CHAPTER 1- INTRODUCTION

1.1 Background

A brain tumor is a gathering or set of irregular cells within the brain. The cranium could

be inflexible wherein the brain is encompassed. There are various issues occurred due

to growth of any tumor inside this restricted space. The malignant and benign are two

main categories of the brain tumor. The skull is pressurized to enlarge from inside in

case of growth of any benign or malignant tumor. This tumor leads to cerebrum lesion

and that could be dangerous to existence also. The brain tumor is divided into two

kinds - primary or secondary. The tumor which happens in the cerebrum is known as

primary brain cancer. Various gray matters are gentle. An optional cerebrum tumor is

additionally metastatic mind tumor. This tumor starts because of spreading of disease

cells spread in the cerebral matter as of an additional limb in which lung or bosom is

included. The encephalon could start inside the brain or it spread to from the rest of the

organs of anatomy. It can broaden to the cerebrum. The growth rate and the position of

a brain tumor investigate its impacts on the function of nervous system. The kind of

brain tumor and also its size and location have assisted in prescribing the treatment

options of brain tumor [1, 64].

At Present, digital images have tremendously influenced and impacted modern society

in multiple ways. Digital image processing has emerged as a vital tool in science

and technology. The whole arena of digital image processing with different techniques

has shown its applicability to the study of medicine. The popular term "medical image

processing" represents the use of digital image processing for medical applications.

Magnetic Resonance Imaging(MRI) is an sophisticated clinical imaging approach.

This approach delivers important information regarding the anatomical structure

1
of human soft tissues. This algorithm is more advantageous than its counterparts since

it offers 3D (three-dimensional) information with better disparity amongst soft matters.

In recent times, detection and categorization of brain cancer is a popular investigatation

scope in scientific or clinical Image Processing. The manual way of brain tumor

detection by doctors is extremely tricky and needs plenty of time. Hence, automatic

brain tumor detection and classification needs to be implemented for avoiding miss-

classification and saving time. In the midst of various imaging methods, the use

of MRI is quite prevalent intended for brain tumor recognition due to its better firmness,

excellent capability to highlight change.

1.2 STAGES INVOLVED IN BRAIN TUMOR DETECTION

The procedure of brain cancer detection is categorized generally into four phases after

the selection of MRI pictures. These phases include pre-processing, cleavage,

characteristic withdrawal and categorization or clustering. A block diagram

describing the above four stages is shown in the Figure 1.1.

Fig.1.1: Segments included in Brain Tumor Recognition

2
The quality of the brain images can be improved by applying different image

preprocessing techniques. In segmentation process, an image is segmented in different

regions. Afterimage segmentation, feature extraction is to be done to extract the

pattern of assessable characteristics or features of the image. Different features i.e.

statistical, texture, on the basis of color and image, etc. are extracted. Classification

methods are executed for classifying affected and non-affected pictures of brain [2].

These methods identify whether the abnormality present in the anomalous pictures is

of malignant (cancerous) or benign (non-cancerous) type.

1.2.2 IMAGE REPROCESSING

This step makes use of MR images for increasing the chances of suspicious region

detection. This step focuses on improving minor image details and denoising of acquired

images. Noise degraded medical MR images are generally less accurate. In general, MR

images degraded by a unwrinkled signal strength dissimilarity known as strength

non-consistency is as well identified as RF inhomogeneity or bias field. This variation

reduces top frequency picture properties including its boundaries and outlines, in

addition makes images blurry. Hence, different image pre-processing methods are

applied for improving the petitioning and characteristic recognition progression of

contribution MRI brain amount s [3]. In this step, various filtering techniques are used

for making images noise free. These techniques suppress the undesired distortions and

improve some image attributes. The key motive of pre- processing is to eradicate image

noise and enhance image contrast level for improving the quality of acquired image. the

tasks performed by preprocessing are:

a. RGB to Gray scale conversion: A gray scale image merely holds gray range

values while MR images consist of RGB content. In ‘gray’ color, all components

including red, green and blue have equal brightness in RGB space. Therefore, it

3
 is essential to assign an individual brightness level to every pixel. MR scans

appear to be black & White but they include some primary shades (colors).

Therefore, conversion of MR images into gray scale images is a must.

b. Filtering: Filtering operation aims to modify or improve an image. Image

processing tasks performed using filtering includes edge enhancement,

sharpening, smoothing etc.

c. Contrast enhancement: the main motive behind improving the image contrast

level is to make unclear image parts clear and accentuate certain image features.

Contrast enhancement is usually performed by eradiating noise and blurriness

from images. Contrast enhancement helps to identify the suspected tumorous

area.

[Link] Techniques involved in Image Preprocessing

A number of approaches are available for image preprocessing. Some of them are

described below [70]:

 Adaptive Histogram Equalization (AHE): In this method, total numbers

of histograms are calculated with respect to a particular part of the picture.

The filminess values of the image are also reallocated with the help of this method.

The information (belongs to the modification of the noisy pixels) and the contrasts

(among background pixel) are very important factors in AHE technique. The noisy

pixels become visible like background information. Hence better accuracy is achieved

with this operation. Additionally, an alternative of AHE technique recognized as

Contrast Limited Adaptive Histogram Equalization (CLAHE) technique can be

utilized for improving the contrast of the medical pictures (particularly blood vessels)

as described in Figure 1.2. It also helps in the normalization of the grey levels by

stretching the edges of the image.

4
 Fig.1.2: Adaptive Histogram Equalization

Advantages:

1. This technique is appropriate for the enhancement of edges in every part

of a picture [4].

2. It also removes associated dark black edge pixels and labels.

Disadvantages:

1. Adaptive Histogram Equalization (AHE) comprises an affinity of over-

amplification of noise in comparatively homogeneous areas of a picture.

CLAHE algorithm avoids through the limitation of amplification.

2. This approach is not able to hold the intensity level of input picture.

 Median Filter: Medial filter is a non-linear digital filtering method. This

filter is utilized for the elimination of pepper and salt noises from the altered gray

5
 scale pictures. This filter is utilized for the replacement of center pixel value with

the median of brightness level in vicinity of that pixel. This filter gives better

performance in comparison with mean filter. Median filers are extremely beneficial in

the presence of impulse noise [5, 71]. Impulse noise is too considered as pepper or

salt noise due to the accessibility of the white and black spots adjoining the image.

Fig.1.3: Median Filter

Advantages:

1. This filtering approach can efficiently preserve image edges’ sharpness.

2. This image is able to perform non-linear smoothing of images.

Disadvantages:

1. It is not capable of understanding the particulars from the noises, which

means it is not able to differentiate between fine details and noise.

 Adaptive Median filters: These filters are improvement over existing Median filter.

This is one of the linear filters, in which optimization algorithm is utilized for the varying

picture parameter. The key objective of using this filter is the smoothness of all noises

6
and the elimination of impulse noise. This filter also minimizes the unnecessary thickness

and thinness of the image edges. The impulse response of this filter is more than two.

Fig.1.4: Adaptive Median Filter

Advantages:

1. Impulse noises of higher possibilities can be controlled using these filters.

2. These filters can efficiently maintain the niceties, and perform smoothening

of non-impulse noises [6].

3. This filter decreases distortion such as extreme thinness or thickness of entity

edges.

Disadvantages:

1. These filters differentiate minute details from the noise in partial manner.

 Weiner filter: The Weiner filter also categorized under linear filters, utilized

to remove or reduce the additive signal and blur as of the degraded pictures. An

7
 optimal tradeoff is executed between noise smoothening and inverse filtering in

this filter. This implies that when a specified low-pass filter blurs the picture

then inverse filtering is used for its recovery. Inverse filtering is one of the

restoration methods.

Fig.1.5: De-blurring image with Weiner Filter

Advantages:

1. This filter de-blurs the image and also eliminates the additive noise from

the picture concurrently.

2. This filter is utilized to determine the contrast amid real picture and low-

pass filter picture.

3. The rate of MSE error can be minimized using this approach [7].

Disadvantages:

1. This filtering approach believes in the linearity of method dynamics.

2. This approach can just control additive and unimodal noise.

 Gaussian filter: This filter is a non-uniform, high cut-off filter. This linear filter

is utilized in favor of the blurring edges and decreases contrast or noise from

the pictures. This is applied to an image to smoothen it by calculating the

8
 weighted averages using the overlaying kernel. This eliminates the high frequency

components from the image.

Fig.1.6: Gaussian Filter

Advantages:

1. It has very fast processing speed.

2. This approach smooths images in highly efficient manner.

Disadvantages:

1. These filters can’t maintain image intensity level.

2. This approach is not efficient in the removal of salt and pepper

noise specifically.

1.2.3 SEGMENTATION
A picture is divided into various segments during segmentation process. A picture

cleavage applies to the procedure of dividing a digital picture into many parts.

This process is generally used to locate image objects and edges for making image

analysis simpler [8]. In this process, a label is assigned to every image pixel. Also,

pixels with similar label have some common visual features. The selection of an image

segmentation algorithm relies on the features to be conserved and retrieved. Several

image segmentation methods are described below:

9
 Fig.1.7: Image Segmentation techniques

[Link] Threshold-based Segmentation:

 Global thresholding: One of the easiest way to partition of picture is the

thresholding techniques. These techniques partition the image elements on the

basis of their lightness intensity. Threshold segmentation is a very popular

segmentation technique in region-based segmentation algorithms. The most

important purpose of this approach is to robotically determine the optimal

threshold based on a certain criterion, and use these pixels acco rding to Gray level

for cluster analysis.

 Local thresholding: In this approach, with the division of an image into sub-

images, threshold values are selected locally. After this, calculation of

threshold value for every part is performed. This approach is more time

consuming in comparison with global thresholding. This approach gives good

outcomes along with background orientations. This approach can only be used for

the extraction of minor areas.

10
  Adaptive Thresholding: It is found that in several cases, instability of gray level

and also the orientation of the contrast inside an image always remain because

the same threshold cannot perform well in all the areas. In order to overcome

this problem, a threshold grey level must be used which is a slowly rotating

function of image location. Several threshold values for a number of local

regions are used in adaptive

thresholding.

[Link] Edge-based Segmentation: Gray Histogram Technique:

The edge-based partition techniques are based with respect to the speedy

variation of intensity in a picture in view of fact that incentive in a solitary

intensity can’t give a high-quality data regarding the boundaries [9]. Rim

discovery methods find the margin or brim wherever each the principal

subordinate of intensity is more noteworthy than a specific limit or the

subsequent subsidiary has zero intersections. In edge-based division strategies,

as a matter of first importance the edges are recognized and afterward are

associated together to frame the item limits to portion the necessary areas.

 Gradient Based Method: In this approach, the dissimilarity among the

brightness values of associate pixels is considered. In the presence of an unusual

transformation in the area of a picture, very small picture noise is observed

in co mparison with gradient relied approaches. In this approach, gradient

operators are implemented on the picture.

 Watershed Segmentation: This approach is applied when the brightness of

front and backdrop is discernible and the allocation of picture contrast is even.

11
 This algorithm is also implemented for finding the comparatively weaker

boundaries.

[Link] Clustering-based Segmentation: The clustering based approaches are the

methods that section the picture into groups having pixels with comparable attributes.

Information bunching is the technique that separates the information components into

groups with the end goal that components in same segment are more like each other

than [Link] popular techniques for clustering are:

 K-means: In this technique, the collection of information is divided into k

number of cluster information. This method categorizes the present data

into k number of displace cluster. Two phases are involved in this technique.

The calculation of k centric is performed in first phase. This technique, in

second phase, positions every point to the closest middle cluster with respect to

data point location.

 Fuzzy C-means (FCM): This technique is utilized for the investigation

and development of models. This is an unsupervised approach. In many

cases, this

technique is more genuine in comparison with harsh clustering.

 Hierarchical clustering: In this approach informative objects are alliance in a

picture in the form of a tree. In this technique, identification about the number

of clusters is not necessary in the early phase. Mainly two types of algorithm,

divisive and agglomerative are used in hierarchical clustering approach [10].

[Link] Region-based Segmentation:

The region-based petitioning techniques are the techniques which section the picture into

different segments possess parallel attributes.

12
  Region growing: This is the most commonly used segmentation technique.

This approach starts from a seed pixel. On the basis of some threshold

value, if no boundaries are recognized, this method increases the area by

implementing the associate pixels. For every edge pixel present in the area, this

procedure is necessary.

 Region splitting and merging: This technique is considered as an up down

method. According to a certain methodology, the picture is divided into various

areas. These regions are combined after the division. The entire picture is seen

as a solitary area. After this, with the help of standard deviation interior likeness

is computed.

1.2.4 FEATURE EXTRACTION

After tumor segmentation, characteristics are taken out from the segmented picture in

favor of determining either the cancer is low grade or high grade. There are various

cancer features such as shape, size etc. that capable of discriminating among diverse

cancerous categories. Tumor with high set shows extra irregularity. Some popular

methods of feature extraction are:

[Link] Feature extraction using DWT

Feature extraction from brain MR images using wavelets is considered as finest

approach that highlights pixels in images for better results. Discrete wavelet transform

(DWT) implements ripple alter by means of a disconnected group of ripple levels along

with transformation that follow a number of described regulation. In another

way, this alteration does the decomposition of the signal into equally rectangular

group of ripples [11].

[Link] Feature extraction using GLCM

13
GLCM (Gray level co-occurrence matrix) is a prominent texture-based characteristic

taking out algorithm. This algorithm performs an operation on the images o n the basis

of second- order statistics to locate textural relationship between pixels. On the whole,

this operation is performed using two pixels. GLCM algorithm finds out the frequency

of combinations of these pixels’ intensity levels [65].

1.2.5 IMAGE CLASSIFICATION

Categorization of brain MR pictures for automated recognition of cancer is quite a

tricky task. Classification must provide information whether the image contain tumors

area or not. Image classification can be performed using a number of available

classification algorithms. Some popular classification algorithms for brain tumor

detection include Support Vector Machine, K-nearest neighbor, random-forest, Naïve-

Bayes etc. All these algorithms have been explained below:

[Link] Artificial Neural Network (ANN)

Artificial Neural Network is a collection of interconnected artificial neurons which

behaves like a human brain. There is a sequence of layers in ANN and a set of neurons

is present in every layer. The weighted connections present on all the neurons present

in the previous and next layers are linked with the neurons of every other layer.

Just like the working of biological brain, each connection transmits a signal from

one neuron to other. Depending upon the structure of network and number of

inputs, the performance parameters are calculated [67].

Advantages:

1. It is a relatively simple learning algorithm (e.g. Gradient Descent and

Back- propagation; compared to some of the Bayesian models)

2. Scales well to larger datasets [12].

3. It out-performs other models in case of superiorly labelled data.

14
 Disadvantages:

1. These models are more computationally expensive than other

traditional algorithms.

2. These networks show good performance with massive volume of data.

3. Neural networks correspond to black box. Therefore, these models

can’t be interpreted easily.

Fig.1.8: Artificial Neural Network

[Link] Convolutional-Neural-Network (CNN)

CNN or Conv-Net is a popular deep learning and feed forwarding algorithm, in

which multiple layers are connected to analyze the images effectively. The main

building blocks of CNN include four layers of Convolutional, Pooling, ReLu and Fully-

connected. In medical image analysis, they play a significant role to classify the tumors

or any affected part of the human body [13,68].

Advantages:

1. It minimizes computation as compared to a regular NN.

15
 2. Convolution abridges calculation to a large level without ruining the

essence of information.

3. It can handle image classification effectively. These models are

able to classify images in a competent manner.

4. These networks apply similar comprehension across all picture positions.

5. After training, these models can generate predictions quickly.

6. These models can perform efficiently with a number of inputs and

layers.

Disadvantages:

1. CNN is computationally expensive. It is possible to perform quicker

implementations using improved computing hardware tools i.e.

GPUs and Neuromorphic chips.

2. The deepness of network slows down its performance in some manner.

3. CNN require huge quantity of stets of information to achieve

satisfactory. These networks show high-quality performance with large

volume of data.

Fig.1.9: Convolutional Neural Network

16
[Link] Decision Tree (DT)

A dataset is divided into uniform subsets repeatedly for calculating the class

membership through DT classifier. In every intermediary state, the acceptations as well

as refusal of set tags are achieved through the level classifier [69]. The node

partitioning, identification of closing knobs along with allocating the set tag to closing

junction are the three major parts of this classifier.

Fig.1.10: Decision Tree

Advantages:

1. These models can be interpreted easily.

2. These models are suitable for constant and unconditional inputs.

3. These models do not require much computation for Classification.

4. These models are able in rule generation for providing support to

specialists so that they can validate their data [14].

Disadvantages:

1. High classification fault rate occurs when training suite is smaller

than the number of classes.

17
 2. Once the issue goes bigger, exponential computation increase occurs.

[Link] Support Vector Machine (SVM)

With the help of Support Vector Machine (SVM), a hyper plane or suit of hyper planes

is generated in high dimensional space for performing classification [66]. The hyper

plane located remotely from the adjacent training data end of some class supports in

the attainment of good division. Fundamentally, the speculation mistake of classifier is

fewer on the off chance that when the edge is huge.

Fig.1.11: Support Vector Machine

It is conceivable to deal with more info information in an effective way by utilizing

non- parametric through double classifier strategy in SVM. Contingent on the

picked hyperplane and bit boundary, superior and exactness are accomplished

Advantages:

1. In contrast to NB, this algorithm gives high accuracy and generates

predictions in lesser time [15].

2. This approach can be utilized for classification as well as regression

issues.

18
 3. This algorithm performs better by means of large dimensional space as

well as obvious division boundary.

Disadvantages:

1. This algorithm can’t be applied on big sets of information.

2. It acquires additional time in preparing. This algorithm consumer

additional time during preparing.

3. This algorithm performs inadequately through intertwining sorts.

19