International Conference on Communication and Signal Processing, April 3-5, 2018, India

Brain Tumor Detection using Image

Segmentation Techniques
Digvijay Reddy, Dheeraj, Kiran, Bhavana.V and Krishnappa H.K

Abstract—Brain Tumor is the unconstrained growth of bizarre tumor discovery since it gives more precise data about size and
cells in brain. In this work, dicom Magnetic Resonance Image position of tumor. K-mean grouping and Level set division
(MRI) is taken as an input and tried to extract tumor cells from (Level set segmentation) [1] is utilized as a part of the
the input image. Pre-processing technique is used to remove noise
procedure. In this work, picture division strategy is used for
from image.To this image, k-means clustering is applied and from
this clustered image, skull was removed using morphological separating tumor, which is done in two stages. In the first stage,
operations to identify tumor cells easily. Finally, image MRI image is taken and after that preprocessing is done to
thresholding is applied to this image followed by levelset expel clamor from the informative picture. Second stage
segmentation to extract tumor cells. Performance matrices like incorporates grouping, thresholding and division.
true positive (TP), true negative (TN), false positive (FP), false
negative (FN), precision and recall to measure the accuracy of our The rest of the paper is organized as below. The related work
results, is also evaluated. of this paper is listed in section II where in section III the
proposed work has been discussed. The methodology is
Index Terms—Brain tumor detection K-means clustering, explained in section IV, the experimental results and
Levelset segmentation, MRI, Morphological operation. performance measures are discussed in section V and VI
respectively. Finally, section VII concludes the paper.

I. INTRODUCTION
II. RELATED WORK
M RI offers the favorable position to be a noninvasive
method that empowers the investigation of cerebrum M.T. El-Melegy and H. M. Mokhtar proposed a new fuzzy
tissues. The early discovery of tumor in the cerebrum leads on [2] approach for the automated segmentation of pathological
sparing the patients' life through appropriate care. Because of brain resonance imaging (MRI). This technique reformulates
the expansion of medicinal information in clinical diagnosis, fuzzy c-means (FCM) algorithm rule required under
the precise location of tumors in MRI turns into an exact consideration of any accessible information regarding the
assignment to perform various operations. Besides, tumor category center. This approach has smart higher segmentation
location in a picture is helpful for therapeutic specialists, as accuracy, strength against noise, and quicker response
well as for different purposes like 3D reproduction etc. compared to other well-known fuzzy based techniques.
The programmed location of tumor area helps in making the Jason J. Corso, Eitan Sharon, and Alan Yuille proposed
determination procedure simple.MRI is favored for cerebrum another technique for programmed division of heterogeneous
picture information [3], which is extremely regular in
therapeutic picture examination. The primary commitment of
the work is a numerical detailing for consolidating delicate
model assignments into the computation of affinities. This
strategy does not unequivocally exploit the class models
Digvijay Reddy is with the Department of Electronics & Communication utilized in the process.
Engineering,Amrita School of Engineering,Bengaluru,Amrita Vishwa
Vidyapeetham,India.(e-mail: [email protected]). Dimitris N Metaxas, Zhen Qian, Xiaolei Huang and Rui
Dheeraj is with the Department of Electronics & Communication Huang Ting Chen, Leon Axel have identified a new class of
Engineering,Amrita School of Engineering, Bengaluru,Amrita Vishwa
deformable models, called hybrid deformable models to
Vidyapeetham,India.
Kiran is with the Department of Electronics & Communication computerize the model instatement process and make changes
Engineering,Amrita School of Engineering,Bengaluru,Amrita Vishwa in division and registration. These strategies incorporate
Vidyapeetham,India. Metamorphs [4], a novel shape and surface mix deformable
Bhavana.V is with the Department of Electronics & Communication
model structure and the combination of deformable models
Engineering,Amrita School of Engineering,Bengaluru,Amrita Vishwa
Vidyapeetham,India.(e-mail: [email protected]). with graphical models and learning techniques.
Krishnappa H.K is with the Department of Computer Science &
Engineering,R.V College of Engineering,Bengaluru,India.
The process of segmentation of Gray matter, White matter and
CSF in the brain portions were done based on clustering
.

978-1-5386-3521-6/18/$31.00 ©2018 IEEE

0018
techniques like clustering process, FCM clustering process 1) FILTERING
and spatial Fuzzy C means. The process of clustering Median filter is employed to reduce noise in the image and to
identifies different portions in the input images based on filter unwanted pixels. Median filter parses through the image
identification of difference in intensities between the input pixel by pixel, replacing each pixel with the median of
image pixels. Number of clusters was selected are based on neighboring pixels thereby providing normalized intensity in
this.For each image pixels randomly assign point coefficients overall image.
for being in the clusters. Calculate the centroid for each
2) CLUSTERING
cluster. For each point, compute its coefficients of being in
Clustering process is employed based on K-means process.
the clusters. Repeat the process till algorithm has converged.
The regions in the images that are having similar intensity is
The convergence is achieved till the coefficients value change
clustered using the K-Means algorithm. Clustering is done for
between two iterations is not more than a particular stopping
skull removal
value.
3) K-MEANS CLUSTERING
The segmented portions were not more accurate since the
The clustering process is employed using K Means approach.
skull portion in the images affects the clustering process
In this approach, the value of k is assumed first and allocates
more. The performances of the process measured were not
centers to each k value. After allocating centers, all the points
much improved since the difference between the exact
that are closest to each center is marked. After marking all
segmentation and the ground truth is more.
points for k clusters, then the mean for each cluster is
determined and the process is repeated until there is further
III. PROPOSED WORK change in the mean.

Fig. 1 shows the process involved in detecting brain tumor by 4) THRESHOLDING

image segmentation method and Fig. 2 represents the flow The specific threshold for the brain tumor portion is
diagram of algorithms involved in each step. identified based on the identification of maximum pixel of the
input image. The image pixels having maximum intensity are
identified to be tumor portions. The threshold level [5] is
adjusted such that the threshold value is 30 pixels lower than
maximum pixel. The image pixels that are greater than the
threshold values are assigned to the value one and others are
given the value zero. The pixels that are having value ‘one’ is
the tumor portion in the images and others are normal image
portions.
5) SKULL REMOVAL
The skull regions captured along with the brain region in the
images are of different intensities and are in larger area. The
Fig. 1. Steps involved in proposed method skull regions are identified and removed based on open and
close morphological operation. A small binary image is the
structuring element, i.e. a small matrix of pixels, each with a
value of zero or one. The size of the structuring element is
specified by matrix dimensions and the pattern of ones and
zeros specifies the shape of the structuring element. The
closing operation of an image by a structuring element is a
dilation followed by erosion and the opening operation of an
image by a structuring element is erosion followed by dilation.
6) MORPHOLOGICAL FILTERING
Morphological Filtering of a two-fold picture is done by
considering operations like opening and closing as channels.
For example, closing with a circle smoothens corners from the
outside and opening with a plate arranging part smoothens
corners from inside. These operations filter through any
inconspicuous components from a photo that are tinier in size
than the sorting output, i.e., opening is isolating the twofold
picture at a scale portrayed by the traverse of the sorting out
Fig. 2. Flow diagram of algorithms involved in each step segment. The fragments of the photo that fits the arranging
The various steps involved in this process are described segment are passed by the channel and tinier structures are
below: dodged from the yield picture. The degree of the sorting out

0019
part is most fundamental to discard hullabaloo yet not to The performance of the process is measured by comparing the
damage objects of interest. segmented portions with the original segmented portions based
on the performance metrics like True positive, True negative,
7) SEGMENTATION False positive, False Negative, Accuracy, Precision and Recall.
Segmentation process is employed by reducing the The performance of the process measured indicates that this
segmented portion till the curve enclosing the regions of approach is more efficient compared to other existing
interest is obtained. The tumor portion obtained using approaches for the segmentation of the tumor portions.
thresholding process is considered as the initial portion of the
segmentation process [6]. The image pixels surrounding the V. EXPERIMENTAL RESULTS AND ANALYSIS
tumor portion that are having similar values nearby the input
region, is then grouped. Contours are extracted around the
The experimental results of this work are described in this
detected regions. The process is repeated till the boundary of
section.
the detected region is within the input region. The final result
obtained is the segmented tumor portion.

Level set Segmentation

The initial phase in Level set is to choose an arrangement of
levels. Level determination depends on some client paradigm.
The underlying locale starts as the correct area of the pixels in
the levels. The pixels that are having comparable furies within
the level are recognized and the pixels are assembled. Before
characterizing, the area homogeneity is confirmed. The two
districts are converged on the premise of the number juggling
mean and the standard deviation.

IV. METHODOLOGY

The input MRI images are initially filtered using median

filter [7]. Skull removal process is then employed based on the
identification of the intensities of the skull potions and removal
of the portions based on morphological operations. The Fig. 3. MRI of Brain
specific threshold for brain tumor portion is detected based on
the identification of the maximum pixels of the input image. Fig. 3 shows input image to be processed. Here input image is
The image pixels having maximum intensities are identified to MRI of brain which is affected by tumor.
be tumor portions.
The threshold value is adjusted such that the threshold value
is 30 pixels lower than maximum pixel. The image pixels that
are greater than the threshold value are assigned label one and
others are given value zero. The pixels that are having value
one is the tumor portion in the images and others are normal
image portions. Level set segmentation process is then applied.
The initial seed for the level set segmentation process is
obtained from thresholding process. Level set segmentation
process is employed by reducing the segmented portion till the
curve enclosing the regions of interest is obtained.
The tumor portion obtained using thresholding process is
considered as the initial portion of the segmentation process.
The images pixels surrounding the tumor portion that are
having similar values nearby the input region is grouped.
Contours are then extracted around the detected regions. The Fig. 4. Noise is removed by using median filter
process is repeated till the boundary detection region is within
the input region. The final result obtained is the segmented
tumor portion.

0020
 Fig. 4 shows the image after preprocessing. Here noise in Fig. 7 shows the image after thresholding where skull
the image is removed by using median filter which normalizes region of brain is removed. The skull region is removed based
the pixel value. on morphological operations.

Fig. 8. Tumor region is detected

Fig. 5. Clustering is done to label image based on color
Fig. 8 shows contour formed around tumor region. Here
Fig. 5 shows the image after clustering. The input image the green outline is contour and levelset segmentation is used
after preprocessing is subjected to clustering using K- here to detect the tumor.
means clustering technique, where different colors are
clustered together.

Fig. 6. Thresholding is done to separate images after clustering

Fig. 6 shows image after thresholding. After k-means Fig. 9. Tumor regions are separated
clustering, different regions of image are clustered based on
color and to separate those, color based regions Fig. 9 shows the image of tumor region which is
thresholding is used. separated from the main image of brain. Thresholding is
used to separate tumor region from brain image after
levelset segmentation.

VI. PERFORMANCE MEASURES

The performance of the process is measured by measuring

the accuracy of the process. The accuracy is measured by
comparing with the ground truth images.
Accuracy = (TP +TN) / (TP+FP+TN+FN)

Precision = TP / (TP+FP)

Fig. 7. Skull region is removed based on morphological operations Recall = TP / (TP+FN)

0021
 VII. CONCLUSION
where, True positive is correctly identified. The methodology presented in this work uses a two-step
False positive is incorrectly identified. procedure for brain tumor detection that combines k-means
True negative is correctly rejected. clustering algorithm followed by levelset segmentation and
False negative is incorrectly rejected. morphological operations. Experimental results have shown
that, this methodology is robust in detecting and bounding the
abnormal cells in MRI images despite the complicate shape of
TABLE I the tumor.
PERFORMANCE EVALUATION MATRICES In future work, various features such as contrast, correlation
Data set Data set Data set etc can be added and also additional information can be
Performance Measures
1 2 3 included to the feature extraction process to make the system
more sensitive to the texture and various other factors.
TP 99.483 99.4247 99.0738

REFERENCES
TN 1.0781 1.2539 2.1406 [1] S. Karthik V.K and K. P. Soman, “Level Set Methodology for Tamil
Document Image Binarization and Segmentation”, Int. Journal of
Computer Applications (IJCA), vol. 39, no. 9, 2012.
FP 0 0 0 [2] M.T. El-Melegy and H. M. Mokhtar, “Tumor segmentation in brain
MRI using a fuzzy approach with class center priors,” EURASIP
Journal on Image and Video Processing, vol. 2014, article no. 21, 2014.
FN 62 59 [3] Jason J. Corso, Eitan Sharon, and Alan Yuille, “Multilevel
56
Segmentation and Integrated Bayesian Model Classification with an
Application to Brain Tumor Segmentation,” MICCAI 2006.
Accuracy 99.9054 99.9146 99.9100 [4] Dimitris N Metaxas, Zhen Qian, Xiaolei Huang and Rui Huang Ting
Chen, Leon Axel, “Hybrid Deformable Models for Medical
Segmentation and Registration,” Ninth International Conference on
Control, Automation, Robotics and Vision, ICARCV 2006, Singapore,
Precision 100 100 100
2006.
[5] Anisha Radhakrishna and Divya, M., “Survey of data fusion and tumor
segmentation techniques using brain images”, International Journal of
Recall 99.9050 99.9141 99.9092 Applied Engineering Research, vol.10, 2015.
[6] R. Hiralal and Menon, H. P., “A survey of brain MRI image
segmentation methods and the issues involved”, Advances in Intelligent
Systems and Computing, vol. 530, pp. 245-259, 2016.
From the test results of 3 datasets, it is clear that this proposed [7] Bhavana. V, Krishnappa H.K, “Multi-Modality Medical Image Fusion
using Discrete Wavelet Transform”, 4th International Conference on
system gives promising test results in terms of accuracy, Eco- friendly Computing and Communication Systems, ICECCS 2015,
precision and recall, it is shown in Table I. ELSEVIER Procedia Computer Science, pp.625-631.

0022