You are on page 1of 6

Bonfring International Journal of Advances in Image Processing, Vol. 6, No.

3, August 2016 22

A Survey of MRI Segmentation Techniques for

Brain Tumor Studies
Prajna Udupa and Sarvesh Vishwakarma

Abstract--- Brain tumor is one of the advancing brain The location, size and shape of the tumor can vary in
illness among the existing population. The causes of the different individuals and the growth rate of these tumors may
condition being non-deterministic, it is found that the survival be affected by several unknown factors. Although the causes
rates of people suffering from brain tumor is very less. This of brain tumors is speculative, some of the causes as identified
medical condition lacks effective treatment and thus by researchers include exposure to ionizing radiations,
approaches that enable efficient treatment planning of the hereditary factors, subjection to industrial chemicals, other
brain tumor is very crucial in the current context. Magnetic medical conditions and corresponding medicine intake, body
resonance image (MRI) scans are widely used in the diagnosis weight and lack of exercise, smoking and alcohol intake and
and subsequent treatment planning of the brain tumors. While so on [2].
in most of the cases only tumor regions are segregated from
People suffering from brain tumors exhibit certain
the MRI scans, segmentation of the tumor along with the
symptoms as the disease progresses some of which include
healthy tissues of the brain such as white matter, gray matter headache, vomiting, seizures, fatigue, vision related problems,
and cerebrospinal fluid enable doctors to study the overall difficulty in speaking, impediment of sensation and walking
medical condition of the patient and treat him accordingly.
and feeling unconscious [3]. The diagnosis of these tumors is
The overall segmentation procedure includes noise removal,
usually by medical examination of symptoms along with the
skull stripping, feature extraction and actual segmentation of
use of medical imaging techniques such as Computed
various brain regions. Several methods used for each of these
Tomography (CT) or Magnetic Resonance imaging for the
stages are discussed in this paper. The advantages and analysis of the tumor affected brain images. It is then followed
drawbacks of each of them is also highlighted. with a biopsy of the sample taken from the tumor affected
Keywords--- Feature Extraction, Noise Removal, brain region that reveals the severity of the tumor [4].
Segmentation, Skull Stripping Identification and classification of tumors using traditional
invasive procedures can damage the brain tissues. To avoid
this, non-invasive methods can be used. These methods take
I. INTRODUCTION the advantage of widely used magnetic resonance images of

T HE human brain forms the most vital part of the central

nervous system which is madeup of tissues such as white
matter, gray matter and the cerebrospinal fluid. This is in turn
the brain to detect the tumor affected brain regions. The
magnetic resonance images contain different intensities for
different regions or tissues of the brain. In the medical field,
protected by a layer of membranes and the skull. The different large number of MR images of the brain are generated for
parts of the human brain are responsible for different functions individuals suffering from brain tumors. Manual segmentation
of the human biological system. Any damage caused to any of of tumors by experts requires substantial time and is usually
these parts can affect the functioning of the corresponding error prone when large number of MR images have to be
human organ. Brain tumor is one of the most commonly found examined by them [5], [6]. To ease their work of segmenting
ailments of the brain in the present day scenario. Although the tumors from MR images, automatic tumor segmentation
survival rate of people suffering from brain tumors is very methods are used. Most of the segmentation algorithms only
scarce, the medical experts mainly focus on the treatment of segment either the healthy tissues [15]-[18] or tumors [7]-[12]
this biological condition that can help elongate the lifespan of from these images with a good accuracy. However,
these patients. segmentation algorithms that segment both the healthy tissues
along with the tumor affected regions of the brain can help in
Brain tumor is the abnormal growth of cells within the efficient medical diagnosis and treatment planning taking into
brain which can be mainly of two types such as malignant and account all of the other biological conditions of patients.
benign. Malignant tumors are high grade tumors that are
cancerous and spread to the other regions of the brain while In this paper, the various methods used for different stages
benign tumors are low grade tumors that are non-cancerous of the brain MRI segmentation procedure are discussed.
and do not spread to the other regions of the brain [1]. Section 2 presents a survey on these methods. Section 3
highlights the advantages and disadvantages of these methods.
Section 4 gives the conclusion to the presented survey.
Prajna Udupa, Student, Dept. of CSE, SCEM, Mangalore, India.
Sarvesh Vishwakarma, Professor, Dept. of CSE, SCEM, Mangalore,
DOI: 10.9756/BIJAIP.10467

ISSN 2277-503X | 2016 Bonfring

Bonfring International Journal of Advances in Image Processing, Vol. 6, No. 3, August 2016 23

II. SURVEY ON MRI SEGMENTATION TECHNIQUES Local Means method for noise elimination automatically
Segmentation of an image is the task of segregating the selects and reduces the search window size for any given pixel
digital image into different fragments or components which i based on the count of fit pixels within a search window. The
makes the analysis of the image much simpler and easier. The fit pixels are the pixels for which the computed weight
segmentation process basically identifies boundaries of denoting the similarity index should be above a certain
various regions within an image and assigns labels to each threshold.
pixel in the image based on the similarity characteristics of the Independent Component Analysis method for noise
pixels. Typical applications of image segmentation include removal represents the observed data as an n-dimensional
medical imaging to locate tumors and pathologies, to plan vector x that contains the pixel information of the image
surgeries, to measure volume of tissues, to diagnose the [23].This data is subjected to pre-processing. Initially the data
anatomical structures, object detection comprising of face is centered by subtracting the data mean from the original data
detection, pedestrian detection, recognition tasks such as face after which the data is exposed to a whitening process which
recognition, fingerprint recognition and so on. aims at removing any of the second-order statistical
The segmentation process basically includes four dependencies that exist in the original data. Standard Principal
important stages namely Noise Removal, Skull Stripping, Component Analysis is usually used to perform whitening of
Feature Extraction and Final Segmentation. The noise removal the data which also accomplishes dimensionality reduction of
step helps eliminate pixel variations in the input data. The the data.
skull stripping stage helps extract the desired brain region The wavelet approach to denoising images comprises of a
from the entire brain MRI. The feature extraction stage helps basis function called the mother wavelet that has a specific
extract prominent features for training the segmentation. The scale denoting the frequency and a unique translation
final segmentation stage is used to obtain the segmented image representing the time domain for the corresponding signal
from the Brain MRI. Each of these stages can be implemented [24]. In addition to this any signal f(x) can be denoted using
with a variety of methods which are discussed in the literature infinite combinations of a discrete wavelet signal, each
survey. corresponding to different scales and translations. On
convolution of the signal f(x) with the mother wavelet
A. Noise Removal
function, wavelet coefficients for the mother wavelet can be
Noise in images is the variation in illumination or color computed. These coefficients can in turn be used to
information of the images which could be caused by circuit reconstruct the original signal f(x). For noise removal of
components of the image capturing device. Noise removal images using wavelet analysis, the wavelet coefficients are
from Magnetic Resonance images is essential in order to subjected to thresholding operations at different frequency
enhance the image quality and help improve the computational values. The coefficients with insignificant energy will be
efficiency of different forms of image processing. The various discarded. This helps to eliminate noise from the signal
segmentation methods available for segmenting the human without altering the signal.
brain MRI are highly sensitive to noise, heterogeneous
intensities found in the image and the low contrast between The Least Mean Square (LMS) adaptive filter makes use
different regions of the image making image segmentation a of the coefficient vector W for the signal, the actual input
difficult task. Thus removal of noise from these images is very signal X and the desired signal d [25]. The LMS algorithm
essential to efficiently segment the brain MRI by overcoming then computes the error signal e from the input signal and the
these challenges. desired signal as their difference and then by using a specific
step size, it computes the corrected input signal that combines
The Gaussian filter is a simple method used to remove the error component into it. This method is iterated until the
noise from images and produce a smoothening effect [20]. It error signal becomes negligible. The Anisotropic Diffusion
uses a kernel function that has the shape of a bell as a filter method for removal of noise from MR images is based on the
and applies this function to the original image that contains the anisotropic diffusion process [26]. In this process different
noise component within it. Median filters can help remove regions of the brain enable different amounts of light to be
impulsive noise from images. They are non-linear in nature diffused into them helping to clearly distinguish between
and function by computing the median of the neighborhood different regions. This method proposed by Perona and Malik
pixels with respect to the pixel of interest [21]. This newly uses the original image and a filter that exploits the local
computed value will be the denoised value for the pixel of features of the original image to produce a resultant image.
interest. This image is obtained after applying transformations that are
The Non Local Means method for removing noise from space variant and non-linear in nature.
images computes the value for a particular pixel i by using all B. Skull Stripping
the pixels within a search window that have a similar
neighborhood in comparison to the neighborhood of the pixel I The human brain is a vital organ comprising of three main
[22]. Weights are assigned to each of the selected pixels based parts namely the forebrain, the midbrain and the hindbrain. In
on the degree of similarity between the patches of selected the study of brain images, separating these tissues associated
pixel with that of the original pixel i. The value for the pixel i with the brain from that of the non-brain tissues is a crucial
is then estimated as the weighted average of all the selected step for processing these brain images. This step is referred to
pixels within the selected search window. The Adaptive Non as skull stripping which helps separate the cerebral tissues
from the other regions like the skull, scalp and the meanings

ISSN 2277-503X | 2016 Bonfring

Bonfring International Journal of Advances in Image Processing, Vol. 6, No. 3, August 2016 24

that protect the brain. There are various techniques employed The Edge based method exploits intensity values along the
for stripping the skull in image processing. edges in an image that differ greatly and thus features
extracted using these edge based values can be very useful for
An elementary method named S3 for brain region
image segmentation [14]. For the edge based intensity values,
extraction uses a median filter to de-noise the image [27]. It
then computes the mean intensity value twice and uses these first order and second order features are determined and then
values to eliminate the skull and membrane regions of the used. Appearance based method for feature extraction
combines both first order intensity features and second order
brain. It then produces the binary image using averaged
spatial features [13]. First order features are effective if the
membrane intensities and applies morphological operations to
contrast of the image is sufficiently good that can help
get the final skull stripped image. Another method to strip the
separate the background from the foreground of the image. In
skull from brain MR images is by combining thresholding
technique with morphological operations [28]. The cerebral cases where images are of low contrast, second order spatial
features are used. These second order features are usually
and non-cerebral tissues are separated from each other by
calculated using the co-occurrence matrices of gray intensities.
using a threshold value for the pixel intensity after which the
The Phase Congruency method operates by computing the
binary image is generated. This binary image then under goes
local phase and energy values for the image [19]. It is a
morphological operations like erosion, dilation and region
filling to extract the desired brain region. frequency dependent technique which computes the Fourier
components for the points in the image and then determines
In the region growing approach, smaller regions of the the points having the same phase values to generate lines,
brain are determined using some criteria which is previously corners or edges as features.
defined and then they are combined together to form larger
regions [29].The Brain Extraction Tool method uses an initial D. Segmentation
surface model for the brain to which the brain surface present MRI image segmentation is necessary in order to
in the image is overlapped [30]. This involves a deformation differentiate between the different regions of the brain from
of the initial model to an optimized model by iterative steps the MR image scans. In threshold based method, global
that use a set of local image features. A robust brain extraction thresholding can be used in order to separate the background
(ROBEX) method for stripping the skull from the brain MR of the image from the foreground objects of similar intensity
images uses a discriminative random forest classifier which is values when the contrast between the background and
trained to identify the boundary regions of the brain [31]. It foreground is high [35].In region based methods, a seed value
then uses the generative point distribution model to find all the is chosen either manually or using an automated
pixels or points within the boundary region. The graph cuts are seed-determining process that corresponds to the region of
then used to obtain the refined final contour. interest [36]. The neighboring pixels of this pixel are then
examined based on some similarity criterion to find the pixels
Another approach to segment brain tissues from non-brain
having similar properties as the seed pixel. Connected regions
tissues is by finding several contour regions within the brain
are formed using these neighboring pixels and the process is
region extracted from a preliminary segmentation [32].
repeated until no more pixels can be added to the region.
Starting from an initial contour, it evolves and expands the
curve and repeatedly finds all the pixels within the specified In the classification method, the training data is used to
contour until a final image mask showcasing the brain region generate manual segmentation which is then used as a
is developed. Hybrid methods constitute a combination of reference for automatic segmentation of the new input data
several techniques used to produce accurate results and thus [37]. This method requires the image to be segmented to
enhance the efficiency of skull stripping [27]. contain distinct features that are quantifiable that allows
correct labeling of image pixels by the training data. The
C. Feature Extraction clustering algorithm is trained to learn on its own from the
Feature extraction in MR images is usually accomplished available data to deduce the segmentation [38]. Clustering
using wavelet transforms. Discrete Wavelet Transform utilizes methods do require an initial segmentation to identify the
a mother wavelet required for the decomposition and then different classes in an image and they are highly sensitive to
alters its scaling and translation parameters to generate this initial segmentation.
wavelet functions from it [34]. This wavelet function is then
Parametric and Geometric deformable models combine the
used to embed the original input signal onto it to generate the
transformed signal. At each stage of the decomposition, a prior knowledge related to the location, size and shape of
down-sampling process is performed to eliminate the different structures with the constraints learnt from the image
data to deform the original model that then produces the
components that do not pass through the filters causing the
segmented model of the brain MRI [39].K-Nearest Neighbors
decomposed signal to be of smaller length than that of the
method can be used to segment the MR image by classifying
original signal. The Stationary Wavelet Transform combines
each pixel using an integer value k that determines the number
all the elements that would be selected from multiple
decomposition steps in epsilon-decimated DWT to a single of neighboring pixels to be examined, a Euclidean distance to
determine the k closest neighbors and a set of labeled
step [33].The decomposed signal has the same length as that
examples for some training data [41].
of the original signal with no elements of the original signal
lost by the decomposition.

ISSN 2277-503X | 2016 Bonfring

Bonfring International Journal of Advances in Image Processing, Vol. 6, No. 3, August 2016 25

Bayesian approach makes use of a supervised and eliminating the smoothening effect across image boundaries
parametric approach [42]. In this method, using the training thus helping to highlight the different brain regions.
dataset, the data and its associated classes are identified for
B. Skull Stripping
which the probability values are also computed. For each new
data that arrives, the posterior probability is computed using The S3 method adaptively computes the threshold value
the Bayes rule along with the prior probability computed for making it more robust. Although it outperforms the other
the training set to determine the class with maximum posterior methods like BET, BSE and ROBEX, it will suffer efficiency
probability. This is the class assigned to the data. Statistical if used with low contrast images or highly noisy images due to
model based approaches can be used for image segmentation erroneous threshold values computed. Although thresholding
[43]. Markov random field models spatial interactions among combined with morphological method is simple and efficient,
the neighboring pixels of an image. It is a statistical model that it fails to clearly differentiate between the brain and non-brain
is used in combination with other segmentation algorithms regions of same intensity values. In order to attain higher
that helps to place pixels with similar properties other than accuracy in skull stripping using this method, the threshold
pixel intensities into one group. value selection plays a major role. The performance of the
region growing approach depends on the seed value selected
Artificial Neural Network based approach can be for the brain and non-brain regions either manually or using
implemented as a classifier where the weights for the ANN are any algorithm.
determined using the training data and then use these weights
to segment the new data [44]. It can also be implemented as a The Brain Extraction Tool method is known for its speed
clustering method or a deformable model. The numerous and eliminates any kind of registration or preprocessing. A
interconnections between the neural network nodes helps potential drawback could be over estimation of the brain
model the spatial interactions among the image pixels. Atlas surface and sensitivity to noise resulting in inaccurate
guided approaches for brain MRI segmentation make use of a stripping [31]. The ROBEX method fails to generate sharp and
standard template or atlas that is generated using the accurate brain regions, especially if the brain comprises of
anatomical information from images which require highly convoluted regions [27]. It requires more time for
segmentation [45]. This atlas is then used as a reference to computation as compared with its counterpart methods.
segment new images taken as input. This approach performs Another drawback is to standardize the image intensity values
an atlas warping where it finds a one-to-one mapping from the in order to produce more accurate final segmentation.
previously segmented atlas image to the target image to be The contour method thus facilitates to strip the skull
segmented. Use of hybrid approaches to segment the brain efficiently by being able to produce genuine results in less
MR images produces more efficient results by overcoming the time. It is also found to be insensitive to noise that is spread
drawbacks of either of the approaches [46]. randomly over the image due to the curvature term not
permitting it in the evolving curve. This method was found to
III. ANALYSIS OF EXISTING METHODS perform well for both normal and tumor affected brain MR
The various methods used for the different stages of brain images. It is found to be accurate and robust to all inputs. But
MRI segmentation have their own benefits and drawbacks it fails for extremely noisy and low contrast datasets.
which are discussed in this section. C. Feature Extraction
A. Noise Removal Although DWT can be used to extract features from MR
Although Gaussian filters are easy to implement, they images, they exhibit the translation variance problem which
result in generation of blurred images that fail to highlight the makes them unfavorable. SWT overcomes the translation
edges from the other brain regions in MR images. Median variance problem of DWT.SWT can also be used to
filters are simple in nature and help preserve the edge accomplish enhanced noise removal when compared with
information and produce efficient smoothening of images. The DWT along with edge detection as it preserves all the features
searching of pixels within the search window employed by the of the original input. It is also found that DWT can be applied
Non Local Means method is an exhaustive process which only to images with a size equal to a power of 2 whereas SWT
increases the computational complexity and results in no allows application to images of random size values[33].
refinement of performance measure. Since the Edge based method depends on edge intensities,
Wavelet analysis helps to eliminate noise for the signals it performs poorly for noisy image and requires removal of
that could be discontinuous in both time and frequency noise for efficient performance. The texture based features
domain. The LMS method for noise elimination enjoys used in Appearance based method are insensitive to intensity
simplicity in both hardware and software implementations. It values and help to efficiently segment the images. The Phase
was found to be robust and helped achieve a good Congruency method is insensitive to intensity variations as it
performance measure with only concise information about the does not determine the edges based on intensity values.
actual input signals and the desired signals. The choice of the
step size and the length of the filter used can affect the
performance of the method. The Anisotropic Diffusion
method of noise elimination facilitates smoothening of image
regions that preserve continuity in the images while

ISSN 2277-503X | 2016 Bonfring

Bonfring International Journal of Advances in Image Processing, Vol. 6, No. 3, August 2016 26

D. Segmentation Selection of parameters that control the strength of spatial

Threshold based methods can fail for low contrast images. interaction is difficult. A very high initialization of parameter
The region based method of segmentation is simple and values can result in loss of structural information producing
effective in accurately segmenting regions of similar over smoothed segmentation results [43].
properties without being computation intensive. At the Due to the highly parallel nature of processing in Artificial
boundary of two different tissue classes, this method fails to Neural Network based approaches, they can turn out to be
correctly classify the image pixels as belonging to one of these very fast and efficient. But since they are usually implemented
two tissue classes due to the partial volume effect that blurs on standard serial computer, the computational advantage is
the intensity of pixels in this region. These methods require reduced [44].A combination of linear and non-linear
appropriate selection of the seed point in order to identify a transformations are used for the mapping in atlas guided
region of interest. It is also sensitive to noise and thus can approach. This approach thus enables efficient labeling of the
result in holes to be present in the segmented image or cause target image using the atlas. Accurate segmentation of
disconnected regions to be developed [36]. complex structures is difficult due to variations in anatomical
The classification method can be applied to images with structures over different subjects of a diverse population [45].
multiple classes. This approach requires manual intervention
to obtain the training data serving as a potential drawback. It IV. CONCLUSION
does not perform any spatial modeling and thus is sensitive to MRI image segmentation is one of the active areas of
heterogeneous intensities. It is also found that using separate research requiring high focus. This survey paper focused on
training set for each possible image to be segmented is a the various methods available to accomplish the different
tedious job and using a limited training set to segment large stages of the segmentation process. It also presented each of
number of input images can result in inaccurate segmentation their benefits and drawbacks. With each of these methods
results [37]. Clustering methods are sensitive to noise and having their own advantages and disadvantages, it is essential
intensity in homogeneities as they do not model spatial to develop efficient segmentation algorithms that can combine
interactions. But absence of spatial modeling has helped in these methods to result in efficient and effective segmentation
producing faster computation results [38]. results.
The constant propagation of contours in parametric
deformable model approach can result in inaccurate REFERENCES
segmentation when faced with missing boundary parts by not [1] Larry E. Davis and Sarah Pirio Richardson, Fundamentals of
being able to identify homogeneous regions of interest [39].To Neurologic Disease, New York: Springer, Pp. 175-187, 2015.
[2] P A McKinney, Brain Tumors: Incidence, Survival and Aetiology,
overcome this problem, Geometric deformable models are Journal of Neurology, Neurosurgery and Psychiatry, Vol. 75, No. 2,
used. They require placing an initial contour at the center of 2004, Pp. 12-17.
the region boundary for the region of interest and then [3] Raymond Liu, Margaretta Page, Karla Solheim, Sherry Fox and Susan
generating successive active contours using a level-set M. Chang, Quality of life in adults with brain tumors: Current
Knowledge and Future Directions, Journal of Neuro Oncology, Vol. 11,
function to obtain the final contour for the region of interest. No.3, Pp. 330-339, 2009.
This method poses difficulties in the practical implementation [4] R. Sauer, Therapy of Malignant Brain Tumors, Vienna: Springer,
as most of the segmentation methods are not capable of Pp. 195-276, 1987.
dealing with objects having regularity in their shapes [40]. [5] M. Kaus, S.K. Warfield, F.A. Jolesz and R. Kikinis, Adaptive template
moderated brain tumor segmentation in MRI, Proc. Bildverarbeitung
The K-Nearest Neighbor method is simple and easy to fur dei Medizin, Pp.102-106, 1999.
implement and turns out to be effective when implemented [6] M. Straka, A.L. Cruz, A. Kochl, M. Sramek, M.E. Groller and
D. Flieschmann, 3D watershed transform combined with a probabilistic
with efficient noise reduction methods. This approach can atlas for medical image segmentation, Proc. MIT, Pp. 1-8, 2003.
result in reduced performance with large trainingset as the [7] J.J. Corso, E. Sharon, S. Dube, S El-Saden, U. Sinha and A. Yuille,
performance is directly proportional to the size of the training Efficient multilevel brain tumor segmentation with integrated Bayesian
set and requires iterating over the entire set. It also requires model classification, IEEE Trans.Med. Imag., Vol. 27, No.5, Pp. 629-
640, 2008.
only limited features as input since classification needs to be [8] M. Prastawa, E. Bullitt, S. Ho and G. Gerig, A brain tumor
done for each and every pixel. It is sensitive to noise and segmentation framework based on outlier detection, Med. Image Anal.,
inhomogeneous intensities due to absence of spatial filtering. Vol. 8, Pp. 275-283, 2004.
Repeated iterations to improve the classification result can [9] H. Khotanlou, 3D brain tumors and internal brain structures
segmentation in MR images, Ph. D. dissertation, Informatique,
make the method slow. It is not favorable for MRI tumor Telecommun. Electron, Telecom ParisTechParis, France, 2008.
segmentation when the training set has some of the under- [10] M.R. Kaus, S.K. Warfield, A. Nabavi, P.M. Black, F.A. Jolesz and
represented classes [41]. R. Kikinis, Automated segmentation of MR images of brain tumors,
Radiology, Vol. 218, Pp.586-591, 2001.
The Bayesian approach considers tumor and edema to [11] S. Ahmed, K.M. Iftekharuddin and A. Vossough, Efficacy of texture,
follow a normal distribution which is not true. It also fails shape, and intensity feature fusion for posterior-fossa tumor
when the brain comprises of large deformations. But this segmentation in MRI, IEEE Trans.Inf. Technol. Biomed., Vol. 15, No.
2, Pp. 206-213, 2011.
method was found to be computationally efficient and [12] R.M. Haralick, K. Shanmugam and I. Dinstein, Textural features for
produced faster segmentation results [42].Since the statistical image classification, IEEE Transactions on Systems, Man and
model approach supports spatial modeling, it is found to be Cybernetics, Vol. 3, No. 6, Pp.610-621, 1973.
insensitive to noise and heterogeneous intensities. Because of [13] R.C. Gonzalez and R.E. Woods, Digital Image Processing, Pearson
Education, 2008.
the spatial modeling, this method is computation intensive.

ISSN 2277-503X | 2016 Bonfring

Bonfring International Journal of Advances in Image Processing, Vol. 6, No. 3, August 2016 27

[14] N. Zhang, S. Ruan, S. Lebonvallet, Q. Liau and Y. Khu, Kernel feature [36] K.P. Wong, Medical Image Segmentation: Methods and Applications
selection tofuse multi-spectral MRI images for brain tumor in functional Imaging, Handbook of Biomedical Image Analysis,
segmentation, Computer Image Vision Understanding, Vol. 115, Springer, Pp. 111-182, 2005.
Pp. 256-269, 2011. [37] O. Chapelle, B. Scholkopf and A. Zien, Semi-Supervised Learning,
[15] J.R. Jimenez-Alaniz, V. Medina-Banuelos and O. Yanez-Suarez, Data MIT Press Cambridge, Vol. 2, 2006.
driven brain MRI segmentation supported on edge confidence and a [38] Popuri K, Cobzas D, Jagers and M, Shah S and Murtha A, 3D
priori tissue information, IEEE Trans. Med. Imag., Vol. 25, No. 1, Pp. variational brain tumor segmentation on a clustered feature set, SPIE
74-83, 2006. Med Imaging, Pp. 14, 2009.
[16] W.E. Reddick, J.O. Glass, E.N. Cook, T.D. Elkin and R.J. Deaton, [39] A. Gooya, K.M. Pohl, M. Bilello, G. Biros and C. Davatzikos, Joint
Automated segmentation and classification of multispectral magnetic segmentation and deformable registration of brain scans guided by a
resonance images of brain using artificial neural networks, IEEE Trans. tumor growth model, in Medical Image Computing and Computer-
Med. Imag., Vol. 16, No. 6, Pp.911-918, 1997. Assisted Intervention-MICCAI 2011, Springer, Pp. 532-540, 2011.
[17] J. Alirezaie, M.E. Jernigan and C. Nahmias, Automatic segmentation of [40] C. Li, R. Huang, Z. Ding, J. Gatenby, D.N. Metaxas and J.C. Gore , A
cerebral MR images using artificial neural networks, IEEE Trans. Nucl. level set method for image segmentation in the presence of intensity in
Sci., Vol. 45, No. 4, Pp. 2174-2182, 1998. homogeneities with application to MRI, IEEE Transactions on Image
[18] P. Kovesi, Image feature from phase congruency, Journal of Computer Processing, Vol. 20, No. 7, Pp. 2007-2016, 2011.
Vision Research, Vol. 1, No. 3, Pp. 126, 1999. [41] S.U. Aswathy, G. Glan, Deva Dhas and S.S. Kumar, A Survey of
[19] T. Song, M.M. Jamshidi, R.R. Lee and M. Huang, A modified Segmentation of Brain Tumor from MRI Brain Images, International
probabilistic neural network for partial volume segmentation in brain Journal of Applied Engineering Research, Vol. 9, Pp. 9267-9275, 2014.
MR image, IEEE Trans. Neural Netw., Vol. 18, No. 5, Pp. 1424-1432, [42] Jin Liu, Min Li, Jianxin Wang, Fangxiang Wu, Tianming Liu and Yi
2007. Pan, A Survey of MRI-Based Brain Tumor Segmentation Methods,
[20] J.E. Lee, M.K. Chung and A.L. Alexander, Evaluation of anisotropic TSINGHUA Science and Technology, Vol. 19, No. 6, Pp. 578-595,
filters for diffusion tensor imaging, 3rd IEEE International Symposium 2014.
on Biomedical Imaging: Nano to Macro, Pp. 77-78, 2006. [43] J. Nie, Z. Xue, T. Liu, G.S. Young, K. Setayesh, L. Guo and S.T. Wong,
[21] M.R. Devi and S. Shenbagavadivu, A comparative analysis of median Automated brain tumor segmentation using spatial accuracy-weighted
filter on various types of images, International Conference on Research hidden MARKOV random field, Computerized Medical Imaging and
Trends in Computer Technologies, Pp. 24-28, 2013. Graphics, Vol. 33, No. 6, Pp.431-441, 2009.
[22] A. Buades, B. Coll and J. Morel, A non-local algorithm for image [44] J. Jiang, P. Trundle and J. Ren, Medical image analysis with artificial
denoising, IEEE Conf. on Comp. Vis. and Patt. Recog, Vol. 2, No. 11, neural networks, Computerized Medical Imaging and Graphics, Vol.
Pp. 60-65, 2005. 34, Pp. 617-631, 2010.
[23] Kedar Phatak, Swapnil Jakhade, Aniket Nene, R.S. Kamathe and [45] M. Cabezas, A. Oliver, X. Llado, J. Freixenet and M. Bach Cuadra, A
K.R. Joshi, Denoising of Magnetic Resonance Images Using review of atlas-based segmentation for magnetic resonance brain
Independent Component Analysis, Recent Advances in Intelligent images, Computer Methods and Programs in Biomedicine, Vol. 104,
Computational Systems (RAICS), Pp. 807-812, 2011. No. 3, Pp. e158-e177, 2011.
[24] I. Diaz, P. Boulanger, R. Greiner and A. Murtha, A critical review of [46] K.B. Vaishnavee and K. Amshakala, Study of Techniques used for
the effects of de-noising algorithms on MRI brain tumor segmentation, Medical Image Segmentation Based on SOM, International Journal of
Annual International Conference of the IEEE Engineering in Medicine Soft Computing and Engineering, Vol. 4, Pp. 40-44, 2014.
and Biology Society, Pp. 3934-3937, 2011.
[25] Shubhra Dixit and Deepak Nagaria, Neural Network Implementation of
Least-Mean Square Adaptive Noise Cancellation, International
Conference on Issues and Challenges in Intelligent Computing
Techniques (ICICT), Pp. 134-139, 2014.
[26] Gang Qiao, Wei Zhang and Sei-Ichiro Kamata, Anisotropic Diffusion
with Edge Projection, 4th International Congress on Image and Signal
Processing (CISP), Pp. 665-667, 2011.
[27] Shaswati Roy and Pradipta Maji, A Simple Skull Stripping Algorithm
for Brain MRI, Proceedings of ICAPR, Pp. 1-6, 2015.
[28] Rosniza Roslan, Nursuriati Jamil and Rozi Mahmud, Skull Stripping of
MRI BrainImages using Mathematical Morphology, Proceedings of
IECBES, Pp. 26-31, 2010.
[29] J.G. Park and C. Lee , Skull Stripping Based on Region Growing for
Magnetic Resonance Brain Images, Neuro Image, Vol. 47, No.4,
Pp. 1394-1407, 2009.
[30] K. Somasundaram and P. Kalavathi, A Hybrid Method for Automatic
Skull Strippingof Magnetic Resonance Images (MRI) of Human Head
Scans, Proceedings of ICCCNT, Pp. 1-5, 2010.
[31] Juan Eugenio Iglesias, Cheng-Yi Liu, Paul M. Thompson and Zhuowen
Tu, Robust Brain Extraction Across Datasets and Comparison With
Publicly Available Methods, IEEE Transactions on Medical Imaging,
Vol. 30, Pp. 1617-1634, 2011.
[32] Gayatri Mirajkar, Suhas Patil and Minal Pawar, Skull Stripping Using
Geodesic ActiveContours In Magnetic Resonance Images, International
Conference in CICSyN, Pp. 301-306, 2012.
[33] Carlos Arizmendi, Alfredo Vellido and Enrique Romero, Binary
Classification of Brain Tumors Using a Discrete Wavelet Transform and
Energy Criteria, Proceedings of LASCAS, Pp. 1-4, 2011.
[34] Yudong Zhang, Zhengchao Dong, Lenan Wu, Shuihua Wang and
Zhenyu Zhou, Feature extraction of brain MRI by stationary wavelet
transform, Proceedings of ICBECS, Pp. 1-4, 2010.
[35] Swati Khurana and Dr.M.L. Garg, MRI Based brain Tumor
Segmentation Methods: A Critical Review, International Journal of
Advance Research in Computer Science and Management Studies, Vol.
3, Pp. 432-446, 2015.

ISSN 2277-503X | 2016 Bonfring