the uncertainties in MR images. A spatial membership function is proposed to impose spatial information and used to modify the traditional MF in interval type-2 FCM. We will mention that in all previous published literature, the validity functions are supposed only available for FCM-based algorithms (Bezdek, 1973; Fukuyama and Sugeno, 1989; Xie and Beni, 1991); therefore here we extend the type-1 validity functions to a generalized form for interval type-2 FCM based algorithms. The rest of this paper goes as follows. Some basic concepts on type-2 fuzzy set are briey introduced and the main steps of IT2FCM are listed in Section 2. In Section 3, the use of spatial constraints is summarized and the interval type-2 FCM with spatial constraints is presented. Section 4 explores the validity functions for interval type-2 FCM based algorithms. The selection of fuzziers is also demonstrated in this section. Experimental comparisons of MR image segmentation procedures are presented in Section 5. Finally, our conclusions and some issues for future study are mentioned in Section 6. 2. Interval type-2 fuzzy C-means algorithm

uncertainty existing in fuzzier m. Hwang and Rhee (2007) used two fuzziers m1 and m2 to dene and manage uncertainty to establish a suitable maximum fuzzy boundary as shown in Fig. 2 case (c). The memberships of pattern sets were extended to interval type-2 fuzzy sets using the two fuzziers and the conventional FCM algorithm was extended to interval type-2 fuzzy C-means algorithm (IT2FCM) accordingly. Both the FCM and IT2FCM algorithms are objective function based clustering methods used to minimize the distance between a pattern and cluster prototype. Similar to FCM, the IT2FCM is an iterative optimization algorithm that minimizes the objective function as:

J m U ; V

c X n X i1 k1

lik m dik

where fuzzier m is replaced by m1 and m2 which represent different fuzzy degrees and give different objective functions. To minimize the objective function Jm, we have

lik 1
2

c X dik =djk 2=m1 j 1

2.1. Type-2 fuzzy set Zadeh (1975) introduced type-2 and higher-type fuzzy sets in 1975 and eliminated the paradox of type-1 fuzzy systems which has the discrepancy that membership grades are themselves precise real numbers. A type-1 fuzzy set has a grade of membership that is crisp while a type-2 fuzzy set has a grade of membership that is fuzzy. A type-2 fuzzy set (Mendel, 2001, 2007; Mendel and John, 2002), denoted , is characterized by a type-2 membership function lA ~ x; u as:

where dik kxk v i k is the distance between input patterns xk and cluster centers vi (k k is the Euclidean norm). In IT2FCM, upper and lower memberships are employed to manage uncertainty for fuzzier m. FOU of fuzzier m is created by two different fuzziers m1 and m2. Upper (lower) membership  ik (lik ) is the upper (lower) bound of the FOU, which can be del rived from (3) as:

lik

8 X c c X > > 1= dik =djk 2=m1 1 ; dik =djk < c > > < j 1 j1
c X > > 2=m2 1 > > ; otherwise : 1= dik =djk

e fx; l; l ~ x; lj8x 2 X ; 8l 2 J # 0; 1g A x A

where l 2 J x is the primary membership of x, which is also the domain of the secondary membership function lA ~ x; u: As shown in Fig. 1, the shaded region shows uncertainty present in the primary memberships of a type-2 fuzzy set, which is the union of all primary memberships. It is also called footprint of uncertainty (FOU). An implementation of general type-2 fuzzy sets needs more computational complexity as compared to type-1 fuzzy sets because we have incorporated secondary grades for each primary MFs. The interval type-2 fuzzy set is a special case of general type-2 fuzzy set (Liang and Mendel, 2000). It enables us to significantly reduce the computational complexity as all secondary grades for an interval type-2 fuzzy set are uniformly weighted (lA ~ x; u 1) as shown in Fig. 1(b). 2.2. It2fcm Consider the case (a) in Fig. 2. Two cluster boundaries C1 and C2, considered to be identical in structure and density, are separated by a vertical line where patterns have equal distance from two cluster centers v1 and v2. The vertical line can be considered as a decision line on the right (left) of patterns belonging to cluster C2 (C1). The shaded boundary is expanded by fuzzier m and maximum fuzzy boundary becomes wider with the increment of m. FCM based methods can perform well on patterns with similar volumes of hyper-spherical shape and density by adjusting the fuzzier m as shown in below gure, case (a). However, the fuzzier m, which can signicantly affect the performances of FCM, treats all clusters equally and has no idea about patterns with different volumes as shown in case (b). It is apparent that FCM based algorithms with only one fuzzier cannot handle

lik

8 X c c X > > 1= dik =djk 2=m1 1 ; dik =djk P c > > < j 1 j1
c X > > 2=m2 1 > > ; otherwise : 1= dik =djk j 1

j 1

Different from FCM, the output of IT2FCM algorithm is an interval type-2 fuzzy set which cannot be transformed to crisp set by defuzzier directly. Type-reduction is the rst and important step of output processing that computes the centroid of a type-2 fuzzy set and reduces type-2 fuzzy set to type-1 fuzzy set (Karnik and Mendel, 2001a). Centroid type-reducer is employed here for getting accurate cluster centers and desirable clustering results. Afterwards a centroid defuzzier is used to obtain a crisp center from the type-reduced interval type-1 fuzzy set (Karnik et al. 1999; Karnik and Mendel 2001b). The key steps of IT2FCM can be summarized as: Step 1. Set the number c of the cluster prototypes (2 6 c 6 N ) and fuzziers m1, m2. Set e > 0 to a very small value and the iterative number l = 0. Initialize randomly prototypes V(l). Step 2. Calculate dik and the upper and lower partition matrices U and U respectively, using (4). Step 3. Update the prototypes to V(l+1). The upper and lower fuzzy partition matrixes are used during the iterative process of IT2FCM and consequently the estimated interval type-1 fuzzy set [vL, vR] for the cluster centers will be obtained. An iterative algorithm proposed in Karnik and Mendel (2001a) is adopted to estimate both ends of the interval fuzzy set. In KM iterative algorithm, we take the calculation of vR for example, the pattern set is arranged in ascending order

C.Y. Qiu et al. / Pattern Recognition Letters 34 (2013) 13291338

1331

Fig. 1. Type-2 membership functions. (a) Gaussian type-2 MF. (b) Interval type-2 MF.

Fig. 2. Decision boundary for two clusters in different cases. (a) Same volume. (b) Different volumes. (c) Different volumes and two fuzziers.

(x1 6 x2 6 6 xN ). Find interval index k such that xk V , and for all patterns set membership lxi lxi if i 6 k  xi if i > k. Then the maximum value vR can and lxi l be calculated using (5). The membership we chose to update vR is called lR ik : The minimum of center vL can be obtained using the same procedure except setting membership  xi if i 6 k and lxi lxi if i > k, where lL lxi l ik is used instead.

Then the crisp center v is simply obtained by a defuzzication method as: v v L v R =2. Step 4. Stop iterating if kV l1 V l k < e is satised, otherwise go to step 2. Step 5. Type-reduce the interval type-2 fuzzy partition matrix as:

lik

R lL ik lik

P l xi xi V P lxi

R where lL ik and lik are left and right memberships obtained during the iterative process in step 3. Finally assign all patterns to clusters

1332

C.Y. Qiu et al. / Pattern Recognition Letters 34 (2013) 13291338

by using the maximum membership value of every pattern. For instance, xk 2 v i if lik maxflik g.
16i6c

3.2. Modied inteval type-2 fuzzy C-means algorithm In the traditional FCM algorithm, the clustering of xk with class i depends on the membership value lik . To take into account the spatial information of correlated neighboring pixels xa, here we propose a spatial membership function for the pixel xk (called lik s ) which can be stated as:

Unfortunately, both the FCM and IT2FCM algorithms are sensitive to noise. Denoising can be used before segmentation, but inevitably the details of images will be lost and it is hard to balance denoising and segmentation. As a result, many extended algorithms using dependable spatial constraints have been proposed to restrict the fuzzy partition matrix. Section 3 demonstrates the use of spatial information. 3. Modied interval type-2 fuzzy C-means algorithms 3.1. Introduction of spatial information In FCM and IT2FCM image segmentation algorithms, features of pixels are assumed to be independent of each other. But in MR images, neighboring pixels have strong correlation and generally dependent on each other. Thus incorporating local spatial interaction among adjacent pixels in the fuzzy clustering process is more practical approach. It is also helpful to reduce noise distortion and intensity inhomogeneity on MR image segmentation. A pixel k and its neighbors are employed to form local image windows. Larger local image windows can be used to incorporate more local spatial information for better segmentation results. Here we use a 3 3 local image window to simplify the calculation. The distances and memberships of adjacent pixels are most widely used while adding spatial information into FCM based algorithms. Wang and Bu, (2010) proposed a modied FCM algorithm using a local similarity measure model which mainly depends on local spatial relationship and local gray-level relationship. Besides, Liu and Zhao, (2012) proposed a novel FCM algorithm called ANLS_FIS using non local spatial information to make the standard FCM more robust to noise. The objective functions of FCM based algorithms using spatial information mentioned above can be summarized and formulated as:

lik s

lia lak

a2N k

where lia represents the fuzzy membership of pixels xa to cluster center i and lak is a function that reects the connection between pixel xk and its neighboring pixels xa. This spatial membership function should satisfy the following conditions: (1) As a membership function, the value of lik s lies in [0, 1] to indicate the importance of the spatial feature. (2) If all neighboring pixels xa completely belong to cluster i (that is lia = 1 with a 2 N k ), the function value will be lik s = 1 correspondingly. This implies that pixel xk is mostly occupied by its neighbors. On the contrary, lik s = 0 for lia = 0. (3) When both the function value lik s and lia equal to 1, then from (8):

X
a2N k

lak 1

Function lak should also satisfy that a longer distance between xk and xa gives rise to a smaller value of lak . To meet the conditions above, we consider the neighboring pixels as cluster centers, and lak can be regarded as the membership function of pixel xk to cluster centers xa. As a result, the function lak can be formulated as:

lak

2 !1=m1 31 Nk 2 X d xa ; xk 4 5 2 d xl ; x k l1

10

J m U ; V

c X n c X n X X aX m 2 l lik m dia 2 ik dik N R i1 k1 a2N i1 k1


k

c X n X b X l m 1 lia m NR i1 k1 ik a2N
k

where Nk is the neighborhood of xk, NR is the area of Nk. Spatial constraints can be incorporated into algorithms for compensation as last two additional terms shown in (7). Distance and membership information (dia and lia ) of adjacent pixels are used in the two terms respectively, which quantize the belongingness of adjacent pixels to one cluster center and exert inuence on the belongingness of pattern xk to that center. The algorithm proposed in Ahmed et al. (2002), Liu and Zhao (2012) had an objective function formed with the penultimate term in (7), where a is a weighting coefcient to modulate the intensity of compensation term. The other algorithms using spatial constraint, as proposed in Pham (2002), Yang et al. (2009), has the objective function as last term in (7) whereas b has the same meaning as a. We also mention here that distance used in conventional FCM based algorithms is Euclidean norm, which is not suitable for revealing non-Euclidean structure of the input data. Zhang and Chen (2004) proposed a novel FCM algorithm with a robust nonEuclidean distance measures induced by the kernel methods. Here we use l ik and dik to denote the modied membership functions and distances. We can observe that if none of these spatial information are used (a = 0, b = 0, l ik , dik return to lik , dik ), these algorithms will return back to conventional FCM algorithm.

while we take into account the two fuzziers m1 and m2 used in IT2FCM to calculate lak , upper and lower spatial membership func ik s and lik s ) will be generated correspondingly according to tions (l (4). The introduction of such spatial functions into IT2FCM causes different combinations in deriving upper and lower fuzzy partition matrix, which makes the algorithm more difcult in analysis and computation. Actually spatial information is an addition term for which a slight variation has little inuence on the segmentation results. Hence we consider a similar spatial membership function for both the upper and lower partition matrix and set m 1=m1 m2 .  ia lia =2 for simplication. Hence funcSimilarly, we take lia l tion lik s can be derived from (8) and (10) as:

lik s

2 !1=m1 31 Nk 2 X d xa ; xk 4 5 lia  2 d xl ; x k a1 l1 ! Nk h i X 2 1=m1 lia d xa ; xk


Nk X

0 Nk X @
l 1

a1

1 d xl ; xk 1
2

!1=m1 11 A

0 @

Nk X l 1

!1=m1 11 N ! k X l ia A 2 2 1=m1 d xl ; xk a1 d xa ; xk

11

lik s is an impact factor of neighboring pixels in Nk to pixel xk and is incorporated into membership functions of the upper and lower fuzzy partition matrix. The distances utilized in FCM and

C.Y. Qiu et al. / Pattern Recognition Letters 34 (2013) 13291338

1333

IT2FCM are formulated as dik kxk v i kA where it is assumed that features for xk to any vi have equal importance. In a practical scenario, this may not be true. Therefore, we modify the distance by incorporating the spatial membership function lik s into it to change the weight of feature for xk to different vi and modied dis tance dik is formulated as:
2 dik

Combining (12) into (13) for the rst expression, we have:

 ik c ik x X
j1

d =lik s 1=m1 1 djk =ljk s 1=m1 1 ! !


2

2 dik =

lik s

12

c X 2 2=m 1 dik =lik s 1=m1 1 1= dlk 1 l 1

These distances have varying degrees of expansion as value of lik s lies in [0, 1]. We observe that this feature is disregarded when lik s 0 and the distance d ik is expanded to innity correspondingly, which implies that pixel xk has innitesimal relation to cluster vi through the inuence of neighboring pixels. Generally distance dik is ranged in [dik, 1], and as closer it is to dik, greater it has impact on clustering. The modied distance dik is used in IT2FCM instead of dik and the new upper and lower membership functions are computed as follows:

c c X X 2 2=m 1 djk =ljk s 1=m1 1 1= dlk 1 j1 c X 2=m 1 2=m 1 dik 1 = dlk 1

l1

lik !

1=m1 1 s

" c X
j1

l1 2=m 1 djk 1 = c X 2=m 1 dlk 1 l 1

# l
1=m1 1 jk s

 ik x

8 X c c X 2=m1 1 > > ; dik =djk < c > 1= dik =djk > < j1 j 1
c X > > 2=m2 1 > > ; otherwise : 1= dik =djk

1=m1 1  ik lik l s c X 11=m1 1  jk ljk s l j1

14

Pc

The

modied

membership

function

still

ensures

that

i1

 ik for example: xik 1, here we take x

xik

8 X c c X 2=m1 1 > > ; dik =djk P c > 1= dik =djk > < j1 j 1 > > > > : 1=
c X j1

j1

13

dik =djk 2=m2 1 ; otherwise

c X 1=m1 1  ik lik l C B s 1=m1 1 c c B X X C i 1  ik lik l s C B  ik x 1 c c C X BX 1=m1 1 A 1=m1 1 i 1 i1 @  jk ljk  l l l jk jk s s j1 j 1

15

Pattern set

Original IT2 membership update Estimate FCM membership (m=m1) Estimate FCM membership (m=m2)

Determine upper and lower memberships

Modified IT2 membership update Calculate spatial membership ik _ s

Update modified IT2 membership ik using (16)

Center update Calculate center V using ik

Type reduction by KM iterative procedure

Defuzzification

V ( l +1) V ( l ) <

No

Yes Hard partition


Fig. 3. Modied interval type-2 fuzzy C-means algorithm.

1334

C.Y. Qiu et al. / Pattern Recognition Letters 34 (2013) 13291338

Finally, the total membership function is deduced as:

 ik x

8 1=m 1 1=m 1  ik l 1  ik l 1 l l > ik s ik s > ; X >1 > c c >X c > > 1 = m 1 1 = m 1 >  jk l 1  jk l 1 l l > jk s jk s > < j 1 j 1
1=m 1 >  ik l 2 l > ik s > ; otherwise > c > X > 1=m 1 > >  jk l 2 l > jk s : j 1 8 1=m 1 1=m 1 lik lik s1 lik lik s1 > > ; X 61 > c c > c X > > 1=m1 1 1=m1 1 > l l l l > jk jk s jk jk s > < j 1 j 1 1=m 1 > lik lik s2 > > ; otherwise > c > X > > 1=m 1 > ljk ljk s2 > :

16

xik

j 1

As it is implicit, the core of FCM is fuzziness, which makes the algorithms more exible and closer to human thinking. For the common use of spatial information, the constraint is a xed value

for a pixel to all the clusters, and the inuence of spatial constraint is adjusted by a given parameter which greatly affects the segmentation results. However, the selection of the tune parameter is generally difcult since that the parameter should keep a balance between insensitiveness to noise and effectiveness of preventing the image details. In our investigation, we have dened a spatial membership function to describe the spatial constraints in fuzzy sense. The spatial MF indicating the degree of a pixel relation to its neighboring pixels is similar to the MFs in FCM indicating the degree of a pixel relation to a cluster center. This is particularly useful when the pixels and their neighbors are not well separated and ambiguous. Moreover, the spatial MF could help us discover more sophisticated relations between a given pixel and its correlated neighboring pixels. We have integrated the spatial MF into partition matrix instead of a compensation term shown in (7) and some experimental comparisons are presented below to demonstrate the effectiveness of our proposed method. Our overall proposed IT2MFCM algorithm can be summarized in Fig. 3.

Fig. 4. Comparison of segmentation results on synthetic image corrupted by different noise.

C.Y. Qiu et al. / Pattern Recognition Letters 34 (2013) 13291338 Table 1 Comparison of validity functions for different images. For each validity function, the best performance is made bold. Segmented image Technique Validity functions

1335

q (%)
Synthetic image with Gaussian 0.01 RFCM IT2RFCM ANLS_FIS IT2MFCM RFCM IT2RFCM ANLS_FIS IT2MFCM RFCM IT2RFCM ANLS_FIS IT2MFCM RFCM IT2RFCM ANLS_FIS IT2MFCM RFCM IT2RFCM ANLS_FIS IT2MFCM RFCM IT2RFCM ANLS_FIS IT2MFCM RFCM IT2RFCM ANLS_FIS IT2MFCM RFCM IT2RFCM ANLS_FIS IT2MFCM 98.41 98.69 99.53 99.37 95.35 96.18 98.90 98.61 97.81 98.03 97.91 98.77 93.30 93.85 93.11 96.82 98.64 98.83 99.13 99.80 98.38 98.61 98.93 99.71

Vpc 0.98203 0.98977 0.98746 0.99641 0.95162 0.95956 0.96631 0.98439 0.96850 0.97367 0.97033 0.99325 0.93405 0.93829 0.93261 0.96103 0.98725 0.98953 0.98933 0.99465 0.97941 0.98823 0.99048 0.99373 0.90274 0.91687 0.92204 0.97408 0.87753 0.91662 0.90506 0.95241

Vpe 0.04667 0.02909 0.03122 0.01571 0.11548 0.09841 0.07334 0.04335 0.08956 0.06645 0.07927 0.01841 0.13443 0.13197 0.13538 0.07101 0.035511 0.028830 0.25316 0.020151 0.039982 0.036014 0.033804 0.022160 0.17458 0.15529 0.14985 0.09825 0.21597 0.17126 0.18335 0.13326

Vfs (103) 12.056 11.576 12.383 12.189 10.664 10.689 11.899 11.664 11.829 11.959 11.872 12.165 8.8513 9.0218 8.6231 10.985 11.663 11.786 11.977 12.047 10.868 11.433 11.535 11.997 2.3857 2.1532 2.3772 2.3674 1.4323 1.4499 1.4384 1.5517

Vxb 0.023463 0.016168 0.008368 0.010904 0.067595 0.054159 0.027734 0.029231 0.039274 0.035383 0.037980 0.023629 0.092235 0.089883 0.093341 0.045132 0.012770 0.012113 0.011942 0.010502 0.018393 0.015583 0.014797 0.011056 0.068311 0.064822 0.061156 0.057438 0.073356 0.065363 0.069476 0.057438

Synthetic image with Gaussian 0.05

Synthetic image with salt and pepper 0.01

Synthetic image with salt and pepper 0.05

MR image in Fig. 6(a)

MR image in Fig. 6(c)

Real MR image in Fig. 6(m)

Real MR image in Fig. 6(r)

4. Validity functions To evaluate the segmentation performance quantitatively, the most intuitive way is comparing the segmentation results with ground truth. Here we employ the similarity index q in our experiment. For a given image, suppose that Ai and Bi represent the sets of pixels belong to cluster i in segmentation result and in ground truth, then the similarity index is dened as
c 1X 2jAi \ Bi j q 100% c i1 jAi j jBi j

When Vpc is maximal (0 6 Vpc 6 1), optimal clustering is achieved. (1) Vpe is the partition entropy value, and small value of Vpe indicate better performance. We will state it as:

( ) c X n hq i 1 X   V pe U ; U ; c lik lik loglik lik 2n i1 k1

19

17

Notice that q is ranged in [0, 1] and the optimal clustering is achieved when q 1. Because the ground truth of segmentation for real MRI images is not usually available, its impossible to use the similarity index to evaluate the segmentation performance of real images. Instead, several popular validity functions to estimate fuzzy clustering methods are proposed in published literature (Bezdek, 1973; Fukuyama and Sugeno, 1989; Xie and Beni, 1991). However, none of these results demonstrate interval type-2 fuzzy clustering algorithms. Here we will extend these functions to a generalized form, which are available for both FCM and IT2FCM based algorithms. (1) Vpc is the partition coefcient, an important indicator of fuzzy partition, and provides better performance with less fuzziness. It is dened as follows:

(2) The two functions above are only related to partition matrix and lack a direct connection to image features. To evaluate the performance of clustering, two validity functions Vfs and Vxb are used, and for interval type-2 fuzzy clustering are dened as:

V fs U ; U ; V ; X

c X n X i1 k1

 ik lik kxk v i k2 kv i v  k2 l  ik lik kxk v i k2 l 

20

Pc Pn V xb U ; U ; V ; X
Pc 1
i1

n minfkv i v j k2 g
i j

k1

21

V pc U ; U ; c

c X n 1X  ik lik l n i1 k1

18

 c i1 v i , Vfs is FukuyamaSugeno function and Vxb is where v XieBeni function. When either Vfs or Vxb is smaller, a better clustering result is achieved. Here we use same names for modied validity functions for an easy comparison between FCM and IT2FCM segmentation results. Here if m1 = m2 = m, IT2FCM will return to FCM and these generalized functions will have a same formulation with their prototypes.

1336

C.Y. Qiu et al. / Pattern Recognition Letters 34 (2013) 13291338

Fig. 5. Comparison of Vpc via varying fuzziers.

5. Experimental results and discussions In this section, we describe some experimental results to compare the segmentation performance of RFCM, IT2RFCM, ANLS_FIS and IT2MFCM algorithms, where RFCM is a robust FCM algorithm proposed in Yang et al. (2009) and mentioned in Section 3. And we extend RFCM to an interval type-2 version called IT2RFCM for comparison. The IT2RFCM has a same spatial constraint as RFCM excepting the utilization of two fuzziers to represent and manage uncertainty. ANLS_FIS proposed in Liu and Zhao (2012) is a modied FCM algorithm using non local spatial information. We test the algorithms on synthetic image with different noises and different MR slices. The segmented images and validity functions are evaluated to compare the degree of effectiveness of four methods. 5.1. Synthetic image segmentation A synthetic image (256 256 pixels) consisting of three classes is generated to detect the noise immunity of three methods. Gaussian white noise with variance 0.01, 0.05 and salt and pepper noise with noise density 0.01, 0.05 are employed in our experiments. The original synthetic image without noise is set as ground truth as shown in Fig. 4(u) and the corresponding segmentation results are shown in Fig. 4(a)(t) and Table 1. Visually, the segmentation performances of all algorithms drop with the increase of noise strength. We observe that some noise still remain in results of RFCM and IT2RFCM, since the weights of spatial information in RFCM and IT2RFCM are adjusted by a coefcient which has to be settled to a large value while segmenting

noisy images. This may lead the weights of spatial information exceed the feature of a pixel itself and some pixels can be mistakenly segmented. The use of non local spatial information makes ANLS_FIS achieve better performance on segmenting images with Gaussian noise and geometrical edges. However, its hard to determine the parameters of non local spatial information in ANLS_FIS while segmenting images with salt and pepper noise. Table 1 gives the validity functions of four methods for a quantitative comparison. As shown in Fig. 4 and Table 1, IT2MFCM is more robust to different noise compared with other algorithms. Notice that because of the injection of interval type-2 fuzzy set and spatial constraint, IT2RFCM and IT2MFCM need more execution time than RFCM and ANLS_FIS. Just like fuzzier m in FCM and RFCM, m1 and m2 in IT2FCM and IT2MFCM are also adjustable parameters. Here we give a comparison with a validity function Vpc of four algorithms by varying fuzziers to demonstrate the inuence of fuzziers on clustering results. As shown in Section 2, the selection of fuzziers has significant effect on the performance of MR image segmentation. As shown in Fig. 5, the clustering results go worse as the fuzziers values increase. In Fig. 5(a) and (b), the partition coefcient in RFCM is a little smaller than that in FCM. The reason is that the using of spatial penalty term in RFCM decreases the values of partition matrix U. In Fig. 5(c) and (d), due to the introduction of spatial membership, IT2MFCM is much more robust to the variation of fuzziers compared with IT2FCM. Consequently the curve of our proposed algorithm is smoother than that of IT2FCM. Besides, the use of interval type-2 fuzzy set enables the algorithm to handle uncertainty in fuzziers while type-1-set based FCM has limitation on expressing such uncertainty.

C.Y. Qiu et al. / Pattern Recognition Letters 34 (2013) 13291338

1337

Fig. 6. Comparison of segmentation results on different MR images.

5.2. MR slices Because that the ground truth of real MR image segmentation is not usually available, we utilize two synthetic MR images produced by an MRI simulator from the Internet Brain Segmentation Repository (IBSR). Since the truth is known, these data enable us to evaluate the performance of different algorithms. Besides, we have obtained a male subject data provided by IBSR of the Center for morphometric Analysis at Massachusetts General Hospital (Worth). The T1weighted MRI scans were performed on a 1.5 T General Electric Signa with the following parameters: TR = 40 ms, TE = 5 ms, ip angle = 40, eld of view = 24 cm, matrix = 256 256, and averages = 1. Fig. 6 presents a comparison of segmentation results among the algorithms when applied on different MR slices mentioned above. The corresponding comparison scores for Fig. 6 are given in Table 1. As shown in Fig. 6(a)(l), all the algorithms with spatial constraints eliminate the effect of noise. However, for RFCM and IT2RFCM, the coefcient has to be settled to a large value to reduce the inuence of noise, consequently the edge blurring effects exist

in the segmentation results of RFCM and IT2RFCM. Some pixels on the edge are mistakenly segmented as marked in red circle in Fig. 6(m)(v). The edge blurring effects are not distinct in Fig. 6(a)(l) but the missegmented pixels indeed exist in the results of RFCM and IT2FCM. This can be shown in the comparison of the similarity index q in Table 1. ANLS_FIS and IT2MFCM have better performance on segmenting pixels on the edge, but for ANLS_FIS its difcult to nd suitable parameters to eliminate the inuence of noise compared with IT2MFCM. 6. Conclusion In this paper, a modied interval type-2 fuzzy C-means algorithm is proposed with applications to MR image segmentation. Two different values of fuzziers (m1 and m2) are employed to represent and manage uncertainty which occurs when determining the fuzzy degree for clusters with different volume. Furthermore, a spatial constraint is used as a spatial membership function and is blended into the membership function of IT2FCM. The spatial MF shows the degree of a pixel belonging to its neighbors and

1338

C.Y. Qiu et al. / Pattern Recognition Letters 34 (2013) 13291338 He, L., Greenshields, I.R., 2008. An MRF spatial fuzzy clustering method for fMRI SPMs. Biomedical Signal Processing and Control 3 (4), 327333. Hwang, C., Rhee, F.C.H., 2007. Uncertain fuzzy clustering: interval type-2 fuzzy approach to C-means. IEEE Transactions on Fuzzy Systems 15 (1), 107120. Kannan, S.R., Ramathilagam, S., Sathya, A., Pandiyarajan, R., 2010. Effective fuzzy Cmeans based kernel function in segmenting medical images. Computers in Biology and Medicine 40 (6), 572579. Karnik, N.N., Mendel, J.M., 2001a. Centroid of a type-2 fuzzy set. Information Science 132 (1), 195220. Karnik, N.N., Mendel, J.M., 2001b. Operations on type-2 fuzzy sets. Fuzzy Sets and Systems 122 (2), 327348. Karnik, N.N., Mendel, J.M., Liang, Q., 1999. Type-2 fuzzy logic systems. IEEE Transactions on Fuzzy Systems 7 (6), 643658. Liang, Q., Mendel, J.M., 2000. Interval type-2 fuzzy logic systems: theory and design. IEEE Transactions on Fuzzy Systems 8 (5), 535550. Liew, A.W.C., Leung, S.H., Lau, W.H., 2000. Fuzzy image clustering incorporating spatial continuity. IEE Proceedings Vision, Image and Signal Processing 147 (2), 185192. Liu, H., Zhao, F., 2012. An adaptive non local spatial fuzzy image segmentation algorithm. Intelligent Computing Technology, 373378. Mendel, J.M., 2001. Uncertain Rule-based Fuzzy Logic Systems: Introduction and New Direction. Prentice Hall, Upper Saddle River, NJ, USA. Mendel, J.M., 2007. Type-2 fuzzy sets and systems: an overview. IEEE Computational Intelligence Magazine 2 (1), 2029. Mendel, J.M., John, R.I.B., 2002. Type-2 fuzzy sets made simple. IEEE Transactions on Fuzzy Systems 10 (2), 117127. Mika, S., Ratsch, G., Weston, J., Scholkopf, B., Mullers, K.R. 1999. Fisher discriminant analysis with kernels. In: Proceedings of the 1999 IEEE Signal Processing Society Workshop, pp. 4148. Pham, D.L. 2002. Fuzzy clustering with spatial constraints. In: International Conference on Image Processing, pp. 6568. Shawe-Taylor, J., Cristianini, N., 2004. Kernel Methods for Pattern Analysis. Cambridge University Press. Shen, S., Sandham, W., Granat, M., Sterr, A., 2005. MRI fuzzy segmentation of brain tissue using neighborhood attraction with neural-network optimization. IEEE Transactions on Information Technology in Biomedicine 9 (3), 459467. Tam, S., Leung, C.C., Tsui, W.K. 2009. A robust segmentation method for the AFCMMRF model in noisy image. In: IEEE International Conference on Fuzzy Systems, pp. 379383. Wang, X.Y., Bu, J., 2010. A fast and robust image segmentation using FCM with spatial information. Digital Signal Processing 20 (4), 11731182. Wells III, W.M., Grimson, W.E.L., Kikinis, R., Jolesz, F.A., 1996. Adaptive segmentation of MRI data. IEEE Transactions on Medical Imaging 15 (4), 429 442. Worth, A.J., 2009. The Internet brain segmentation repository (IBSR). Xie, X.L., Beni, G., 1991. A validity measure for fuzzy clustering. IEEE Transactions on Pattern Analysis and Machine Intelligence 13 (8), 841847. Yang, Z., Chung, F.L., Shitong, W., 2009. Robust fuzzy clustering-based image segmentation. Applied Soft Computing 9 (1), 8084. Zadeh, L.A., 1975. The concept of a linguistic variable and its application to approximate reasoning I. Information Science 8 (3), 199249. Zhang, D.Q., Chen, S.C., 2004. A novel kernelized fuzzy fuzzy C-means algorithm (FCM) algorithm with application in medical image segmentation. Articial Intelligence in Medicine 32 (1), 3750. Zhang, Y., Chi, Z.X., Liu, X.D., Wang, X.H., 2007. A novel fuzzy compensation multiclass support vector machine. Applied Intelligence 27 (1), 2128.

can effectively improve the performance of algorithm on MR image segmentation. Experimental results indicate that the proposed IT2MFCM outperforms RFCM, IT2RFCM and ANLS_FIS for MR image segmentation in terms of suppressing noise and clustering validity functions. The benets of type-2 fuzzy theory have been demonstrated in many areas, but in the eld of fuzzy clustering it is still a novice. In fact FCM can be seen as a specic case of IT2FCM, and all the modied methods of FCM algorithms can be applied in a reasonably straightforward manner to improve the performance of these algorithms. Acknowledgements This work was supported by the National Science Foundation of China (Grant Nos. 51177137, 61134001) and the Fundamental Research Funds for the Central Universities (Grant No. SWJTU 11CX034). References
Ahmed, M.N., Yamany, S.M., Mohamed, N., Farag, A.A., Moriarty, T., 2002. A modied fuzzy C-means algorithm for bias eld estimation and segmentation of MRI data. IEEE Transactions on Medical Imaging 21 (3), 193199. Balafar, M.A., Ramli, A.R., Saripan, M.I., Mashohor, S., 2010. Review of brain MRI image segmentation methods. Articial Intelligence Review 33 (3), 261274. Bezdek, J.C., 1973. Cluster validity with fuzzy sets. Journal of Cybernetics 3 (3), 58 73. Bezdek, J.C., 1981. Pattern Recognition with Fuzzy Objective Function Algorithms. Kluwer Academic Publishers. Bezdek, J.C., Hall, L.O., Clarke, L.P., 1993. Review of MR image segmentation techniques using pattern recognition. Medical Physics 20, 1033. Birgani, P.M., Ashtiyani, M., Asadi, S. 2008. MRI segmentation using fuzzy C-means clustering algorithm basis neural network. In: Third International Conference on ICTTA, Damascus, pp. 15. Chuang, K.S., Tzeng, H.L., Chen, S., Wu, J., Chen, T.J., 2006. Fuzzy C-means clustering with spatial information for image segmentation. Computerized Medical Imaging and Graphics 30 (1), 915. Clark, M.C., Hall, L.O., Goldgof, D.B., Clarke, L.P., Velthuizen, R.P., Silbiger, M.S., 1994. MRI segmentation using fuzzy clustering techniques. IEEE Engineering in Medicine and Biology Magazine 13 (5), 730742. Clarke, L.P., Velthuizen, R.P., Camacho, M.A., Heine, J.J., Vaidyanathan, M., Hall, L.O., Thatcher, R.W., Silbiger, M.L., 1995. MRI segmentation: methods and applications. Magnetic Resonance Imaging 13 (3), 343368. Cristianini, N., Shawe-Taylor, J., 2000. An Introduction to Support Vector Machines and Other Kernel-based Learning Methods. Cambridge University Press. Fukuyama, Y., Sugeno, M. 1989. A new method of choosing the number of clusters for the fuzzy C-means method. In: Proceedings of Fifth Fuzzy Systems Symposium, pp. 247250.

Вам также может понравиться