2 LITERATURE SURVEY

2.1 Introduction

The last few years have seen a marked rise in medical

cases involving brain tumors. It is the 10th most common type of
tumor that affects both adults and children alike. Nevertheless,
if diagnosed well on time, it is also one of the most curable types
of tumors. Scientists and researchers have therefore been
working towards the development of sophisticated techniques
and methods for the identification of the tumor type and stage.

Brain imaging involves the use of different techniques to

view the structure, function or pharmacology of the nervous
system, either directly or indirectly. Within medicine,
neuroscience and psychology, it is a modern field. Magnetic
Resonance Imaging (MRI) is the commonly used for resection
and analysis of anomalies in terms of brain tissue form, size or
location, which in turn help to identify tumors.

2.1.1 Intensity inhomogeneity

The most vital limiting in MR image classification is the

intensity inhomogeneity, called the intensity non-uniformity or
shading. It alludes to the MRI artifact that the forces of a similar
tissue are not steady finished the picture spatial space, i.e. the
pixels of a similar tissue may display very extraordinary
intensities and the pixels from the diverse tissues may seem to
have comparative intensities.

At times, this intensity inhomogeneity may deliver over

30% varieties of image intensity for a similar cerebrum tissue,
and is particularly extreme for high-field MR system. "Body coil

23
images at 1.5 Tesla, for instance, regularly seem to have a
position subordinate power (shading). At field qualities of 3
Tesla, or more, the picture power consistency, even in the head
imaging, frequently is hazardous." The intensity inhomogeneity
fundamentally shows up on the pictures procured from a wide
range of MR scanners with various serious levels and can't be
considered as an irregular noise.

2.1.2 Partial volume effects (PVE)

Partial Volume Effect (PVE) is an artifact that one voxel in

the MR brain volume contains more than one tissue, and stem
from the insufficient resolution of MR scanner. This results in
the blur boundary between tissues and increases the
misclassification ratio.

2.1.3 Neuroanatomical Structure Complexity and Variation

The brain tissue, especially GM (also called the cortex),

exhibits significantly complex topological variance from one
person to another.

2.1.4 Noise

The random noise in the MR images causes the

misclassification and fragmented edges. The noise can be arised
due to the acquisition of the image from the MR image scanner
or compression of the image.

2.2 Image Classification

Classification utilizes information with known labels/
indices to classify images into hyperspace. Image features are
identified with intensity or other texture properties.

24
 Classification techniques can be both supervised and
unsupervised. By the manual collaboration that is difficult and
tedious. Additionally, the utilization of a similar preparing set for
an expansive number of images can prompt one-sided comes
about, which don't consider anatomical and physiological
inconstancy between various subjects.

2.3 Related work in Brain Tumor Classification

There are numerous algorithms accessible for image
classifying the medical images. In the following are some of the
main methods accessible for image classification are reviewed.

M. H. Fazel Zarandi et al [24] proposed an expert system

approach, type-II fuzzy logic based for image classification,
which results in better brain tumors image diagnosis. This
approach is the integration of all better image pre-processing
methods and paves way for type-II fuzzy logic development for
approximate reasoning.

M. Zarinbal et al [25] proposed a method-using image

processing proficient classification of type-II to classify
astrocytomas in brain images through MRI. The performance of
this system showed better results as consecutive planes to MRI
scan images are taken. This method produced an accuracy of
89%,recall of 89% and true negative rate of 90%.

Hesheng Wang et al [26] proposed multiscale fuzzy C-

means which is fully automated to classify brain tumors. This
paper introduces three methods, i.e., Conventional fuzzy c-
means, modified fuzzy c-means and multi scale fuzzy c-means as
well as performance comparisons among these methods on
McGill brain dataset.

25
 Patel S. A. et al., [27], have used clustering method for
brain tumor detection in abnormal MR images. The Fuzzy C-
means (FCM) algorithm is shown to be superior than other
clustering procedures in the form of segmentation competence
and identified as FCM produces better segmentation but it fails
to identify the tumor lesions

Shanthakumar et.al., [28] uses Adaptive Neuro Fuzzy

Interference System (ANFIS) in computerized seed point
assurance. The pixel drive of the proposed count is not subject
to the tumor sort. The tumor division comes to fruition are
surveyed in perspective of various criteria, measuring the
likeness Index (SI), Overlap portion (OF), Extra part (EF) which
contrasted with estimations of 0.817, 0.817, and 0.182
independently.

Fazel Zarandi M. H. et.al [29] have suggested a aggregating

the existing filtering methods based on fuzzy rule for Pre-
processing step. By using the Type-II fuzzy concepts the
Possibilistic C-Mean (PCM) method, Mahalanobis distance, and
Kwon validity index are extended are used for Segmentation.
Thresholding technique is used for Feature Extraction and
developed Brain MRI tumor grade classification system using
Type-II Approximate Reasoning method.

Pan Zhigeng et.al [30], have proposed a bayes-based

district developing calculation that appraisals parameters by
examining the attributes in neighbourhood areas and builds the
bayes factor as a grouping paradigm. Both of these techniques
are blundering sensitive in light of the fact that, a disgraceful or
bogus depiction of starting arrangement and wrong choice of the
seed picture will prompt affecting outcomes. Factual strategies

26
and fuzzy logic approaches are by all accounts reliable and are
the best contender for the substitution of the previously
mentioned procedures.

S. Thara and K.Jasmine [31] proposed a new algorithm to

classify a normal image and tumor image. Probabilistic neural
network and Fuzzy Probabilistic Neural Network Classifier for
Radial Bias function which yields an accuracies of 91.3%, 96.3%
and 97.2% respectively are used.

M. Monica Subhashini et al [32] proposed a machine

learning algorithm called naive bayes. Firstly classification is
done on SVM, Naive bayes, and learning vector quantization
based on accuracy. The naives bayes algorithm showed an
accuracy of 91%.

Yudong Zhang et al [33] classifies tumors as normal or

abnormal automatically and accurately by a neural network
based method. The computational time of an image is 0.0451s
and the accuracy over both training and testing phases are
100%. The dataset is collected from the Harvard Medical School
website.

Yuehao Pan et al [34] introduced a grading system of brain

tumors through convolutional neural networks. They also
compared these with neural networks in terms of sensitivity and
specificity. This showed 18% improvement of performance
through CNN. The dataset used is BRATS 2014.

El-Sayed Ahmed El-Dahshan et al [35] proposed a hybrid

technique for MRI classification using ANN and KNN as
classifiers. The accuracies obtained for ANN and KNN are 97%

27
and 98% respectively. The dataset used is collected from the
Harvard Medical School website.

B.Sudha et al [36] classifies brain tumor grades using

FFNN, MLP and BPN. These three methods are compared
through their performance metrics like accuracy, specificity and
sensitivity. BPN tops the accuracy of 96.7%, specificity of 84%
and a sensitivity of 72%. The accuracy completely depends on
optimal features extracted in the training phase.

Eltaher Mohamed Hussein et al [37] proposed an ANN

technique for brain tumor detection. This work also involves
comparisons between different ANN techniques. The three ANN
techniques are BPNN, RNN and Elman Network. Out of these
Elman Network stood first with an overcoming performance ratio
of 88.24%.

Madhubanti et al [38] proposed an algorithm to classify a

normal brain and a person’s brain with Alzheimer’s disease. The
binary classification is achieved by supervised neural network
based classifier. Through this an 100% classification accuracy is
obtained. The dataset is collected from the Harvard Medical
School website.

Y.Zhang et al [39] proposed a forward neural network for

automatic and accurate classification of MRI brain images as
normal and tumorous. The features are extracted from an image
using wavelet transform and dimensions of features are done by
principle component analysis (PCA). After passing these features
into FNN the optimization is done by ACPSO i.e., Adaptive
Chaotic Particle Swarm Optimization. The obtained classification
accuracy is 98.75%.

28
 Sandeep Chaplot et al [40] taking wavelets as an input
proposed a method using both SVM and neural networks. This
method classifies images as normal or abnormal. The accuracy
obtained by SVM is 98%, which is more than that of self-
organizing maps i.e., 94%.

Salim Lahmiri [41] proposed a method for glioma grades

detection using particle swarm optimization techniques. After
segmenting by particle swarm optimization techniques some
multi-fractal features are obtained. These features are used for
the classification process. There are three types of particle
swarm optimization techniques used: PSO, DPSO and FODPSO.
The obtained accuracy is 98% and sensitivity is 100%.

Deepa Subramaniam Nachimuthu et al [42] proposed an

extreme machine learning particle swarm optimization classifier
to improve the classification accuracy to classify low grade and
high grade gliomas. Performance validation of this method is
done by a test set. The accuracy obtained by co-occurrence
matrices is 86.5% and with both volumetric features and
spectroscopic features is 99.15%.

Abdu Gumaei et al [43] proposed a method in which

features are extracted using a novel and efficient hybrid method,
referred to as PCA-NGIST. Finally, brain tumors are classified
usingRLEM classifier.

Sudipta Roy et al [44] proposed a classification algorithm

using ANFIS classifier and compared with other two classifiers
namely: ANN with back propagation Learning model and K-
nearest neighbors. The accuracies obtained are 95% for ANFIS
classifier,90%for ANN and 90% for KNN.The dataset is collected
from Med Harvard.
29
 Minakshi Sharma et al [45] proposed a method to classify
between Astrocytoma tumors and a hybrid genetic technique for
brain tumor segmentation. The obtained accuracy is 97%,
sensitivity is 96.9% and specificity is 95.6%. The dataset is
collected from a web source.

Nooshin Nabizadeh et al [46] proposed a classification

method using four techniques, namely: SVM, KNN, NSC, SRC
and k-means. Out of these SVM produces the highest accuracy
of 94.8%, then KNN produces an accuracy of 89.6% and lastly
NSC gives an accuracy of 63.5%. The dataset consists of T1-w
brain images and FLAIR sequences.

Megha.P.Arakeri et al [47] proposed a classification

algorithm using SVM, ANN and KNN techniques. Classification
accuracy of SVM is 81.37%, ANN is 75.04% and KNN is 70.46%.

Betsabeh Tanoori et al [48] proposed a technique based on

SVM and active contour models to find brain structures in brain
MR images. Depending on the extracted features SVM classifier
gets trained for each brain tissue. Validation of given method is
done by gold standard brain MRI dataset.

Frank G Zolner et al [49] proposed an SVM algorithm

which includes feature reduction analysis. The processing time
of the algorithm has been measured in which speed is increased
as inputs to SVM are decreased. In this method SVM is selected
with a Radial Bias Function.

Evangelia I.Zacharaki et al [50] proposed a machine

learning scheme using support vector machine to classify grades
of gliomas. The binary SVM classification gives an accuracy,
sensitivity and specificity of 85%, 87% and 79% when classifying

30
metastases from gliomas and 85%, 88% and 96% when
classifying high grade gliomas from low grade neoplasms. Multi-
classification is also been carried out through this method.

Selvaraj et al. [51] have proposed a propelled grouping

procedure in view of Least Squares Support Vector Machines
(LS-SVM) and connected to cerebrum picture cuts arrangement
utilizing highlights got from cuts.

Gayatri Chavan et al., [52] have proposed the SVM (bolster

vector machine) classifier took after by KNN. Segmentation based
Fractal Texture Analysis or SFTA calculation are utilized to
remove the surface component from pictures. The
characterization is done on the database to arrange cerebrum
MRI into three classifications, for example, normal, benign and
malignant.

Zacharaki et al. [53] have proposed MRI-based

characterization of brain tumor sort utilizing SVM-RFE. The
evaluations of brain tumor locales are broken down and
analyzed in gentle, direct and extreme.

Jainy Sachdeva et.al., [54] develop a natural CAD system

for helping radiologists in multiclass mind tumor classification.
The tumor ranges are separated by Content Based Active
Contour (CBAC). The ranges are then spared as Segmented
Region of Interest (SROIs). Intensity and surface rundown of
abilities are isolated from these SROIs. The elements are
especially picked in light of the obsessive purposes of
enthusiasm of cerebrum tumors given by the radiologist. Genetic
Algorithm (GA) picks the arrangement of ideal features from this
information set. Two blend machine learning models are realized

31
using GA with support vector machine (SVM) and artificial
neural framework (ANN).

Quratul Ain et.al., [55] suggested a computerized system

for the finding of the brain tumor. The Proposed framework is
multi-arranged framework for brain tumor conclusion and tumor
district extraction. The Next phase of the proposed framework is
an arrangement that depends on these extricated features. The
Ensemble based SVM arrangement is used. More than
99%accuracy is accomplished by the classification. After
classification, the proposed method extricates tumor district
from tumorous images utilizing multi-step segmentation.

El-Dahshan et al [56] proposed a method-using computer

aided diagnosis through MRI to treat brain tumors. This
technique is a hybrid machine learning technique for brain
tumor classifications and gives accurate results in minimum
time. The proposed algorithm gives an accuracy of 99% on both
training and testing set, sensitivity of 100% and specificity of
92%. The experiments are made on Brain tumor MRI dataset
collected from the Harvard Medical School website.

K A Smitha et al [57] proposed three significant methods

for glioma classification and characterization. The three methods
are :rPSR, rCBV and rCBF and the variation of values between
LGG and HGG are statistically significant. The ROC analysis for
each of them will be 96% recall and 71.8% true negative rate,
88% recall and 69.2% specificity and 72.5 recall and 66.7% true
negative rate.

Norhashimah Mohd Saad et al [58] proposed a paper

comparing Neuro-imaging techniques, MRI techniques and CAD
techniques for classifying brain tumors into benign or malignant
32
or infections. Brain imaging helps physicians for better
treatment and diagnosis. The dataset for this imaging technique
trials are available online.

Arti Tiwari et al [59] proposed an paper which includes

comparisons of different brain tumor classification methods like
SVM, fuzzy c-means, Particle swarm optimization and neural
networks. The performance of these methods is evaluated by
confusion matrix, accuracy, sensitivity, specificity and f1 score.
The dataset is used from different sources like Local, Harvard,
Rider, MICCAI etc.,

Massimo Caulo et al [60] proposed a data-driven

methodology for classification of glioma grades. A multi model
magnetic resonance imaging technique is used and compared
with current radiological reporting methods. The obtained AUC
is high i.e., 0.95 which helps physicians and patients for fast
treatment.

Karimaghaloo Z. et.al., [61] has implemented Continuous

Random Field (CRF), Temporal hierarchical adaptive texture-
based method to analyze the image variations. Results show that
the large training data limited the classification performance

Xiao Xuan et al., [62], have proposed an AdaBoost

algorithm to perform and to select the most discriminative
features and classify tumor and normal structures.
Classification accuracy is improved. Large training data limited
the performance.

Anitha V. et al., [63], have proposed integration of adaptive

pillar K-means algorithm with the two-tier classification

33
approach and found better computation efficiency and
segmentation quality due to two tier classification.

Huang M. et al., [64], have proposed Local Independent

Projection-based Classification (LIPC) and discussed better
classification due to reducing the feature vectors. This algorithm
fails to classify some lesions due to unreliable retrieval
misalignment.

Bezdek J. C. et.al [65] exhibited Supervised and

Unsupervised strategies for picture preparing an actual point by
point survey of these systems.

Babak Afshin-Pour et al., [66] have presented image

Classification method in the fMRI brain image. In this study,
they state that a hybrid combination of LDA and GCCA to get the
better classification rate than traditional algorithms. In this
study, with the help of feature optimization, it improves better
Classification result than only by estimating pixel variation and
also provides correct treatment to the patient.

Harati et al. [67] have proposed a PC helped arrangement

technique which joins customary MRI and perfusion MRI for
differential analysis. Their technique comprises a few stages,
including ROI definition, feature extraction, feature choice and
classification.

The separated components incorporate tumor shape and

intensity attributes and in addition rotation invariant surface
elements. Elements subset determination is performed utilizing
support vector machines (SVMs) with recursive element disposal.
The twofold SVM characterization exactness, affectability, and
specificity, evaluated by forgetting one cross-approval on 102

34
brain tumors, are individually 87%, 89%, and 79% for
segregation of metastases from gliomas, and 87%, 83%, and 96%
for separation of high review from poor quality neoplasms.

Pim Moeskops et.al [68] have developed Automatic

segmentation in MR brain image is critical for quantitative
investigation in huge scale thinks about with image procured at
all ages. This work shows a strategy for the automatic
segmentation of brain MR image into various classes of tissues
by a CNN. The technique is not reliant on unequivocal
components, but rather figures out how to perceive the data that
are essential for classification based on the training data set.

Xavier Tomas-Fernandez et al., [69] propose another

calculation that accomplishes lesion and brain tissue
segmentation through the synchronous estimation of a spatially
global inside the-subject intensity conveyance and a spatially
nearby intensity dispersion got from a sound reference populace.

Reynoso J. L. et.al. [70] have proposed artificial intelligence

apparatus and, in result, lessens era of futile information. It
focuses on the representation of magnetic resonances of different
parts of the human body, such as knees, spinal column, arms,
elbows, etc., using ontologies.

Tianming Zhan et.al.,[71] have proposed level set strategy

here is connected in light of local Gaussian dissemination to
accurately recognize the limits of the white issue lesions in the
ROI. The level set technique in view of nearby Gaussian
dispersion fitting vitality is vigorous to the force inhomogeneity of
MR information and along these lines prepared to do decisively
extricating the limits of white issue sores.

35
 Cheng H. D. et.al. [72] have introduced Computer Aided
Diagnosis framework (CAD) for bosom growth pictures and
comprises preprocessing, division, highlight extraction, including
characterization and arrangement.

Yaniv Bar. et.al [73] have Presented a method for medical

application of identification of chest pathology in radiographic
images using CNN trained from a non-medical data collection. In
this case the descriptor for the Decaf5 baseline is the leading
representation.

Bram van Ginneken. et.al [74] proposed a method for

features from pre-trained CNN OverFeat, trained to detect
objects in natural images are used in computed tomography
scans for nodule detection. For each nodule candidate, the 2D
sagittal, coronal, and axial patches extracted 4096 features from
the penultimate layer of OverFeat and identified them with linear
vector support machines.

T. Schlegl et.al. [75] Proposed and evaluated a semi-

supervised learning approach, using multi-site data. By
Adaptation of the domain via unsupervised pre-training of
convolutional neural networks to inject image classes for which
no annotations are available.

Hoo-Chang Shin et.al. [76] proposed a method by taking

advantage of three important factors of using deep neural
convolution networks for classification of thoraco-bdominal
lymph node (LN) and interstitial lung disease (ILD).

Uri Avni et.al. [77] presented an efficient image

classification and retrieval system for medical image databases,
especially large x-ray archives. The methodology is based on a

36
local patch representation of the content of the image, using a
visual word bag approach and SVM for classification.

Gustavo Carneiro et.al. [78] proved that CNN models are

pre-trained with computer vision databases In medical imaging
applications. With a focus on the classification of mammograms
using craniocaudal (CC) and mediolateral oblique (MLO) views
and their respective mass and micro-calcification segmentations
of the same breast.

Hao Chen et.al. [79] proposed a learning-based approach

for fetal abdominal standard location in US images, plane (FASP)
constructs a deep convolutional neural network transferred
domain. To reduce the problem of overfitting caused by the small
amount of training samples.

2.4 Conclusion
Traditional image processing and machine learning
techniques follow the steps like pre-processing, feature
extraction, feature selection and classification. Pre-processing
stage typically includes the elimination of noise, skull-stripping
and correction of bias in intensity.

Image processing techniques are used after pre-processing

to extract features that usually describe each distinct tissue.
Some examples include features such as intensity, texture,
asymmetry-related features; discrete wavelet transforms (DWT),
textures, multi-fractal brownian motion features, first-order
statistical features, speed gradients, and edge-based features.
Different kinds of classifiers are implemented by using these
features; like SVM, Neural Networks (NN), KNN, SOM, RF and so
on.

37
 Research work shows a better performance when used a
CNN-based models [80], particularly in the classification of 2D
data. CNN's benefit is that each kernel is learned spontaneously
in different layers, so there is no need for the feature setting in
advance, which makes the number of training examples
important.

Contrary to conventional methods of classification in which

features extracted from images are fed to the networks,
Convolutional Neural Networks (CNN) automatically learn from
the data itself representative complex features directly. Deep
learning models are becoming more popular in recent years in
classification tasks. Medical image classification with deep
learning models at different levels can improve the prediction
task and accuracy as well.

38