Fuzzy C-Means Algorithm for Medical Image Segmentation

M.C.Jobin Christ Dr.R.M.S.Parvathi

Dept of Biomedical Engineering Principal
Adhiyamaan College of Engineering Sengunthar College of Engineering for Women
Hosur, India Tiruchengode, India
[email protected] [email protected]

Abstract-Clustering of data is a method by which large sets of (2) An additional parameter m has been introduced as a
data are grouped into clusters of smaller sets of similar data. weight exponent in fuzzy membership. The extended
Fuzzy c-means (FCM) clustering algorithm is one of the most objective function, denoted by Jm , is:
commonly used unsupervised clustering technique in the field
of medical imaging. Medical image segmentation refers to the
segmentation of known anatomic structures from medical
images. Fuzzy C-means (FCM) is a method of clustering which
allows one piece of data to belong to two or more clusters.
Fuzzy logic is a multi-valued logic derived from fuzzy set where P is fuzzy partition of dataset X formed by C1, C2,
theory. FCM is popularly used for soft segmentations like ….., Ck and k is number of clusters. The parameter m is
brain tissue model. And also FCM can provide better results weight that determines the degree to which partial members
than other clustering algorithms like KM, EM, and KNN. In of cluster affect the clustering result. Like hard c-means,
this paper we presented the medical image segmentation fuzzy c-means also tries to find good partition by searching
techniques based on various type of FCM algorithms. for prototype vi that minimizes the objective function Jm.
Unlike hard c-means, however, the fuzzy c-means algorithm
Keywords – FCM, Segmentation, Silhouette, Spatial FCM, also needs to search for membership function ȝci that
HMRF model. minimizes Jm. A constrained fuzzy partition {C1, C2, …..,
Ck } can be local minimum of the objective function Jm
I INTRODUCTION
only if the following conditions are satisfied:
1.1 Overview of Fuzzy C-Means
The Fuzzy C-Means algorithm is an unsupervised fuzzy
clustering algorithm. Conventional clustering algorithm
finds “hard partition” of a given dataset based on certain
criteria that evaluate the goodness of partition. By “hard
partition” we mean that each datum belong to exactly one
cluster of the partition. While the soft clustering algorithm
finds “soft partition” of a given dataset. In “soft partition”
datum can partially belong to multiple clusters. A soft
partition is not necessarily a fuzzy partition, since the input
space can be larger than the dataset. However, most soft
clustering algorithms do generate a soft partition that also
forms fuzzy partition. A type of soft clustering of special
interest is one that ensures membership degree of point x in
all clusters adding up to one, i.e., Few important points regarding the FCM algorithm:
It guarantees converge for m > 1.
It finds local minimum of the objective function Jm.
The result of applying FCM to a given dataset depends
not only upon the choice of parameter m and c, but also on
the choice of initial prototype.
A soft partition that satisfies this additional condition is
called a constrained soft partition. The Fuzzy C-Means 1.2 Medical Image Segmentation
algorithm, which is best known fuzzy clustering algorithm, Segmentation of images holds an important position in
produces constrained soft partition. In order to produce the area of image processing. It becomes more important
constrained soft partition, the objective function J1 of hard c- while typically dealing with medical images where pre-
means has been extended in two ways: surgery and post surgery decisions are required for the
(1) The fuzzy membership degree in cluster has been purpose of initiating and speeding up the recovery process.
incorporated in the formula. Computer aided detection of abnormal growth of tissues is
primarily motivated by the necessity of achieving maximum
___________________________________
978-1-4244 -8679-3/11/$26.00 ©2011 IEEE

33
possible accuracy. Manual segmentation of these abnormal approach for the segmentation of brain tumor. But the degree
tissues cannot be compared with modern day’s high speed of freedom for application of this method is limited too. The
computing machines which enable us to visually observe the algorithm demands an initial contour plan for extracting the
volume and location of unwanted tissues. Medical imaging region of interest. Therefore, like region growing approach,
types mostly are ultrasound images; X-ray computed this method is also semi automatic. Both of these methods
tomography, digital mammography, magnetic resonance are error sensitive because, an improper or false description
image (MRI), and so on. MRI images have good contrast in of initial plan and wrong selection of the seed image will
compare to computerized tomography (CT). Therefore, most lead to disastrous results. Statistical methods and fuzzy logic
of researches in medical image processing use MRI images. approaches seems to be reliable and are the best candidates
A well known segmentation problem within MRI is the task for the replacement of the above mentioned techniques.
of labeling voxels according to their tissue type which
include White Matter (WM), Grey Matter (GM), C] Supervised and Un-Supervised Segmentation Methods
Cerebrospinal Fluid (CSF) and sometimes pathological
tissues like tumor etc. Supervised and un-supervised methods for image processing
are frequently applied. Segmentation of volume using KNN
1.2.1 Various Segmentation Methods and both hard and fuzzy c-means clustering methods results
A] Segmentation by Thresholding that there appears to be enough data non-uniformity between
Thresholding method is frequently used for image slices to prevent satisfactory segmentation. Supervised
segmentation. This is simple and effective segmentation classification enables us to have sufficient known pixels to
method for images with different intensities. The technique generate representative parameters for each class of interest.
basically attempts for finding a threshold value, which In an un-supervised classification pre hand knowledge of
enables the classification of pixels into different categories. classes is not required. It usually employees some clustering
A major weakness of this segmentation mode is that: it algorithm for classifying an image data. KNN, ML and
generates only two classes. Therefore, this method fails to Parzen window classifiers are supervised classification
deal with multichannel images. Beside, it also ignores the algorithm. Whereas, un-supervised classification algorithm
spatial characteristics due to which an image becomes noise includes: K-Means, FCM, minimum distance, maximum
sensitive and undergoes intensity in-homogeneity problem, distance and hierarchical clustering etc.
which are expected to be found in MRI. Both these features
create the possibility for corrupting the histogram of the
image. For overcoming these problems various versions of II FCM FOR MR BRAIN IMAGE SEGMENTATION
thresholding technique have been introduced that segments Here we used some existing algoritms for MRI brain
medical images by using the information based on local image segmentation using different types of FCM
intensities and connectivity. Though this is a simple algorithms.
technique, still there are some factors that can complicate the
thresholding operation, for example, nonstationary and 2.1 Silhouette method
correlated noise, ambient illumination, busyness of gray
This method combines FCM, Silhouette Method and
levels within the object and its background, inadequate
Programming language R. This method gives a easiest way
contrast, and object size not commensurate with the scene. In
to find appropriate structure in the data of MRI. The data of
some methods, the objective function is constructed using
MRI T1, T2, PD images used as training set, and the new
the divergence function between the classes, the object and
method is performed to show the ability of finding structure
the background. The required threshold is found where this
in the data of training set. All the data of MRI T1, T2, PD
divergence function shows a global minimum.
combined together, and then a combined multi MRI
introduced using program in R. The steps are given below,
B] Region Growing Method 1. Coverts from image to data matrix, which is to be
Due to high reliability and accurate measurement of the classified.
dimensions and location of tumor, MRI is frequently used 2. Separate matrices interm of Images, if more than one
for observing brain pathologies. Previously, region growing image to be classified.
and shape based methods were heavily relied upon for
3. Construct a new data matrix with three columns to be
observing the brain pathologies. Bayes based region growing
algorithm that estimates parameters by studying structured, the columns are correspond to data of T1,
characteristics in local regions and constructs the Bayes T2, PD.
factor as a classifying criterion. The technique is not fully 4. Initialize the cluster centers at first time [n classes].
automatic, i.e. it requires user interaction for the selection of 5. Use FCM to make partition(clusters) into data matrix.
a seed and secondly the method fails in producing acceptable 6. If clusters Average width is > = 0.6 silhouette (stop
results in a natural image. It only works in homogeneous algorithm) or otherwise, repeat from step 4 until the
areas. Since this technique is noise sensitive, therefore, the clusters average width reach 0.6. (clusters average
extracted regions might have holes or even some width may be changed according to the problem).
discontinuities. Shape based method provides an alternative 7. Separate the elements of Matrix according to clusters.

34
8. Get segmented MR image using data matrices of Step
7.
2.2 Spatial FCM 2.3 HMRF-FCM
One of the important characteristics of an image is that The hidden Markov random field model (HMRF) to
neighboring pixels are highly correlated. In other words, model the image class labels, which takes into account the
these neighboring pixels possess similar feature values, and mutual influences of neighbouring sites formulated on the
the probability that they belong to the same cluster is great. basis of fuzzy clustering principle. In this method the
This spatial relationship is important in clustering, but it is explicit assumptions of the HMRF model is incorporated
not utilized in a standard FCM algorithm. To exploit the into fuzzy clustering procedure, an efficient fuzzy clustering-
spatial information, a spatial function is defined as type treatment is yielded. This combines the benefits from
the spatial coherency modelling capabilities of the HMRF
model, and the enhanced flexibility obtained by the fuzzy
clustering algorithm, i.e. fuzzy c-means algorithm (FCM).
The HMRF-FCM segmentation framework is validated with
where NB(xj) represents a square window centered on pixel noisy synthesis as well as brain MR images. This method is
xj in the spatial domain. A 5x5 window was used throughout formulated using an HMRF-FCM algorithm which offers an
this work. Just like the membership function, the spatial FCM-type treatment of the HMRF model. This combines the
function hij represents the probability that pixel xj belongs to benefits from the spatial coherency modelling capabilities of
ith cluster. The spatial function of a pixel for a cluster is the HMRF model, and the enhanced flexibility obtained by
large if the majority of its neighborhood belongs to the same the fuzzy clustering algorithm. The algorithm is given
clusters. The spatial function is incorporated into below,
membership function as follows: Derive an estimation of x(k) using

III Using x(k), compute the point wise prior probabilities of the
MRF ʌij (k) given by
where p and q are parameters to control the relative
importance of both functions. In a homogenous region, the
spatial functions simply fortify the original membership, and
the clustering result remains unchanged. However, for a
noisy pixel, this formula reduces the weighting of a noisy
cluster by the labels of its neighboring pixels. As a result, Compute the fuzzy membership functions rij(k) using
misclassified pixels from noisy regions or spurious blobs can
easily be corrected. The spatial FCM with parameter p and q
is denoted sFCMp,q. Note that sFCM1,0 is identical to the
conventional FCM.
The clustering is a two-pass process at each iteration. (k + 1)
The first pass is the same as that in standard FCM to IV Compute the estimator updates ȝij and ıij (k+1) by
calculate the membership function in the spectral domain. In
the second pass, the membership information of each pixel is
mapped to the spatial domain, and the spatial function is
computed from that. The FCM iteration proceeds with the
new membership that is incorporated with the spatial
function. The iteration is stopped when the maximum
difference between two cluster centers at two successive
iterations is less than a threshold (=0.02). After the
convergence, defuzzification is applied to assign each pixel
to a specific cluster for which the membership is maximal.
The spatial function modifies the membership function of a
pixel according to the membership statistics of its and
neighborhood. Such neighboring effect biases the solution
toward piecewise-homogeneous labeling. This technique
reduces the number of spurious blobs, and the segmented
images are more homogeneous. The sFCM algorithm with a
higher q parameter provides a better smoothing effect. The
possible disadvantages of using higher spatial weighting are
the blurring of some of the finer details.
V In case of convergence, i.e.

35
 REFERENCES
[1]. Balaji K, Juby N Zacharias’s Seminar at IIITM-K on 26/11/2007.
[2]. Chowdhury, M.H.; Little, W.D,"Image thresholding techniques" IEEE
Pacific Rim Conference on Communications, Computers, and Signal
Processing, 1995. Proceedings. 17-19 May 1995 Page(s):585 – 589.
Where Tc is the convergence threshold, exit; otherwise k [3]. Guillermo N. Abras and Virginia L. Ballarin,; "A Weighted K-means
Algorithm applied to Brain Tissue Classification", JCS&T Vol. 5 No.
= k +1 and return to 1. 3, October 2005.
[4]. Keh-Shih Chuang, Hong-Long Tzeng, Sharon Chen, Jay Wu, Tzong-
III CONCLUSION Jer Chen, “Fuzzy c-means clustering with spatial information for
In this work we have made a analysis about the image segmentation”, Computerized Medical Imaging and Graphics 30
(2006) ,9–15.
performance of three segmentation methods Silhouette [5]. Matthew C. Clark “Segmenting MRI Volumes of the Brain With
method, Spatial FCM, HMRF-FCM. The HMRF-FCM Knowledge- Based Clustering” MS Thesis, Department of Computer
algorithm combines the advantages of HMRF model, in Science and Engineering, University of South Florida, 1994.
terms of spatially correlated data clustering effectiveness, [6]. M. Masroor Ahmed, Dzulkifli Bin Mohamad, “Segmentation of Brain
and the increased flexibility of FCM-type method. It gives MR Images for Tumor Extraction by Combining K-means Clustering
less misclassification error, faster convergence and better and Perona-Malik Anisotropic Diffusion Model”, International Journal
of Image Processing, Volume (2) : Issue(1), Pages: 27-34.
accuracy. Silhouette method gave a easiest way to find [7]. M. Sezgin, B. Sankur, " Survey over image thresholding techniques
appropriate structure in the data of MRI. The data of MRI and quantitative performance evaluation" J. Electron. Imaging 13 (1)
T1,T2, PD images combined together, and then a combined (2004) 146-165.
multi MRI introduced using program in R. The training set [8]. Proceedings of the International MultiConference of Engineers and
originally had large dimension of data matrix, so the Computer Scientists 2010 Vol I, IMECS 2010, March 17 - 19, 2010,
Hong Kong.
program used to reduce the dimension of training set and the [9]. Rousseeuw, PJ, Silhouettes: a Graphical Aid to the Interpretation and
new algorithm method is applied to show the ability of the Validation of Cluster Analysis, Journal of Computational and Applied
method. Spatial FCM that incorporates the spatial Mathematics 20, 53-65, 1987.
information into the membership function to improve the [10]. Smita Pradhan, Dipti Patra, “Unsupervised Brain Magnetic Resonance
segmentation results. The membership functions of the Image Segmentation using HMRF-FCM framework”, 2009.
[11]. S.S. Mehta, C.R.Trivedi, N.S. Lingayat, “Identification and delineation
neighbors centered on a pixel in the spatial domain are
of QRS Complexes in electrocardiogram using fuzzy C-means
enumerated to obtain the cluster distribution statistics. These algorithm”, Journal of Theoretical and Applied Information
statistics are transformed into a weighting function and Technology 2009, Vol. 5, No 5, Pages :609-617.
incorporated into the membership function. This neighboring [12]. V. Deelport and D liesch, “Fuzzy C-means algorithm for code book
effect reduces the number of spurious blobs and biases the design in vector quantization”, Electronic letter 1994, Vol. 30, No.13.
solution toward piecewise homogeneous labeling. This [13]. www.icgst.com
[14]. Zhou, J.; Chan, K.L.; Chong, V.F.H.; Krishnan, S.M “Extraction of
method provided that the effect of noise in segmentation was Brain Tumor from MR Images Using One-Class Support Vector
considerably less with the new algorithm than with the Machine” 27th Annual International Conference of the Engineering in
conventional FCM. Medicine and Biology Society, 2005. IEEE-EMBS 2005, Page(s):6411
– 6414.

36