Intuitionistic Fuzzy Clustering Based Segmentation of Spine MR Images

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 04 Issue: 07 | July -2017 www.irjet.net p-ISSN: 2395-0072
© 2017, IRJET | Impact Factor value: 5.181 | ISO 9001:2008 Certified Journal | Page 790
Intuitionistic Fuzzy Clustering Based Segmentation of Spine MR Images
Binita Dutta1, Prof. J. V. Shinde2
1 P.G. Student, Department of Computer Engineering, Late G. N. Sapkal COE and Management, Nashik, India
2 Asst. Professor, Department of Computer Engineering, Late G. N. Sapkal COE and Management, Nashik, India
---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract - In medical field segmentation of image is very
difficult because of poor resolution and limited contrast.
Traditional segmentation techniques and thresholding
methods which are commonly used suffers from drawbacks
due to heavy dependenceonuserinteraction. Thesetechniques
cannot show the uncertainties present in an image. The
degradation of results takes place when the images are
effected by noise, outliers and artifacts. The main motive is to
overcome the drawbacks of traditional Fuzzy C Mean, the
objective functions of traditional Fuzzy C Mean approaches
has been modified and the spatial information has been
introduced in the objective function of the traditional Fuzzy C
Means. An automatedtechniquehasbeenproposedtoregulate
the initial values of the centroid of the cluster using
intuitionistic fuzzy clustering. A new complement function of
intuitionistic Fuzzy Clustering is mentioned for intuitionistic
fuzzy image representation which takes the inhomogeneity of
intensity and spine Magnetic Resonance image noise into
consideration. The evaluation of which is done using Dice
Coefficient, Jaccard coefficient, Recall and Precession.
Key Words: Intuitionistic fuzzy clustering, labelling, MRI,
Pre-processing, Vertebra segmentation.
1. INTRODUCTION
The basic building block of image analysis tool kit is the
image segmentation. Image segmentation in medical field is
tedious as the images are mostly affected by the noise.
Automation of thesegmentationprocessofmedical imagesis
a difficult task which is complex and does not contain any of
the simple linear characteristics features. Moreover the
results obtained by segmentation algorithm is severely
affected by the inhomogeneity of the intensity and the
volume effect especially when the MRI (magnetic resonance
images) is considered.Mostcomplex bonystructurein entire
human body is spine which bears the load of entire body and
consists of 26 irregular bones which are connected so as to
provide a curved and flexible structure. The backboneorthe
vertebral column has 5 major divisions and is 70cm long.
The first 7 vertebrae which is present in the neck make the
cervical vertebrae, the next 12 form the thoracic vertebrae
and the lumbar vertebrae are 5 in number and are the one
supporting the lower back. Sacrum is portion inferior to the
lumbar which articulates with hipbones of the pelvis. The
termination of entire vertebral column is done by the tiny
coccyx. The extension of spine is done by the help of
intervertebral disc whose function is that of a shock
absorber. These discs are thickest in the cervical andlumbar
region making these regions more flexible. Due to wear and
tear with age degeneration of discisa commonphenomenon
which results in back pain. Few types of disc wear tear are
Herniate disc, degenerated disc and spinal stenosis. The
spine Magnetic Resonance imagingisidentifiedbyT1andT2
weighted images. The water contents appears bright in T2
weighted image (in medical terms its hyper Intense which
can be clearly seen in the spinal canal) and vice versa (i.e.
hypo Intense) in that of T1 weighted images.
The signs degeneration of bone marrow can be clearly
detected by a MR with high spatial resolution where water
protons and fat are found in abundance. Even though
degenerative changes occurring in the spine which are
imaged using a radiological equipment are biological
phenomena, certain irrelevant processes are also captured.
These are the artifacts which are produced by the intensity
inhomogeneity. The process of segmentation is highly
affected by the MR image complexities. As a result of which
accurate diagnosis remains a challenge without manual
interference in segmenting the vertebral features.Diagnosis
using a computer therefore requires a Robust automated
segmentation of vertebrae from MR images.
There are many automated segmentationtechniquesamong
which clustering techniques are more predominantly used.
Clustering techniques are classified into two types first one
is hard clustering i.e. k means and second one is soft
clustering i.e. Fuzzy C-Means Clustering. Various
segmentation techniques of MR image of vertebrae using
fuzzy c mean (FCM) clustering are used to identify and label
the individual vertebral structures. But FCM clustering fails
to deal with local spatial property of images which results in
strong noise sensibility. As it’s considered that the
uncertainty and unknown noise are always included in
medical images, this leads to degradation in segmentation
quality generally. [1]
The approaches of Intuitionistic fuzzy set (IFS) is also used
for segmentation of image. IFS takes non- membership and
hesitancy values into account along with membership value
and these sets serve useful to deal with uncertainty and
vagueness in the image pixel intensities .FCM based
clustering methods are mostly dependent on selection of
parameter, noise sensitivity and their convergence mainly
depends on cluster centroid initialization . Due to the

limitations there can be stucking of the algorithm in a local
optimum which in turn reduces the accuracy of
segmentation. In this work, an intuitionistic fuzzy clustering
(IFC) algorithm (which is rough set based) is proposed for
the segmentation of spine MR images. The main
contributions of present work is: To achieve best possible
result than the existing system by applying Intuitionistic
Fuzzy C Mean algorithm and help the medical specialists to
diagnose deformities.
The work in this paper is divided in two parts. 1) Pre-
processing 2) Segmentation and post-processing. Pre-
processing is done by applying filters i.e. anisotropic
diffusion filter which helps to remove the acquisition noise
which is present in MRI image. Thereafter, thesegmentation
is been done using intuitionistic Fuzzy C Means clustering
algorithm on the pre-processed mages followed by
morphological processing of erosion and dilation. Then,
components are labelled and a colouredvertebraeisgot. The
paper organisation is as follows. Section 2 shows the
previous work done in this field. The steps of the algorithm
and the flow diagram is described in section 3 and section 4
presents experimental results showing results of tested
mages. Finally, Section 5 shows the conclusion part.
2. RELATED WORK
Global thresholding, thresholding at multilevel and
supervised techniques of clustering are the commonly used
segmentation methods. The intensity level is resolved from
the grey scale image histogram in intensity thresholding. In
medical images the intensity distributions are random
especially in MRI images which results in failure of global
thresholding methods due to absence of resolving the
optimal threshold. Disadvantage of spatial uncertaintytakes
place in intensity thresholdingmethodsastheinformationof
pixel location is overseen [3]. For identifying the contour
boundaries of the ROI, an edge detection technique can be
implemented. The probability of these lines being continues
is minimum. And these methodsusuallyneedexpensive post
processing to obtain objects without any holes.
Segmentation of MRI vertebral body using a graph based
approach is proposed in [4]. A Nystrom approximation is
used here to decrease the processing speed of NCut
(normalized cut) method. Above methods maximize the
dissimilarity between two subsets and similarity within a
subset in an image. This problem as proposed by Malik and
Shi, minimizes to solve the generalized Eigen vector which
requires a diagonal matrix and a weighted matrix that is
initiated using the feature image strength. Nystrom
approximation simplifies the computationally intensive
calculation of the Eigen vectors which uses few random
samples to calculate the diagonal matrix and the solution is
extended to the complete set. Here anisotropic diffusion is
used for the pre-processing which gives a coil corrected
output. A semiautomatic classificationofspinedisorders disc
degeneration, spinal stenosis and disc herniation is given in
[5]. The authors have represented MRI response in T2
weighted images for classification of disc degeneration. The
level of degeneration s denoted by reduction in mean
intensity and Pfirmann scale is used to classify it.
Segmentation of disc is done using gradient vector flow
(GVF) on an image after pre-processing which isfollowed by
skeletonization. The length is marked by drawing a vertical
axis from the mid-point. The abnormalities in disc is
classified by changes in intensity and length. About 90% of
herniation [6] happen in Lumbar region, which leads to the
processing of discs located in the lumbosacral area.
Evaluation for disc herniation, is done by extracting the
contour from the axial slice. The herniated discs can be
either classified as focal based (extrusion) or broad based
(protrusion). An automatic segmentationmethod(quasi)for
intervertebral disc (IVD) is given in [7].
The only involvement for selection of initial point is to be
deformed by snakes for evolving an elliptical contours. FCM
is used for further refinement of the ROI(Regionofinterest).
An equation is evolved using a new energy term called
geometric energy, which is included withactivecontourthat
shapes the disc and controls it to a certain limit. In this case,
since IVD appears to have an elliptical structure,so evolution
of contour is reduced when it goes beyond the mentioned
shape. Here the Adaboost classifier is used for classification
of discs and the root of mean square error (RMSE) and Dice
similarity coefficient is used for performing validation
between the segmented disc and the ground truth. The
thresholding is extended by region growing methods by
integrating it with the connectivity by an intensitysimilarity
measure. An initial origin position is assumed in these
methods and the intensity column is expanded using
connected neighborhood, over the neighboring region.
However, these are highly affected by the initial origin
points and noise. Among all others FCM algorithm is found
out to be more effective and efficient with significant
benefits for classification-based segmentation method [8].
FCM algorithm allows pixels dependencies with different
membership degrees with multiple clusters, unlike the hard
clustering methods, like K-means algorithm,that exclusively
assign pixels to only one cluster and is found out to be more
worthy in real applications.Introductionofnon-membership
is done along with its values in segmentation of MR images
of brain using intuitionistic fuzzy clustering (IFC) [9] along
with that of membership (MF). The degree of hesitation for
each pixel is a significant part in heuristic based
segmentation. Several general grey scale images are
presented in [10] where the merging of membership
functions takes place using IFC, to show the uncertainty in
selecting the best MF. Detection of Modic changes (MC) can
remarkably present the lesion detection on the bone
marrow. Disc herniation [2] can also be associated with it. It
is seen that higher the severity of Modic changes, higher is
the level of herniation. The burden of manual differentiation

of Modic Changes levels for a radiologistcanbedecreasedby
automating the process of classificationanddetection.Italso
provides an accurate percentage of degeneration based on
intensity change in signal contrasts to the normal level.
Though there are articles on MR segmentation of vertebral
bodies [11][12][13] , due to the complexities involved in
delineating VB’s, the paper focuses on utilizing the
intuitionistic fuzzy concept for perfectly isolating them. The
paper focuses on implementation of IFC based method for
segmentation of vertebrae withfurtherpost-processing. The
vertebrae are also labelled to reduce the burden of
radiologists to classify the degenerations involved.
3. SYSTEM ARCHITECTURE
The proposed system is schematically showninfigure1. The
MR Spine images of patients have been collected. Total 14
MRI images of spine is been used with a dimension of 512 x
512 from Samarth diagnostics centre. The T1 weighted
images, are used as the initial dataset for the proposed
algorithm. Firstly, the image is smoothed using the edge
preserving anisotropic diffusion filter. The MRI images are
prone to various types of noise among which acquisition
noise is common, to remove this noise anisotropic diffusion
filter has been used. This pre-processing serves both the
purpose of removing inhomogeneity and as contrast
enhancer as well. After pre-processing the pre-processed
images are segmented using Intuitionistic Fuzzy clustering
algorithm followed by some of themorphological operations
which are executed for extracting the vertebral bodies (VB)
from the clustered output. The hole filling is the preliminary
step followed by removing islands using an erosion. Finally
the vertebrae are got which are completely devoid of inter-
vertebral discs in between them and other muscles present
near the vertebrae. After which the colour code is been
obtained by using an area metric for labelling of the
vertebrae.
Fig 1. : System Architecture Diagram
In the existing system after pre-processing segmentation is
done by using Fuzzy C Mean clustering. Fuzzy C Mean
Clustering (FCM) is the most successful algorithm for data
clustering but the main drawback of FCM for segmentation
of image is that its objective function does not take into
consideration anyspatial dependenceamongpixelsofimage.
Secondly, the membership functionofFCMismostlydecided
by the Euclidean distance which measuresthepixel intensity
similarity and the center of the cluster. Closer intensity
values to the cluster center is defined by the higher
membership. Hence the sensitivity to noiseofa membership
function is high. In MR image with noise and intensity
inhomogeneity, the performance of FCM methodsdecreases
and results in improper segmentation. To overcome these
issues we have used Euclidean distance based on
intuitionistic fuzzy set,using thisintuitionisticfuzzydistance
measure, partition matrix and cluster centroid are updated.
While updating membership, membership function, non-
membership and hesitation of pixel value are considered
which results into compensation for the inhomogeneity in
the intensity and noise in spine MR images and results in
more appropriate tissue segmentation. This provide us
better results when compared with the existing system.
4. EXPERIMENTAL RESULTS
Figures shows the results of pre-processing of the sagittal
view of the Spine MRI image and segmentation using
Intuitionistic Fuzzy C Mean Clustering algorithm. Figs. 2 (a)
shows the original MRI Spine image. (b) is the image
obtained by applying Pre-processing. Image shown in fig 2
(c) is the image obtained by Fuzzy C Means algorithm (d) is
the lumbar portion of the spine vertebra obtained after
morphological processing (e) is the image of lumbarportion
of the vertebrae obtained by applying Intuitionistic Fuzzy C
Mean Clustering algorithm. Graphical representation of the
result of image P9 of Dice coefficient, Jaccard Coefficient ,
Precision and Recall is shown in fig 3(a),( b) and (c)
respectively. The results of the proposed system is found to
be more efficient than the existing system.
(a) (b)
(c) (d)

(e) (f)
Fig. 2. Pre-processing and segmentation of P9 (a) Original
image (b) Image after applying pre-processing(c) Image
after applying FCM clustering (d) Image of vertebrae after
applying FCM (e) ) Image of vertebrae after applying IFC (f)
Coloured vertebrae or labelled vertebrae.
(a)
(b)
(c)
Fig. 3. Graphical representation of P9 (a) Dice Coefficient
(b) Jaccard coefficient (c) Precision and Recall
In the proposed system we have used the datasetofa clinical
centre named Samarth diagnostic centre for Image
segmentation and since the existing system and the
proposed system both deals finding out the deformities in
the lumbar part of spine so the modality that is been chosen
is Magnetic Resonance Imaging or MRI. T1 weighted MRI
images are considered as the bestimagestofindoutdiseases
related to bones. The results of both the existing system and
proposed system are been compared by using Dice
Coefficients, Jaccard Coefficient, Recall and Precision as
metrics for evaluation. The results of which are been shown
below in table 1.
Table -1: Comparative Results of FCM and IFC
Comparative Results for IFC and FCM
Image
Dice
Coefficient
Jaccard
Coefficien
t
Recall Precision
FCM IFC FCM IFC FCM IFC FCM IFC
T1_sag_P
1
0.91 0.95 0.91 0.9
3
90.9
1
93.1
7
55.6
7
63.4
5
T1_sag_P
2
0.81 0.85 0.71 0.7
7
70.8
3
77.4
1
71.8
5
88.1
6
T1_sag_P
3
0.86 0.90 0.81 0.8
5
81.2
0
85.4
5
60.0
6
71.8
7
T1_sag_P
4
0.86 0.92 0.83 0.8
7
82.8
9
86.7
7
59.9
0
68.2
5
T1_sag_P
5
0.75 0.82 0.63 0.7
5
63.1
0
75.1
5
62.6
0
70.1
0
T1_sag_P
6
0.85 0.89 0.76 0.8
1
76.3
0
80.7
5
65.7
2
73.2
1
T1_sag_P
7
0.84 0.87 0.74 0.7
9
73.8
7
79.0
2
61.4
8
74.6
0
T1_sag_P
8
0.84 0.90 0.78 0.8
2
78.0
9
81.7
8
57.8
9
63.4
1
T1_sag_P
9
0.85 0.89 0.81 0.8
6
80.5
5
85.8
6
60.6
8
71.3
2
T1_sag_P
10
0.68 0.71 0.51 0.7
0
51.0
0
69.6
9
58.1
6
66.5
7
T1_sag_P
11
0.81 0.82 0.63 0.7
8
63.2
5
77.7
8
80.2
5
87.7
5
T1_sag_P
12
0.63 0.69 0.46 0.6
6
46.0
7
65.5
0
70.5
3
77.7
7
T1_sag_P
13
0.87 0.92 0.78 0.8
0
78.4
0
80.0
8
57.2
2
62.5
8
T1_sag_P
14
0.71 0.79 0.53 0.7
7
53.0
9
77.2
7
59.1
3
68.8
8

5. CONCLUSIONS
In this paper an IFC algorithm followed by morphological
operations and labeling has been presented for
segmentation of vertebra from spine MR images, which is
compared to the simple FCM clustering. Thestudyincludesa
dataset of 14 patients of age 29 to 65 of T1-weighted MR
images of spine. It is seen that the proposed system proved
to be more efficient than the existing system and provides
improved segmentation results as compared to the existing
system.
ACKNOWLEDGEMENT
I would also like to thank Samarth Diagnostics Centre for
their support for sharing the data of the patients.
REFERENCES
[1] Jiyo Athertya,Saravana Kumar Gurunathan ”Fuzzy
Clustering Based Segmentation Of Vertebrae in T1-
Weighted Spinal MR Images,”International Journal of
Fuzzy Logic Systems (IJFLS) Vol.6, No.2,April 2016.
[2] H. B. Albert, P. Kjaer, T. S. Jensen, J. S. Sorensen, T.
Bendix, and C. Manniche, “Modic changes, possible
causes and relation to low back pain,” Med.
Hypotheses, vol. 70, no. 2, pp. 361–368, 2008.Yassin M.
Y. Hasan and Lina J. Karam, “Morphological Text
Extraction from Images”, IEEE Transactions On Image
Processing, vol. 9, No. 11, 2000
[3] S. R. Kannan, S. Ramathilagam, a. Sathya, and R.
Pandiyarajan, “Effective fuzzy c-means based kernel
function in segmenting medical images,” Comput. Biol.
Med., vol. 40, no. 6, pp. 572–579, 2010.
[4] J. Carballido-Gamio, S. J. Belongie, and S. Majumdar,
“Normalized cuts in 3-D for spinal MRI segmentation.,”
IEEE Trans. Med. Imaging, vol.23, no. 1, pp. 36–44, Jan.
2004
[5] S. Ruiz-españa and E. Arana, “Semiautomatic computer-
aided classification of degenerative lumbar spine
disease in magnetic resonance imaging,” Comput.
Biol. Med., 2015.
[6] C. Dora, M. R. Schmid, A. Elfering, M. Zanetti, J. Hodler,
and N. Boos, “Lumbar disk herniation: do MR imaging
findings predict recurrence aftersurgical diskectomy?,”
Radiology, vol. 235, no. 2, pp. 562–567, 2005.
[7] I. Castro-Mateos, J. M. Pozo, A. Lazary, and A. F. Frangi,
“2D segmentation of intervertebral discs and its degree
of degeneration from T2-weighted magnetic resonance
images,” Spie Med. Imaging Comput. Aided Diagnosis,
vol. 9035, no. figure 1, p. 903517, 2014.
[8] T. Chaira, “A novel intuitionistic fuzzy C means
clustering algorithm and its application to medical
images,” Appl. Soft Comput. J., vol. 11, no. 2, pp. 1711–
1717, 2011.
[9] Y. K. Dubey and M. M. Mushrif, “Segmentation of brain
MR images using intuitionistic fuzzy clustering
algorithm,” Proc. Eighth Indian Conf. Comput. Vision,
Graph. Image Process. – ICVGIP ’12, pp. 1–6, 2012.
[10] V. P. Ananthi, P. Balasubramaniam, and C. P. Lim,
“Segmentation of grayscaleimagebasedonintuitionistic
fuzzy sets constructed from several membership
functions,” Pattern Recognit., vol. 47, no. 12, pp. 3870–
3880, 2014.
[11] I. Ben Ayed, K. Punithakumar, and R. Minhas, “Vertebral
Body Segmentation in MRI via Convex Relaxation and
Distributipn Matching,” MICCAI, pp. 520–527, 2012.
[12] P. D. Barbieri, G. V. Pedrosa, A. J. M. Traina, and M. H.
Nogueira-Barbosa, “Vertebral Body Segmentation of
Spine MR Images Using Superpixels,” 2015 IEEE 28th
Int. Symp. Comput. Med. Syst., pp. 44–49, 2015.
[13] M. Lootus, T. Kadir, and A. Zisserman, “Vertebrae
Detection and Labelling in Lumbar MR Images,” Lect.
Notes Comput. Vis. Biomech., vol.17,pp.219–230,2014.
[14] C. kong chui Bing Nan li, “Integrating spatial fuzzy
clustering with level set methodsforautomated medical
image segmentation,” Comput. Biol. Med., 2011.
[15] C. I. Nedeljkovic, “Image Classification Based on Fuzzy
Logic,” pp. 1–6, 2004.
[16] M. Gong, Y. Liang, J. Shi, W. Ma, and J. Ma, “Fuzzy C-
means clustering with local information and kernel
metric for image segmentation,” IEEE Trans. Image
Process., vol. 22, no. 2, pp. 573–584, 2013.

Intuitionistic Fuzzy Clustering Based Segmentation of Spine MR Images

More Related Content

What's hot (18)

Similar to Intuitionistic Fuzzy Clustering Based Segmentation of Spine MR Images (20)

More from IRJET Journal (20)

Recently uploaded (20)

Intuitionistic Fuzzy Clustering Based Segmentation of Spine MR Images