``

Chapter-2

Literature Survey
2.1 Review of MRI Segmentation and Classification Techniques
In recent days segmentation scheme utilizes membership grades of components to
relegate pixels into auspicious segments which is predicated on a rudimental section growing
method. “FCM” algorithm attains consistency and contextual information from the facile
images, but disastrous in the case of composite strepitous images.

Image segmentation has been used for detection of encephalon tumor which is an
astronomically consequential process for cull right therapy at the precise time. Different
classifiers proposed for relegation such as FCM, SVM, ANN, expectation-maximization
procedure to extract the paramount information from the medical imaging procedures. An
overview and findings of a few researches are presented.

W. M. Wells et al.(1996)[33] offered Intensity-predicated classification of MR images

intra-scan in homogeneities method with utilization of the EM algorithm to a fully automatic
process for segmentation of tissue from MRI data.

Yannis A. Tolias et al.(1998)[34] presented a novel flexible fuzzy clustering

proposal for image segmentation with spatially varying sample vectors for different window
dimensions in which amultire solution representation is utilized.

Further,Yannis A. Tolias et al(1998)[35],proposes a new approach for enhancing the

outcomes of fuzzy clustering by striking spatial constraints for resolving image segmentation
troubles. The author have evolved a Sugeno-type rule-based scheme with three inputs and 11
rules that cooperates with the clustering outcomes obtained by the popular FCM and/or PCM
algorithms which offers good image partitions in terms of region softness and elimination of
the consequence of noise. The results of the suggested rule-based neighborhood development
(RB-NE) system are measure up to well-known segmentation algorithms via stochastic field
modeling.

6
``

Dzung L. Pham et al.(1999)[36]presented fuzzy segmentations of images that are

created with modification of the impartial function in the FCM algorithm, which permits
every class to vary over the image through the centroids.
A.W.C.Liew et al.(2000)[37] utilized the spatial background information utilizing a 3
x 3 window. The algorithm is compatible to the image comfortable in the logic that influence
since the neighboring pixels is concealed in non-homogeneous areas in the image. A cluster
merging proposal that merges two clusters depend on their closeness and their level of
overlap is presented. Investigational results with synthetic as well as genuine images
designate that the proposed algorithm reduces noise and more proficiently adept at
determining relegation ambiguity and duplication with different cluster assortment than the
prognostic able FCM Algorithm.

Mohamed N. Ahmed et al.(2002)[38]presented an algorithm by modifying the

objective functions of the typical FCM algorithm to recompense by the label in its instant
neighborhood. This acts as a regularized for salt and pepper noise.

Dzung L. Pham et al.(2002)[39] presented a new approach to fuzzy clustering which

leads to a constant algorithm by utilizing consequence term. To resolve the strength of the
consequence term, a criterion depend on cross-validation is employed, which reduces the
noise from images.

Yanling Li et al.(2010)[40] presented an expeditious FCM clustering algorithm

through spatial constraints (FFCM-S)to surmount the quandary that FCM-S algorithm faced a
time consuming aspect. The FFCM-S algorithm has employed to have better convergence
capability.

Damodharan et al. (2016)[41] proposed a encephalon tumor apperception and

relegation neural network predicated method. In this process, the quality rate is composed
discretely for sectionalisation of gray matter (GM), white matter (WM), and cerebrospinal
fluid (CSF) as well as tumor region and claims an exactness of 83% utilizing NN predicated
analysis.

7
``

Salem et al. (2016) [42], utilized SVM, fast Fourier transform (FFT) and Minimal
Redundancy-Maximal-Relevance (MRMR) procedure to improve the meticulousness of the
classifier. This technique obtained an accuracy of 98.9%.

Chaddad et al. (2015) [43], employed Gaussian mixture model (GMM) for feature
removal from MR images. The performance of the GMM feature extraction is improved by
using PCA and wavelet based characters, and achieved an “accuracy” of 97.05% for “T1-
weighted” with “T2-weighted” and 94.11% for “FLAIR weighted MR” pictures are obtained.
Zanaty et al.(2012) [44], proposed a hybrid FCM and Jaccard resemblance coefficient
algorithm to determine segmented gray matter and Kumar and Vijayakumar et al(2015) [45],
uses PCA along with SVM and obtains similarity index of 96.20%. Cui et al. (2013) [46],
appealed a localized fuzzy clustering scheme, by utilizing Jaccard resemblance index as
quantity of the segmentation correctness and claim 83% to 95% accurateness. Wang et al.
(2014)[47] have obtained a segmentation method predicated with the quandary of intensity in
homogeneities in image partition.

Sachdeva et al. (2013) [48], utilized a dataset of 428 MR pictures for a multiclass
brain tumor relegation, segmentation, with feature extraction. In this scheme, ANN, PCA-
ANN has been employed and found the growth in relegation accurateness from77%to91%.

Nilesh Bhaskar rao Bahadureet et al. (2017)[49], has obtained dice similarity
catalogue, which is individual of the important factors to find the accuracy segmentation and
support vector machine on behalf of classification and attained 96.51% accuracy, 94.2%
meticulosity, and 97.72% sensitivity.

Azian Azamimi Abdullah et al. (2012) [50], presented the diagnosis of the brain
tumour in MRI images using Graphical User Interface (GUI) in MATLAB by cellular neural
network (CNN) algorithm. CNN simulator is capable to detect the location of brain tumor
fruitfully with in shorter time and reported 100% confirmation of the present of tumor.

Ramzi A.Haraty et al.(2015)[51], presented an improvement to k-means clustering,

that can cluster enormous data and a distributed form of the 𝐺-means algorithm has been
applied to improve the computational power.

8
``

DinhThuan Nguyen et al.(2013) [52], presented a superior algorithm on behalf of

learning k means clustering to improve performance based on speed.

Christos Boutsidis et al. (2015)[53], presented the study of dimensionality reduction

for k-means clustering which creates a small set of features and provides constant-factor
estimation regarding k-means objective assessment. Mingjun Song et al.(2015)[54], proposed
three steady approximation algorithms for k-means clustering. In this study, they have
projected three O(1)-approximation algorithms whose run time is self-determining for the k-
median difficulty.
EffatNaaz et al. (2016)[55], applied the K-means to acquire different views of medical
notes which progress the visibility of medical note. This observation shows in amplify of the
accuracy of medications linked with symptoms.

Leonard WafulaWakoli et al. (2014) [56], presented a customized K-Meanss

algorithm to flag out distrustful claims.

Kai Zhang et al. (2017) [57], presented high usefulness in image denoising
responsibilities, by training Deep CNN model. .Adrian Ion-Margineanu, et al. [58], presented
on finding a classifier that separates between tumour succession and regression regarding
multi parametric MR data recovered from follow-up GBM victims. It also determines the
performance by calculating BER for each time point and wBER for every time points.

William Thomas H.M (2015) [59], presents MRI image for tumor detection using
morphological filtering which is removed from the input image and also presented based on
morphological operation and segmentation algorithm. The experiment was conducted for the
detection of tumor by using morphological operation to image and later segmentation in
particular watershed for s canned MRI image from human brain for different samples the
algorithm.Sayali, V Roeli, et al. (2015) [60], presented how to extract easily information and
various features from the images segmentation for detection of brain tumor. Garima Singh, et
al. (2016) [61], presented a new idea for recognition of the brain cancer using normalized
histogram and segmentation with K-means clustering algorithm and also proposed well-
organized Naïve Bayes Classifier and Support Vector Machine (SVM) for accurate prediction
and classification.

9
``

Nabizadeh, et al. (2015) [62], presented a novel technique which includes

Normalization of Histogram along with K-means Segmentation to eliminate the unwanted
noise and extract the tumor from the MR image. Krupali D. Mistry, et al. [63], presented an
approach of detecting Enchondroma bone cancer from MRI images via image processing,
segmentation clustering methods, i.e. K-means integrated through FCM clustering.

PriyaPatil, et al.(2012) [64], presented two methods that are used to plan the image in
MRI. By this assist of design how to distinguish the boundaries of brain cancer and calculate
the genuine area of tumor and also the f-transform is worn to give the definite information
like rebuilt of misplaced edges and extracting the hushed edges. Accuracy and simplicity in
an MRI Images is reliant on each other. KailashD.Kharat, et al.(2012) [65], proposed an
approach which utilizes a blend of Multi-Layer Perceptron (MLP), Radial Basis Function
(RBF)procedures for segmentation, quality vector extraction and model learning. These two
techniques points out distrustful regions that have related property as the tumor regions.

Zhongyu Li, et al. (2017)[66], introduced the evaluation set of rules with a variety of
investigative and diagnostic scenarios in the directions of large-scale retrieval, which can
more improve the enactment of medical image investigation. T. Sathies Kumar et al. (2017)
[67], presented the encephalon tumor detection that allows confining a mass of abnormal
cells in a slice of MR using “SVM Classifier” and segmentation to know about the dimension
of the tumor in that segmented area, also the extracted features of the segmented portion will
be trained using ANN to display the category of tumor.

Emre Dandi, et al. (2017)[68], developed software which is intended with a user
interface for this reason. Thus the case of culpable by physicians will be offered.
Consequently, the application software will avoid errors and may be worn as a secondary
means for brain tumor segmentation. It has been shown in detailed test experiments on image
dataset that designed application can detect brain tumors successfully.

Aqhsa Q, et al. (2014) [69], presented classification by taking input as energy,

entropy, contrast and a few other statistic features like mean, median, variance and
association for MLP and they have also shown a feature selection method to decrease the
feature space by using neural network for classification.

10
``

M.Jasmine, et al.(2012) [70], presented the inspection of the tumor segmentation in

every year that is established for brain tumors might be easily noticed through MR images
and showed diagnosis along with analysis will be supported by the level of segmentation
accurateness. They have also explored the MR images of the encephalon tumor segmentation
techniques and imaging modalities and they also showed how the operations progression
before segmentation is examined with the comparison of dissimilar brain images and the
virtues of different approaches are occupied into account.

Nicolas Sauwen, et al.(2015) [71], presented a technique for detection of tumor

substructures (i.e. possible tumor, necrosis and edemas from twenty-four glioma victims.
Identifying tissue-specific patterns across patients is hampered due to the coexistence of
dissimilar tissue types, specially when mixed “tissue” classes are considered.

Neha Rani, et al. (2016)[72], presented neural network with back propagation to
categorize the performance of tumors component and showed that this technique results high
accurateness and apperception which further minimizes the spending time.

Sanghamitra T. Kamble, et al.(2015) [73], presented different types of tumors some

and applied K- means algorithm which is sufficient to extract from the encephalon cell. The
noise free image is specified as a input to the “k-means” and tumor for accurate partition of
brain tumor.

WeidongJi, et al. (2016)[74], presented the basic principles of RBF, and then
comprehensively accounts the RBF domestic and overseas study status, discussed the
supremacy and disadvantages of different methods, and collates the explanation, also
analysed the existing problems and the development trend in the future.

M. Masroor Ahmed et al.(2015) [75], presented anisotropic dissemination model for

image improvement and K-means clustering method for grouping tissues. The proposed
technique considered weighted gray level intensity images and formed appreciative results.

Piyush M. Patel, et al. (2013)[76], presented an algorithm that utilizes K-means

algorithm in cluster investigation, particularly for huge volume dataset in partitioning to
adoption of multi-sampling.
11
``

Malathi R, et al. (2015)[77], presented image segmentation takes part a significant

role in image dispensation as it helps in the removal of suspicious regions since the medical
images. They have proposed an algorithm for tumor detection regarding segmentation of
brain MRI images using K-Means clustering.

Nimeesha K M, et al.(2013) [78], presented that the execution time is less in K-Means
compared to FCM clustering technique, because the amount of iterations of K-Means is less
than FCM clustering and also reported that the tumour area is calculated for some accurate
result. Also presented that the results from K-Means clustering method is superior than the
Fuzzy C Means method because, fuzzy C Means is semi-supervised method. Therefore, pre-
processing is required, Whereas K-Means Clustering does not require pre-processing since it
is unsupervised method and number of iteration is less. Maximum lossless compression is
achieved by K-Means Clustering. It also provided accurate results with minimal amount of
data.

G. Santhosh krishnan, et al. (2016)[79], used median filter to eliminate noise from the
image, GLCM is used to remove the needed features from the image to locate whether the
tumor is there or not, then used fuzzy c-means for image partition and artificial neuro fuzzy
skill and also ANFIS is used for categorization of type of brain tumor. The research focussed
on the exposure of brain abnormalities along with tumor from the MRI images and to classify
which type of brain tumor it is. Also focussed on a variety of image pre-processing methods
for managing the medical images, then image partition is used to successfully segment tumor
elements from rest of the brain image so that the tumor is detected from the segmented image
and then the feature extraction methods are studied.

V.Sagar Anil Kumar, et al.(2014) [80], presented relate degree economical K-Means
for Morphological Image Processing (MIP). Segmentation resolves separating an image into
displace in varied regions of a medical image.

Yudong Zhang, et al. (2016)[81], presented how to notice abnormal brains in the early
hours by applying wavelet-energy (WE). The wavelet-energy utilized feature description and
automated unusual detection as well as quantum-behaved particle swarm optimization to
optimize the weights of the SVM.
12
``

MingwenZheng et al.(2014) [82], presented superior density method to find out the
RBFNN’s data centre regarding the improvement of density technique and estimate the
expansion constant of every centre, then trains the network with the method of gradient
descent scheme.
EmanAbdel-Maksoud et al.(2015) [83], presented K-means clustering method
integrated with FCM algorithm to achieve accurateness, processing time, along with
performance.

P.Maffezzoni, et al.(2012) [84], introduced a robust deterministic annealing procedure

to determine the RBF centres which improves good noise rejection capability.Yiping Jiao et
al. (2015)[85], proposedan preliminary cluster centres selection approach for K-means
algorithm to enhance the performance of RBFNN. Antonios D, et al. (2011)[86], presented a
hybrid clustering algorithm that combines the merits of fuzzy and crisp clustering to train
Gaussian RBFNN. C.V.Subbulakshmi, et al.(2015) [87], presentedPSO algorithm among the
extreme learning machine classifier for five UCI Machine Learning Repository for handling
medical dataset.

MohdFauzi Bin Othman et al.[88] in 2011, reported the performance of classification

results of brain cancer using wavelet (Daubechies (db4)) with Support Vector Machine
(SVM) and found accuracy of 65%. MohdFauzi Othman and MohdAriffananMohdBasri,
2011, [89], used Principal Component study and Probabilistic Neural Network (PNN) and
reported precision of 73 to 100% with varying spread values from 1 to 3.

Rezaei and Agahi et al. (2017)[90] proposed weighted SVM and achieves an
exactness of 89.92% for cancerous tumors. orheim et.al [29], uses SVM, “texture features”
and attained a classification “accuracy” of 87% for “3DMR” images. Krishna et al.(2018)[91]
proposed the hybrid replica PSO- LLRBFNN Algorithm for habitual brain tumor detection
and achieves an accuracy of 98% for ADNI dataset.

There are diverse hybrid models like LLRBFNN model with TLBO weight
optimization [92] for power signal classification, financial predicting etc. Further the hybrid
optimization algorithm PSO-GAby Garg [97] and De Fátima Araújoa and Uturbey [98]

13
``

utilizes “PSO-DE(PSO-Differential Evolution)” for dispatch of generation , Dipankar Santra

et al. (2016)[99], utilizes PSO-ACO for economic load dispatch dilemma.

In this research, the performance of combination of biologically inspired hybrid

algorithms utilized to optimize the weights of Radial basis function network, LLWNN,
LLRBFNN model to get better accuracy.

14