Sie sind auf Seite 1von 6

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/322682689

Nonsubsampled Contourlet Transform based Expectation Maximization


Method with Adaptive Mean Shift for Automatic Segmentation of MR Brain
Images

Article  in  International Journal of Applied Engineering Research · January 2018

CITATIONS READS
0 20

1 author:

Meena Prakash R.
Sethu Institute of Technology
18 PUBLICATIONS   26 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

MR Brains Segmentation View project

All content following this page was uploaded by Meena Prakash R. on 24 January 2018.

The user has requested enhancement of the downloaded file.


2014 International Conference on Electronics and Communication Systems (lCECS -2014), Feb.13 -14, 2014, Coimbatore,
INDIA

Nonsubsampled Contourlet Transform based


Expectation Maximization Method with Adaptive
Mean Shift for Automatic Segmentation of
MR Brain Images

R.Meena Prakash R.Shantha Selva Kumari


Department of ECE Department of ECE
P.S.R.Engineering College Mepco Schlenk Engineering College
Sivakasi, India Sivakasi, India
meenaprakash73@gmail.com r sskp@yahoo.com

Abstract- An automatic method of MR brain image An image can be modeled as the union of c homogenous
segmentation into three classes White Matter, Gray Matter and regions U�=l Ak where each homogeneous region is specified
Cerebrospinal fluid is presented. The intensity non uniformity or by Ak (x, y) = bkPk + nk where Pk represents signal
bias field and noise present in the MR brain images pose major intensity, bk represents multiplicative bias field and nk
limitations to the accuracy of traditional EM segmentation
represents additive, zero mean random noise component. The
algorithm. To overcome these drawbacks, Nonsubsampled
difficulties associated with MR brain image segmentation
Contourlet Transform low pass filter is used as preprocessing
arise from three main factors which are (a) Partial
step. Since the bias field is found to be smoothly varying, it is
Volume Effect (b) Intensity Inhomogeneity and (c) Noise.
proposed and applied that the GMM is preserved locally in the
Partial Volume Effects occur because of multiple pixels
image blocks of appropriate size. Hence the image is divided into
blocks and then EM segmentation is applied. To ensure
contributing to a single pixel resulting in blurring of intensity
smoothness among the segmentation output of the successive
across boundaries. These effects occur due to the limited
blocks, an adaptive mean shift followed by pixel stretching is resolution of the imaging device and due to these, finer
proposed. The algorithm is evaluated on TI weighted simulated anatomical details are lost in the image. The Intensity Non­
brain MR images and 20 normal Tl-weighted 3-D brain MR Uniformity (INU) artifact arises due to the in-homogeneity in
images from IBSR database. Results ensure that there is around the magnetic field and results in an unwanted low frequency
4% improvement in accuracy in Gray Matter Segmentation for bias term modulating the signal. All MR images are affected
3-D brain MR images compared to fuzzy local Gaussian mixture by random noise. The noise comes from the stray currents in
model. Also the computational costs are reduced in this method. the detector coil due to the fluctuating magnetic fields arising
Keywords- MR brain image segmentation, Expectation Maxim­ from random ionic currents in the body, or the thermal
ization, Nonsubsampled Contourlet Transform. Adaptive mean. fluctuations in the detector coil itself. The most prevalent
method for segmentation of MR brain images is the
I. INTRODUCTION
Expectation Maximization method. But it does not take into
The advances in imaging technology aid the doctors in account the above said drawbacks specific to MR image.
analyzing the diseases and in proper treatment planning. Koen Van Leemput et al. developed the model based
Magnetic Resonance Imaging is more commonly deployed for method for automated classification of MR brain images[l],
imaging the brain to categorize the widespread brain diseases. [2], [3].The method applies an iterative Expectation Maxi­
MR imaging technology is non-invasive and it shows good mization algorithm that interleaves pixel classification with
soft tissue contrast in brain images. Hence it is extensively estimation of class distribution and bias field parameters,
used for imaging the human brain to identify the diseases. The improving the likelihood of the parameters at each iteration.
outcome of Segmentation of brain image is classification of Benoit Scherrer et al. proposed distributed local MRF models
tissues into one of the three major classes: White Matter, Gray in which tissue segmentation is done by partitioning the
Matter and Cerebro-Spinal fluid. The quantification of these volume into sub volumes and local MRFs are estimated in
tissue volumes help to assess the progression of brain diseases, cooperation with their neighbours [4].
such as Alzheimer's disease, epilepsy, multiple sclerosis and Thanh Minh Nguyen and Q. M. Jonathan Wu proposed an
schizophrenia. Manual segmentation of brain tissues is time algorithm which incorporates the spatial relationship between
consuming and also it leads to inter and intra observer neighboring pixels by replacing each pixel value in an image
variability. Hence automated segmentation methods are in with the average value of its neighbors including itself [5].
research focus.
Zexuan Ji, Yong Xia et al. proposed the fuzzy local
Gaussian mixture model in which a truncated Gaussian kernel

[IEEE copyright notice]


2014 International Conference on Electronics and Communication Systems (lCECS -2014), Feb.I3 -14, 2014, Coimbatore,
INDIA
function is used to impose the spatial constraint and fuzzy
memberships are employed to balance the contribution of each p(y / (}) n P(Yi (})
= /
GMM [6]. With reference to the works in the literature, (3)
different approaches have been adapted to overcome the (y, I r = Bj ) (r = )
P .
i }, .
P , }
effects of bias field and noise in segmentation accuracy. From (r' = }. I B) =
P 1 Yi' . (4)
I r/ = } , B) p r/ = } )
LP(Y/ (
.
[4], it is inferred that the Gaussian Mixture Model is preserved
in the local image data or the sub volwnes of the image since }
the bias field is smoothly varying over the entire image. The expressions for mean and variance are
In this paper, block based EM segmentation is employed to LYiP(r, = j / Yi'(})
remove the effect of bias field. To ensure smoothness among
Ji}
the segmentation output of the neighbouring blocks, the
method of adaptive mean shift and pixel stretching is proposed
LP(ri = j / Yi.(})
(5)
which involves fewer computations. Further, as the standard
GMM does not take into account the spatial correlation LP(ri = j / Y,.(})(Yi - Jij)2
between the neighboring pixels, Nonsubsampled Contourlet
transform is employed as preprocessing which shows good
performance in segmentation accuracy with the presence of
LP(r/ = j / Yi'(})
(6)
noise and INU when compared to the other methods.
Equation 4 performs classification while 5 and 6 are
The chapters are organized as follows. In Chapter II, parameter estimates. The algorithm fills in the missing data
background of Gaussian Mixture Model, EM segmentation during step 1 and then finds the parameters that maximize the
and Nonsubsampled Contourlet transform is given. In chapter likelihood for the complete data during step 2. The likelihood
III, the proposed method is explained along with the metrics of is guaranteed to increase at each iteration. The drawbacks
evaluation of the algorithm. In Chapter IV, the results are observed in the method are that it treats each pixel as
discussed and compared with the other state of the art independent and it does not take into account the spatial
techniques. relationship between pixels.

II. BACKGROUND B. N onsubsampled Contourlet Transform

A. Gaussian Mixture Model and EM Segmentation


The EM algorithm is a general technique for finding
Maximum Likelihood parameter estimates in problems with
hidden data [1], [2], [3]. The EM algorithm consists of two
steps E step and M step. The E step computes expectation of
the unobserved data. The M step computes Maximwn Like­
lihood estimates of the unknown parameters. The process is
repeated until it converges. The likelihood increases with each -t-+-i__H-+--+ � WI
iteration and finally converges. For segmentation problem, the
observed data are the signal intensities, the missing data are
the classes to which the signal intensities belong, and the
parameters are mean and variance of the classes. Each class is
modeled b1' a normal distribution G(J(Y-Il) with mean 11 and Figure 1. Nonsubsampled Contourlet Transform
variance (J .The probability density that class j has generated
the voxel value Yi at position i is Nonsubsampled Contourlet transform exhibits multidirec­
tional and multiresolution properties. Images are analyzed at
(1) different resolutions from coarse to fine using multiresolution
property. Multidirectional property allows the representation
to contain basis elements oriented at different directions. A
with ri £{j/j=l...K} the tissue class at position i and double filter bank structure is used in the Nonsubsampled
8j = {llj,(Jj} the distribution parameters for class j. Defining Contourlet Transform. The Nonsubsampled Contourlet
9= {9j} as the model parameters, the overall probability Transform has the important characteristic that it is shift
density for Yi is invariant since there are no up samplers and down samplers in
the filter structure. Hence the position of each pixel in the sub
band corresponds to the original location in the image.
(2)
The NSCT is constructed by combining nonsubsampled
pyramids and nonsubsampled directional filter banks as shown
which is a mixture of normal distributions. Since all the voxel in fig (1). The Nonsubsampled pyramid structure results the
intensities are assumed to be statistically independent, the multi resolution property. In this paper, the Nonsubsampled
probability density for the image y given the model is pyramid decomposition is used as preprocessing to remove the
effect of noise and INU in segmentation.
2014 International Conference on Electronics and Communication Systems (lCECS -2014), Feb.I3 -14, 2014, Coimbatore,
INDIA
III. METHOD B. Nonsubsampled Contourlet Transform Low Pass Filter
The MR brain image is corrupted with noise which occurs
from stray currents in the detector coil. The presence of noise
in MR brain image affects the segmentation process.
Preprocessing techniques like homomorphic filtering methods
may be employed to remove the noise in MR brain images. In
NSCT low pass filter
I n p ut MR Brain
this paper, the NSCT low pass decomposition is used as
Image with INU preprocessing to eliminate the effect of noise and INU in
a nd noise segmentation. The features of NSCT which make it better
suited for removing noise artifacts in the segmentation process
is discussed. The NSCT is constructed from two filter
structures.

1. A nonsubsampled pyramid structure which


ensures the multiscale property.
MR Brain Image divided
EM Segmentation 2. A nonsubsampled directional filter bank structure
into blocks
which gives the directionality.
A Nonsubsampled filter bank has no down sampling or up
sampling and hence it is shift invariant. The perfect recon­
struction condition is given as Ho(z)Go(z)+HJ(z)GJ(z) 1. In
=

this paper, the multi scale decomposition using Nonsubsam­


pled pyramid is used to eliminate the effect of noise in
Adaptive Mean shift and pixel
Segmented output
segmentation. Multi scale decomposition is achieved by
stretching iterated Nonsubsampled filter banks. The standard GMM
considers each pixel as independent and does not take into
account the spatial correlation between the neighboring pixels.
Figure 2. Proposed Method
The efficient filter bank of the Nonsubsampled Contourlet
The algorithm is summarized in the following steps. transform and its shift invariant property make it an efficient
frame work for image processing applications.[7] The
1. The brain region is extracted from the neighbouring
proposed algorithm effectively utilizes the Nonsubsampled
pixels using mask and morphological operations.
pyramid decompositon to incorporate the spatial correlation
2. The Nonsubsampled Contourlet transform is applied
between the neighbouring pixels. Depending upon the level of
to decompose the image into low pass sub band and
noise and RF inhomogeneity present, the level of
band pass directional sub bands.
decomposition may be chosen.
3. The low pass sub band output of NSCT is taken and
the image is divided into square blocks (2D C. Block Based Expectation Maximization Segmentation
representation) or cubical volumes (3D The distribution of the tissue classes in MR brain image is
representation) of appropriate size. commonly modeled as Gaussian Mixture Model. The
4. Then the Expectation Maximization algorithm is parameters of GMM are estimated via Expectation
applied to each block to segment the block into Maximization algorithm. The bias field or Intensity Non
different classes with varying pixel intervals. Uniformity Artifact in MR brain images refers to the slowly
5. Then adaptive mean shift and pixel stretching is varying multiplicative field that occurs due to the
applied to the means of each block to ensure inhomogeneities in the magnetic fields of the MRI machine.
consistency among the segmentation output of The input image is divided into blocks since the GMM is
neighbouring pixels. preserved locally. If the size of the block is chosen to be large,
6. Finally the tissues are labeled as Gray Matter, White the effect of bias field cannot be eliminated from
Matter and Cerebrospinal fluid depending on the segmentation. If the size of the block is chosen to be small, it
class intervals of the means. results in poor model estimation. Hence the block size is to be
A. Brain Extraction chosen carefully. In this paper, for the IBSR data volumes of
real images, a block size of 64 x 64 pixels is chosen. Then
The brain extraction is an important step before
Expectation Maximization segmentation is applied to each
segmentation is to be done. The pixels lying outside the brain
block which is explained in the following steps.
contour will also have similar intensities with that of the brain
region. The segmentation time will be reduced if segmentation 1. The image block of size 64x64 and the number of
is restricted to the brain region alone. First a brain mask is classes (here selected best as 5 for MR brain image)
derived from the image itself by threshold operation. Then are given as inputs to the algorithm.
morphological operations erode and clean are used to extract 2. The initial parameters - mean, variance and
the brain region.
proportion of the classes are initialized assuming
equal distribution for all the classes.
2014 International Conference on Electronics and Communication Systems (lCECS -2014), Feb.13 -14, 2014, Coimbatore,
INDIA
3. The E step and the M steps are performed iteratively = . v�e .
until convergence. The E step assigns probabilities to TeCi) vJ+vi-vJe (8)
each label at each pixel. The M step updates the
parameter estimates namely mean, variance and vde
is the nwnber of voxels that are assigned to tissue i by both
proportion using Maximwn Likelihood estimation. the ground truth and the automated algorithm.
4. Each block of the image is segmented into five vd
is the number of voxels that are assigned to tissue i by
classes with five means obtained. automated algorithm
l/�/ is the number of voxels that are assigned to tissue i by
D. Adaptive Mean Shift and Pixel stretching Algorithm ground truth.

IV. RESULTS AND DISCUSSION


In this section, the performance of the algorithm on 3D

I I
Set mean offifth class mu (5) =224 simulated and real data sets is presented. The system configu­
ration used is Intel Core 2 Duo CPU @2.53GHz with 1.98GB

! of RAM. The algorithm is executed using MATLAB.


A. Segmentation of MR Brain Images from simulated
Adjust the means of other classes as
mu(i)=abs(mu(i»*224/mu(5) brain database
EM segmentation with adaptive mean shift and pixel

!
stretching is implemented in Tl-weighted simulated brain MR
images selected from the brain web simulated brain database
For every mu ( i), classify th e p ixe ls as (Brain Web). The method is validated on simulated images
Ifmu(i)<=IO, pixel value=O with 40% inhomogeneity and 9% noise, 181 x 217 x 181
If IO<mu(i)<=70, pixel value=128 dimension 1 x 1 x 1 mm3 spacing. For these images, the initial
If 70<mu(i)<=220, pixel value= 192 step of division into blocks is eliminated. Also, the mean
If220<mu(i)<=255, pixel value=254 intervals are selected as 10 to 128, 128 to 180 and 180 to 255.
The ground truth for the Brain Web dataset is the phantom
atlas used to generate the simulated scans. The ground truth
for the Tlw images is known for comparisons. The
Figure.3 Adaptive Mean Shift and Pixel Stretching segmentation results are shown in figure 4. The qualitative and
quantitative results show that there is significant improvement
For every block, do the steps described in fig. (3). The EM in segmentation accuracy of about 8.5% due to the
algorithm considers each pixel as independent and does not employment of NSCT low pass filter as preprocessing.
take into account the spatial correlation between the
neighbouring pixels. To overcome the bias field which is a (a) (b)
slowly varying component, the image is divided into blocks
and then the EM segmentation is done. To provide consistency


among the neighbouring blocks, adaptive mean shift and pixel
stretching is implemented. This algorithm involves very much
less computational complexity when compared to the other
methods and hence the time of execution is greatly reduced.
E. Tisse Labeling
Finally the tissues are labeled with pixel values as 128 for (c) (d)
Cerebrospinal fluid, 192 for Gray Matter and 254 for White
Matter. To evaluate the performance metrics, confusion matrix


is formed which gives the matrix of the predicted classes and
actual classes of the three tissue types.
F. Performance Metrics
1. Accuracy: The accuracy of a class is computed using
TN+TP
the expression, Accuracy = (7)
TN+TP+FN+FP
Figure. 4. Segmentation of Simulated image from MRI brain phantom
where TN stands for True Negative, TP for True
with 9% noise rf bias 40% Slice 115
Positive, FN for False Negative and FP for False
(a) Original image (b) Manual segmented image
Positive. (c)EM Segmentation, Segmentation Accuracy=81. 5%
2. Tanimoto coefficient: The overlap between the (d) NSCT based EM segmentation with adaptive mean shift and pixel
automatic segmentation and the ground truth is stretching, Segmentation Accuracy=9\ %
measured by Tanmimoto coefficient Tc. The Tc for
each tissue type i is defined by
2014 International Conference on Electronics and Communication Systems (lCECS -2014), Feb.13 -14, 2014, Coimbatore,
INDIA
B. Segmentation of 3D MR Images from MR IBSR real Nonsumsampled Contourlet transform employs an efficient
dataset filter bank structure for decomposition and reconstruction.
The method is also validated on the real brain volwnes. A The low pass output of the Nonsubsampled Contourlet
set of 20 normal Tl-weighted real brain data was downloaded transform incorporates the spatial correlation between the
from the IBSR repository. Each volwne consists of around 60 neighboring pixels and hence gives better segmentation
coronal Tl slices. Each of the brain volumes in the IBSR site accuracy when subjected to Expectation Maximization.
is provided with manual segmentation by expert clinician. The Results ensure that it achieves an improvement in
segmentation accuracy of the algorithm is compared to that of segmentation accuracy of around 8.5% for a noise level of
six standard segmentation algorithms provided by the IBSR 9%.
which include MAP, adaptive MAP,BMAP, FCM, tree­ Table I. Comparison of average execution time for segmentation of 20
structure k-mean, and maximum-likelihood classifier 3D brain image volumes with the proposed method and fuzzy local Gaussian
mixture model
algorithms and also with the recently proposed fuzzy local
Gaussian mixture model[6].The Tanimoto performance metric
for GM segmentation and WM segmentation for these METHOD FUZZY LOCAL PROPOSED METHOD
algorithms is plotted in figures 5 and 6.It shows that there is GAUSSIAN
4% increase in segmentation accuracy for GM segmentation MIXTURE MODEL
and little decrease in accuracy for WM segmentation.
TIME (SECS) 655 137
..
0.9 The greater the nOIse, the more nwnber of decompOSItIOn
0.8 levels to be employed. Since the GMM is preserved in the
� 0.7
image blocks of appropriate size, a block based EM
UJ
segmentation followed by an adaptive mean shift and pixel
Ii! 0.6
tt stretching is presented. It is inferred from the results that the
8 05
. method is efficient and better based for segmentation of any
o
b 0.4 real and simulated images from IBSR database. The results
� 0.3 may very well be extended to clinical MR brain images also
<t
>-- 0.2 and for identification of brain diseases. Qualitative and
0.1 - FlGMM
Quantitative analysis depicts that the method outperforms the
o -r--r--r---'---'--r---r---r---r---r---r-=�--,P+,
+--.---.---r- ""dTM
0 0..
r =T' th=.,
""."" od other methods in literature with respect to both computation
cost and also accuracy.

VI. REFERENCES
[I] Koen Van Leemput, Frederik Maes, Dirk Vandermeulen, and Paul
Fig. 5. Tanimoto performance metric of GM segmentation obtained by
Suetens, "Automated Model-Based Tissue Classification of MR Images
applying eight segmentation algorithms to 20 3-D IBSR brain MR images. of the Brain", IEEE Transactions On Medical Imaging, Vol. 18, No. 10,
l.9 October 1999 , pp.897 -908.
[2] Koen Van Leemput, Frederik Maes, Dirk Vandermeulen, and Paul
l.8 Suetens, "A UnifYing Framework for Partial Volume Segmentation of
>-- l.7 Brain MR Images", IEEE Transactions On Medical Imaging, Vol. 22,
z No. I, January 2003, pp.105-119.
w
Ol.6 [3] Koen Van Leemput, Frederik Maes, Dirk Vandermeulen, and Paul
it Suetens, "Automated Model-Based Bias Field Correction of MR Images
�l.5 -6MAP
u of the Brain", IEEE Transactions on Medical Imaging, Vo1.l8, No.IO,
-KM
�l.4 October 1999, pp. 885-896.
o -MAP
� l3 MlC
[4] Benoit Scherrer, Florence Forbes, Catherine Garbay, and Michel Dojat,
-
Senior Member, IEEE,"Distributed Local MRF Models for Tissue and
;:
l.2 - rn(MEANS Structure Brain Segmentation" IEEE Transactions on Medical Imag­
-R.GMM ing,Vol. 28,No. 8, August 2009, pp 1278-1295
l.1
-Proposed Method [5] Thanh Minh Nguyen and Q. M. Jonathan Wu,"Gaussian-Mixture-Model­
Based Spatial Neighborhood Relationships for Pixel Labeling Problem",
� 'I> • "II � ." ." • 'I> ." ." ." ",'" � ." � ." � ." ." IEEE Transactions on Systems, Man, and CyberneticsPART B: Cyber­
<" .... '\.1 to' ",'" ",fe'�' ttl' ""..,"'0' ..,'t"" ..,....'\./��, �"I ..,ot"" ,,"'�' '\. .. ,,0'" ...... '
netics, VOL. 42, No. I, February 2012.
IBSR BRAIN VOLUMES [6] Zexuan Ji, Yong Xia, Member, IEEE, Quansen Sun, Qiang Chen, Mem­
ber, IEEE, Deshen Xia, and David Dagan Feng, Fellow, IEEE,"Fuzzy
Figure 6. Tanimoto performance metric of WM segmentation obtained by Local Gaussian Mixture Model for Brain MR Image Segmentation",
applying eight segmentation algorithms to 20 3-D IBSR brain MR images. IEEE Transactions on Information Technology in Biomedicine, vol. 16,
no. 3, MAY 2012, pp.339 - 347.
V. CONCLUSION [7] Arthur L. da Cunha, Jianping Zhou, Minh N. DO,"The Nonsubsampled
Contourlet Transform:Theory, Design and Applications", IEEE Transac­
In this paper, a new algorithm of Block based Expectation tions on Image Processing, Vo1.l5, No.IO, October 2006.
Maximization Segmentation of MR brain images with
adaptive mean shift and pixel stretching is presented. The
[%1 �p://www.cma.mgh. harvard.edU!ibsr/

View publication stats

Das könnte Ihnen auch gefallen