Classification And Improved Segmentation of MR Images
For Radiosurgery Treatment
Gopal.G , Karthik.K , Lakshmanan.M, Palpandian. G & Arasa Kumar
Department of ECE,
Kalasalingam Institute of Technology
E-mail : mainprojectdip@gmail.com
Abstract – This paper is all about presenting an idea
I.
which will be useful for radiosurgery (especially brain
tumor).We present a fast and robust practical tool for
segmentation of solid tumors with minimal user interaction
to assist clinicians and researchers in radiosurgery
planning and assessment of the response to the therapy.
With a previous step of segmentation we are making a
classification between tumor-affected and non-affected
images. Classification is done by feature extraction.A
cellular automata (CA) based seeded tumor segmentation
method on contrast enhanced T1 weighted magnetic
resonance (MR) images ,which standardizes the volume of
interest (VOI) and seed selection, is proposed.
Furthermore, a sensitivity parameter is introduced to
adapt to the heterogeneous tumor segmentation problem,
and an implicit level set surface is evolved on a tumor
probability map constructed from CA states to impose
spatial smoothness. Sufficient information to initialize the
algorithm is gathered from the user simply by a line drawn
on the maximum diameter of the tumor, in line with the
clinical practice. Furthermore, an algorithm based on CA
is presented to differentiate necrotic and enhancing tumor
tissue content, which gains importance for a detailed
assessment of radiation therapy response. Validation
studies on both clinical and synthetic brain tumor datasets
demonstrate 80%–90% overlap performance of the
proposed algorithm with an emphasis on less sensitivity to
seed initialization, robustness with respect to different and
heterogeneous tumor types, and its efficiency.
INTRODUCTION
With the advancement in the field of medical
science even the treatment of extremely complicated
diseases has become easy and flexible. One such
advancement is the radiosurgery (treatment is done
through radiation).But radiosurgery for treating diseases
like brain tumor. Radiosurgery treatment gives a very
high dose of radiation to a very small, precise area. The
radiation is capable of destroying the cells that comes in
it’s path. Thus radiosurgery needs clear area of damaged
tissues so that the non- cancerous cells don’t get
damaged. Thus Segmentation plays a vital role in
treating the brain tumor. In addition a classification
process is being used to identify presence of tumor in
the image. This is followed by segmentation so as to
clearly obtain the tumor region. The computation time
for segmentation is a factor of interest. In this proposed
model the computation time is greatly reduced and is
achieved with less human intrusion. Through this
method the necrotic part from the tumor. Next, as our
application is in the clinical radiosurgery planning,
where manual segmentation of tumors are carried out on
contrast enhanced T1-MR images by a radio oncology
expert, we modify the CA segmentation towards the
nature of\the tumor properties undergoing radiation
therapy by adapting relevant transition rules.
Keywords—Brain tumor classification and segmentation,
cellular automata, contrast enhanced magnetic resonance
imaging (MRI), necrotic tissue segmentation, radiosurgery,
seeded segmentation.
ISSN (Print) : 2319 – 2526, Volume-2, Issue-1, 2013
127
Special Issue of International Journal on Advanced Computer Theory and Engineering (IJACTE)
II. CLASSIFICATION
INPUT: (a) Already labeled training data
The use of classification is to detect the presence of
tumor in the Magnetic Resonance Image. Initially the
image is filtered using Gaussian filter. The filter
removes any noise present in the image. The next step is
feature extraction. Before feature extraction the RGB
component is converted into Grey level. The extraction
process involve extracting 12 features such as
Autocorrelation (out.autoc), Contrast (out.contr),
Correlation (out.corrm), Correlation (out.corrp), Cluster
Prominence(out.cprom), Cluster Shade (out.cshad),
Dissimilarity (out.dissi), Energy (out.energ),Entropy
(out.entro),,Homogeneity (out.homom), Homogeneity
(out.homop), Maximum probability (out.maxpr). The
so obtained features of the input image is compared with
the pre-evaluated values of an normal image(image of
non-tumor image).This is further proceeded with
classification of image. K-NN (K-Nearest Neighbor)
classifier is used for classification. K-NN is a
classification process that involves comparing the
neighbor cell’s status before making a decision.
{Xi| i=1,2,…,n}.
(b) The test datum X.
ALGORITHM:
FOR i=1,2,…upto n
Determine the distance between x and xi.
IF(i ≤ K)
Include Xi in the set of K-nearest neighbors.
ELSE IF (xi is closer to x than any previous nearest
Neighbors)
2.1 K- NEAREST NEIGHBORS ALGORITHM:
Delete the farthest of the K-nearest neighbors.
In this method, for each test datum, the Euclidean
distances between the test datum and all the training
data are calculated, and the test datum is assigned the
class label that most of the K closest training data have.
The K-NN algorithm assumes that all the data
correspond to points in the N- dimensional space KN.
Let the test datum xi be represented by the feature vector
[ x1i ,x2i ,x3i ,…,xNi] , where xki denotes the the value of
Kth attribute of the test datum xi, and x’i is the transpose
of xi. The distance between xi and xj is defined as
d(xi,xj)= √(∑Nk=1(xki-xkj)2. If the number of training data
is n, then n such distances will be calculated, and the
closest K training data are identified as neighbors. If
K=1, then the class label of the test datum is equal to the
closest training datum. If K >1, then the class label of
the test datum is equal to the class label that most of the
neighbors have. If there is a tie, then the tie is resolved
arbitrarily. The neighbors are taken from a set of objects
for which the correct classification (or, in the case of
regression, the value of the property) is known. This can
be thought of as the training set for the algorithm,
though no explicit training step is required. The knearest neighbor algorithm is sensitive to the local
structure of the data. Nearest neighbor rules in effect
compute the decision boundary in an implicit manner.
Include xi in the set of k- nearest neighbors.
END IF
END FOR
FOR c= 1 to c
Pc(x)=(1/K)(no. of neighbors in class c) (3)
END FOR
Crisp class label of x is j
When pj= max{p1,p2,…….,pc}
OUTPUT: (a) Class label of x.
(b) Class confidence value pc for all c.
ISSN (Print) : 2319 – 2526, Volume-2, Issue-1, 2013
128
Special Issue of International Journal on Advanced Computer Theory and Engineering (IJACTE)
III. METHODOLOGY
to Graph-cuts for, random walker for, and shortest paths
for . In image segmentation problems, vertices are
corresponding to image pixels, while edge weights are
similarity measures between neighboring pixels based
on image features.
It is well known that the medical image is difficult
to segment automatically by the computer for its
complexity. The pixel value of the interested organ in
the image may vary greatly. The boundary may be
blurred, or have some gaps. Thus, some low level image
segmentation methods cannot fulfill the complex
medical image segmentation, particularly in brain tumor
segmentation. It becomes more important while
typically dealing with medical images where presurgery and post surgery decisions are required for the
purpose of initiating and speeding up the recovery
process. Computer aided detection of abnormal growth
of tissues is primarily motivated by the necessity of
achieving maximum possible accuracy. Manual
segmentation of these abnormal tissues cannot be
compared with modern day’s high speed computing
machines which enable us to visually observe the
volume and location of unwanted tissues. A brain Image
consists of four regions i.e. gray matter (GM), white
matter (WM), cerebrospinal fluid (CSF) and
background. These regions can be considered as four
different classes. Therefore, an input image needs to be
divided into these four classes. In order to avoid the
chances of misclassification, the outer eleptical shaped
object should be removed. By removing this object we
will get rid of non brain tissues and will be left with
only soft tissues. In this experiment we have used T1,
T2 and PD weighted brain MRIs. These images posses
same size and same pixel intensity values . The pixels
from the image under consideration is supposed to be
3.2. Seed Selection
Seed selection algorithm employs the same idea to
follow the familiar clinical routine to which the
clinicians are used to: the volume of interest (VOI), the
tumor seeds and the background seeds are determined
by using the line already drawn by the user to measure
the longest diameter of the solid tumor. Focusing on
tumor segmentation problem, the seed selection
procedure starts with a single line drawn by the user
along the longest visible diameter of the tumor.
Afterwards, the VOI and the seeds are computed as
follows: 1) The line is cropped by 15% from each end
and thickened to three pixels wide to obtain tumor
seeds; 2) VOI is selected as the bounding box of the
sphere having a diameter 35% longer than the line; 3)
One-voxel-wide border of this VOI is used as
background seeds.
grouped in any one of the aforementioned class. Finally,
by applying certain post processing operations, the
tumerous region can be extracted
3.3 Cellular Automata in Image Segmentation:
3.1. Seeded Image Segmentation
Cellular Automata consists of a regular grid of
cells, each in one of a finite number of states, such as on
and off (in contrast to a coupled map lattice). The grid
can be in any finite number of dimensions. For each
cell, a set of cells called its neighborhood (usually
including the cell itself) is defined relative to the
specified cell. An initial state (time t=0) is selected by
assigning a state for each cell. A new generation is
created (advancing t by 1), according to some fixed rule
(generally, a mathematical function) that determines the
new state of each cell in terms of the current state of the
cell and the states of the cells in its neighborhood.
Typically, the rule for updating the state of cells is the
same for each cell and does not change over time, and is
applied to the whole grid simultaneously, though
exceptions are known, such as the probabilistic cellular
automata and asynchronous cellular automaton. Growcut method uses continuous state cellular automata to
interactively label images using user supplied seeds. The
As an initial step the Volume Of Intrest(VOI) from
the image is considered. The image is being classified
based on seeds as foreground(tumor part) and
background using seed selection techniques as follows.
Given an undirected graph with vertices and edges, a
weighted graph assigns a value (typically real and
nonnegative) to each edge between vertices and in
image
segmentation
problems,
vertices
are
corresponding to image pixels, while edge weights are
similarity measures between neighboring pixels based
on image features (e.g., intensities). Each vertex has an
attribute , which is an indicator of the probability of a
label (e.g., a foreground and a background label). With
the foreground F and background B seeds supplied by
the user, the labeling problem is solved. In the final
solution, the vertices which have the value are labeled as
foreground and those without values are labeled as
background. The solution has been shown to converge
ISSN (Print) : 2319 – 2526, Volume-2, Issue-1, 2013
129
Special Issue of International Journal on Advanced Computer Theory and Engineering (IJACTE)
cells are corresponding to image pixels, and the feature
vector is RGB or gray scale intensities. The state set for
each image pixel consists of a “strength” value in a
continuous interval, a label and an image feature vector.
The automata are initialized by assigning corresponding
labels at seeds with a strength value between 0 and 1
where a higher value reflects a higher confidence in
choosing the seed. Strengths for unlabeled cells are set
to 0. Since the VOI is completely bounded by the
background seeds, each path connecting inside and
outside the VOI is blocked by a seed. Then, the result of
labeling using only the data inside the region is
equivalent to using the whole volume whereas the
computation time is significantly reduced.
characteristics are likely to be healthy. Secondly, it is
possible to include misclassified necrotic regions to
tumor region, which are usually surrounded by
enhanced Tissue. CA algorithm has the advantage of
finding distance of each cell to the nearest seed in a
simultaneous iteration. The probability for tumor is
given by,
Ptumor =
DB / (DB+DT)
Where,
DB = distances for background
DT = distances for tumor
Fig: Effect of smoothing. Example of tumor slice with
vascularization and necrotic part. Tumor probability
map obtained by CA algorithm. Segmentation result
before smoothing (red), after smoothing (blue), and
expert (yellow)
The level-set-based smoothing over the constructed
tumor probability map in constitutes an important part
of the proposed method, as the clinical expert
segmentation, particularly in radiation oncology, mainly
outlines the tumor borders using contouring for
radiotherapy planning as opposed to pixel by pixel
labeling of the tumor carried out in some validation
studies. As a result, our interactive tumor segmentation
includes an appropriate intelligent smoothing of the
tumor borders based on the labeling results obtained
from a graph-theoretic approach. This is a process that is
expected to simulate the expert’s manual contouring.
3.5.Enhancing/NecroticSegmentation
In the seeded tumor segmentation application over
contrast enhanced T1-weighted MRI for heterogeneous
tumors, which mostly consist of a ring enhancing region
around a dark necrotic core (and also irregular borders),
most of the foreground seeds fall in the necrotic region.
This sometimes causes the segmentation algorithm to
get stuck at necrotic to enhancing tumor transition
borders. To overcome such problems, prior knowledge
that tumor voxels are brighter in post-contrast T1-MRI
can be initialized.
Quantification of the necrotic regions within a
whole tumor is an important problem in assessment of
the tumor progress. Delayed radiation necrosis, which
typically occurs three months or more after treatment, is
the primary risk associated with stereotactic
radiosurgery. Necrosis of the tumor can occur as a result
of the radiosurgery as well as by the tumor progress
itself like in high grade gliomas. Necrotic class naturally
arises in segmentation using multi protocol (T1, CE-T1,
T2, DWI, etc.) intensity classifiers due to its different
intensity characteristics in different modalities.
However, our aim in this study is to quantify the
necrotic and enhanced parts of the tumor using solely
contrast enhanced T1weighted MRI volumes.
3.4 Level Set Evolution on Constructed Tumor
Probability Map
Smoothing is an important prior in segmentation of
brain tumors from post-contrast T1 images, because of
three main reasons: First, an area surrounded by tumor
tissue is considered as a tumor region even the intensity
ISSN (Print) : 2319 – 2526, Volume-2, Issue-1, 2013
130
Special Issue of International Journal on Advanced Computer Theory and Engineering (IJACTE)
Fig :A level set surface of PT = 0.5 is initialized and
evolved on the tumor probability map.
Fig: Enhanced and necrotic seeds determined by the proposed method,
used as an input to the CA segmentation algorithm (Necrotic seeds in
red and enhanced seeds in blue.
The necrotic part of the tumor is the region where
the lesions are secreted. The lesion is dangerous enough
to spread the tumor to the organs and parts of the body.
In CE-T1 MR images, necrotic parts of the tumor are
observed as hypo-intense for there is no blood flow into
these regions where enhanced parts are hyper-intense.
Without any prior information, segmentation using an
intensity threshold can be applied by assigning necrotic
label to the voxels lower than the chosen threshold and
enhanced label to those that are higher. The utilized data
simulates contrast-enhanced T1-weighted MR images
with synthetically generated tumors. The tumor
probability maps are also available with the dataset for
validation purposes. This dataset is included in the
performance evaluations since the ground truth
segmentation is readily available. To choose the
threshold, we explored using expectation maximization
and Otsu’s methods. However, usually the two classes
are not separable on the intensity histogram even though
they could be separated easily on the image. Instead of
using simple thresholding , connectedness was imposed
by using the CA algorithm with two thresholds as
follows: Initially the voxels lower than a necrotic
threshold are labeled as necrotic seeds and higher than
an enhanced threshold are labeled as enhanced seeds .
Next, the voxels at remaining mid-intensities are labeled
by assigning the label of the nearest seed using the CA
algorithm. An algorithm to choose the two thresholds is
devised as follows: The necrotic part is obtained using
region growing from the selected seed. First the number
of necrotic voxels and the number of enhanced voxels
are roughly calculated by using Otsu’smethod. In CE-T1
MR images, necrotic parts of the tumor are observed as
hypo-intense for there is no blood flow into these
regions where enhanced parts are hyper-intense.
Without any prior information, segmentation using an
intensity threshold can be applied by assigning necrotic
label to the voxels lower than the chosen threshold and
enhanced label to those that are higher.
Fig : The necrotic part inside the tumor obtained from
region growing
IV. CONCLUSION
We presented a segmentation algorithm for the
problem of tumor delineation which exhibit varying
tissue characteristics. As the change in necrotic and
enhancing part of the tumor after radiation therapy
becomes important, we also applied the Tumor-cut
segmentation to partition the tumor tissue further into its
necrotic and enhancing parts. We presented validation
studies over a synthetic tumor database and two real
tumor databases: one from Harvard tumor repository
and another from a clinical database of tumors that
underwent radiosurgery planning at Radiation Oncology
Department of ASM. Strengths of the proposed method
include its simple interaction over a single slice and less
sensitivity to the initialization (demonstrated by lower
coefficient of variation values), its efficiency in terms of
computation time, and robustness with respect to
different and heterogeneous tumor types. Choosing the
contrast enhanced T1 modality limits the application to
ISSN (Print) : 2319 – 2526, Volume-2, Issue-1, 2013
131
Special Issue of International Journal on Advanced Computer Theory and Engineering (IJACTE)
the tumors that are enhanced with the contrast agent,
excluding the edema/infiltration region around the
tumor. For the targeted clinical application of
radiosurgery planning, using a single modality is an
advantage due to the computational efficiency and ease
of use
[3].
Chunyan Jiang, Xinhua Zhang, Wanjun Huang,
Christoph
Meinel“Segmentation
and
Quantification of Brain Tumor” VECIMS 2004 –
IEEE International Conference on Virtual
Environments, Human-Computer Interfaces, and
Measurement Systems Boston, MD, USA, 12-14
July 2004
[4].
Andac Hamamci*, Nadir Kucuk, Kutlay
Karaman, Kayihan Engin, and Gozde Unal,”
Cellular Automata Segmentation of Brain Tumors
on Post Contrast MR Images”
[5].
Manish sarkar,Tze-Yen Leong,”application of KNearest Neighbors on breast cancer diagnosis
problem”
[6].
M. Masroor Ahmed , Dzulkifli Bin Mohamad,”
Segmentation of Brain MR Images for Tumor
Extraction by Combining Kmeans Clustering
anPerona-Malik Anisotropic Diffusion Model”
V. FUTURE ENHANCEMENTS
In the future, the system should be improved by
adapting more segmentation algorithms to suit the
different medical image segmentation aims. For setting
the deformation region in level-set method, the intensity
range can reveal some target area. However, the brain
tumor case is complex. We need more sophisticated
strategy to set deformation region where the level set
model’s propagation terminates in. In order to balance
the expert hand work and computer automatic work, the
system should enhance the precision of the algorithm
computation, and also improve the user interface to
facilitate the user control. This tool can be used in the
segmentation and quantification of many types of
medical images. For the brain tumor case, we should
add some other features for the aid in the diagnoses
VI. REFERENCES
[1].
M.-R. Nazem-Zadeh, E. Davoodi-Bojd, and H.
Soltanian-Zadeh,“Atlasbased
fiber
bundle
segmentation using principal diffusion directions
and spherical harmonic coefficients,” Neuro
Image, vol. 54, pp. S146–S164, 2011.
[2].
Andac Hamamci*, Nadir Kucuk, Kutlay
Karaman, Kayihan Engin, and Gozde Unal
“tumor-cut: segmentation of brain tumors on
contrast enhanced MR images for radiosurgery
applications”. ieee transactions on medical
imaging, vol.31, no.3, march 2012

ISSN (Print) : 2319 – 2526, Volume-2, Issue-1, 2013
132