ABSTRACT

The improvement of technology and Machine Learning can help radiologists in tumor
diagnostics without invasive measures. A Machine-learning algorithm that has
achieved substantial results in image segmentation and classification is the
convolutional neural network (CNN). A new CNN architecture is proposed for brain
tumor classification of three stages of tumor types. Initially, the image is filtered using
2D Adaptive Median Filter (2D-AMF) for noise removal. The de-noised image is then
enhanced using Contrast Limited Adaptive Histogram Equalization (CLAHE)
algorithm. From the enhanced image, the Region Of Interest (ROI) is segmented using
clustering and threshold methods. Clustering is performed using Fast Fuzzy C Means
Clustering (FFCM) and threshold is performed using Otsu threshold (OT) algorithm.
From the ROI, features are extracted using Gray Level Co-Occurrence Matrix
(GLCM). GLCM is a feature that computes the occurrence of pixel pairs in specific
spatial coordinates of an image. Finally, the obtained features are classified using
machine learning based Convolutional Neural Network (CNN) techniques.

v
 CONTENTS
CHAPTER No. TITLE PAGE No.

ABSTRACT V
LIST OF TABLES VII

1. INTRODUCTION
1.1 OBJECTIVES 01
1.2 EXISTING SYSTEM 02
1.2.1 DISADVANTAGES OF EXISTING
ALGORITHMS 02
1.3 PROPOSED SYSTEM 03
1.3.1 IMAGE RESTORATION 03
1.3.2 IMAGE ENHANCEMENT 03
a) Adaptive histogram adjustment 03
b) Contrast Enhancement 04
1.4 DEGRADATION MODEL 05
1.5 IMPULSE DETECTION 07
2. LITERATURE SURVEY 09
3. SYSTEM MODEL 13
3.1 SEGMENTATION 14
3.1.1 IMAGE SEGMENTATION 14
a)Image Segmentation by Clustering 15

3.2 CLUSTERING ALGORITHMS 17

3.3 K MEANS SEGMENTATION 20

3.3.1 METHODOLOGY 21

vi
 3.3.2 K-Means Algorithm 24

3.4 ADVANTAGES OF PROPOSED SYSTEMS 32

3.5 FCM Algorithm 35

3.6 The Fuzzy c-means Algorithm 38

3.7 The FCM Algorithm 38

3.8 Results 43

3.9 ADAPTIVE MEAN SHIFT THRESHOLDING FOR

ALZHEIMER SEGMENTATION 47

4. RESULTS AND DISCUSSION 48

5. CONCLUSSION 55
FUTURE WORK 55
REFERENCES 56

vii
 LIST OF FIGURES

FIG NO: TITLE PAGE NO.

1.1 block diagram of general degradation model 04

3.1 Similarity between image regions or pixels 16

3.2 Flow-chart of an image segmentation method 18

3.3 K-means clustering method 19

3.4 M-FCM Segmentation on synthetic images corrupted

with multiplicative gain and real T1 brain MRI images 30

3.5 FCM with low SNR images 45

4.1 Input Image 48

4.2 Filter Image using Anisotropic Diffusion Filter 49

4.3 Clustering 50
4.4 Modified Fuzzy C Means Segmentation 51
4.5 Tumor Image 52
4.6 Super-Imposed image of Tumor Region 53

viii
ix
 CHAPTER 1

INTRODUCTION

We propose another technique for deformable enlistment of pre-usable and post-repeat mind
MR sweeps of patients. Playing out this sort of intra-subject enlistment is trying as
Alzheimer, resection, repeat, and edema cause enormous misshapenings, missing
correspondences, and conflicting force profiles between the outputs. Populace investigation
of cerebrum morphology from attractive reverberation pictures adds to the examination and
comprehension of neurological infections. Such investigation ordinarily includes division of
an enormous arrangement of pictures and correlations of these divisions between significant
subgroups of pictures (e.g., "typical" versus "sick").

The pictures of every subgroup are normally chosen ahead of time in a regulated
manner dependent on clinical information. Their divisions are commonly guided by at least
one accessible map books, thought to be appropriate for the current pictures. Bunching
approach is broadly utilized in biomedical applications especially for cerebrum tumor
identification in unusual attractive reverberation (MRI) pictures.

Fuzzy clustering using modified fuzzy C-means (FCM) calculation end up being
better preposterous bunching approaches as far as division proficiency. Yet, the significant
disadvantage of the changed FCM calculation is the enormous computational time needed for
intermingling. The viability of the adjusted FCM calculation as far as computational rate is
improved by changing the group community and participation esteem refreshing model.

In this paper, the utilization of adjusted FCM calculation for MR cerebrum tumor
identification is investigated. An extensive element vector space is utilized for the division
method. Near investigation as far as division effectiveness and combination rate is performed
between the traditional FCM and the changed FCM. Exploratory outcomes show unrivaled
outcomes for the changed FCM calculation as far as the exhibition measures.

1.1 OBJECTIVES
• To develop an Efficient Alzheimer’s disease segmentation on MRI brain image
and classification of stages of tumor using machine learning
Detailed Objectives:
• Initially, the image is filtered using 2D Adaptive Median Filter (2D-AMF) for
noise removal.
• The de-noised image is then enhanced using Contrast Limited Adaptive
Histogram Equalization (CLAHE) algorithm.

1
 • From the enhanced image, the Region Of Interest (ROI) is segmented using
clustering and threshold methods.
• Clustering is performed using Fast Fuzzy C Means Clustering (FFCM) and
threshold is performed using Otsu threshold (OT) algorithm.
• From the ROI, features are extracted using Gray Level Co-Occurrence Matrix
(GLCM).
• GLCM is a feature that computes the occurrence of pixel pairs in specific
spatial coordinates of an image.
• Finally, the obtained features are classified using machine learning based
Convolutional Neural Network (CNN) techniques.

1.2 EXISTING SYSTEM

Force based Region-developing division

1.2.1 DISADVANTAGES OF EXISTING ALGORITHM

• The locale developing division isn't liked for its restricted scope of utilizations and
programmed highlights are not having precise qualities.

• Preprocessing tests are expected to discover which kind of sifting will be more
advantageous. This builds the impact of the dot clamor and Gaussian commotion the
ultrasound, mammogram Images and MRI Images.

• The wanted Alzheimer territory is chosen from the portioned picture to ascertain the
volume. The volume of the ideal Alzheimer zone is more noteworthy than the first Alzheimer
region. The locale developing calculation will fragment the Alzheimer zone as well as the
non-Alzheimer territory which has extreme focus proportion.

• This calculation completely relies upon the force of the picture not the shape and
surface. So the exactness and affectability is low.

2
1.3 PROPOSED SYSTEM

Preprocessing

Algorithm 1

1.3.1 Image Restoration

Proposed Algorithm 1

(Image Pre-processing Step 1 – Image Denoising (Restoration))

Image De-noising utilizing Anisotropic Diffusion combination Filter

The least difficult and best explored dissemination strategy for smoothing pictures is
to apply a straight dispersion measure. We will zero in on the connection between
straight dissemination separating and the convolution with a Gaussian, investigate its
smoothing properties for the picture just as its subordinates, and survey the major
properties of the Gaussian scale-space incited by direct dispersion sifting.

Proposed Algorithm 2

1.3.2 Image Enhancement

Image Enhancement utilizing Contrast Limited Adaptive Histogram

Equalization (CLAHE calculation)

a)Adaptive histogram adjustment

(AHE) is a PC picture handling strategy used to improve contrast in pictures. It varies

from customary histogram leveling in the regard that the versatile technique processes
a few histograms, each comparing to a particular segment of the picture, and uses
them to reallocate the gentility estimations of the picture. It is subsequently
appropriate for improving the neighborhood difference of a picture and bringing out
more detail. In any case, AHE tends to over-enhance commotion in moderately
homogeneous districts of a picture. A variation of versatile histogram balance called
contrast restricted versatile histogram balance (CLAHE) forestalls this by restricting
the intensification.

3
Head part examination (PCA) is a factual technique that utilizes a symmetrical change
to change over a bunch of perceptions of potentially associated factors into a bunch of
estimations of directly uncorrelated factors called head segments. PCA is the least
complex of the genuine eigenvector-based multivariate examinations. Regularly, its
activity can be considered as uncovering the interior construction of the information
such that best clarifies the change in the information. PCA is numerically
characterized as a symmetrical direct change that changes the information to another
organize framework to such an extent that the best fluctuation by some projection of
the information comes to lie on the main arrange (called the primary head segment),
the second most prominent difference on the subsequent facilitate, etc.

b)Contrast Enhancement

There is a solid impact of differentiation proportion on settling force and recognition

capacity of pictures. Strategies for improving picture contrast are among the most
broadly utilized upgrade measures. The affectability scope of any distant detecting
locator is intended to record a wide scope of territory brilliance from dark basalt levels
to White Sea beds under a wide scope of lighting conditions. Barely any individual
scenes have a brilliance range that uses the fu l s en s I tivity scope of these identifiers.
To deliver a picture with the ideal differentiation proportion, it is imperative to u t
ilize the whole brilliance scope of the presentation medium, which is for the most part
film.

The motivation behind picture reclamation is to "make up for" or "fix" deserts which
debase a picture. Corruption comes in numerous structures, for example, movement
obscure, commotion, and camera misfocus. In cases like movement obscure, it is
feasible to concoct an awesome gauge of the real obscuring capacity and "fix" the
haze to reestablish the first picture. In situations where the picture is defiled by
commotion, the best we may want to do is to make up for the corruption it caused. In
this task, we will present and execute a few of the strategies utilized in the picture
preparing world to reestablish pictures.

1.4 Degradation Model

The block diagram for our general degradation model is

4
Fig. 1.1: block diagram of general degradation model
where g is the corrputed picture gotten by passing the first picture f through a low pass
channel (obscuring fuction) b and adding clamor to it.

To comprehend what is the issue here, one first requirements to comprehend what a
middle channel is and what it does. In a wide range of sorts of advanced picture
handling, the fundamental activity is as per the following: at every pixel in a
computerized picture we place a neighborhood around that point, break down the
estimations of the relative multitude of pixels in the area as indicated by some
calculation, and afterward supplant the first pixel's worth with one dependent on the
investigation performed on the pixels in the area. The local at that point moves
progressively over each pixel in the picture, rehashing the interaction.

The BDND calculation initially orders the pixels of a restricted window, focusing on
the current pixel, into three gatherings lower force motivation commotion,
uncorrupted pixels, and higher power drive clamor. The middle pixel will at that point
be considered as uncorrupted, given that it has a place with the uncorrupted‖ pixel
bunch, or undermined. For that, two limits that segregate these three gatherings need
to be precisely decided for yielding high commotion identification exactness.

The BDND calculation is applied to every pixel of the boisterous picture to

distinguish whether it is uncorrupted or tainted. After such an application to the whole
picture, a two-dimensional paired choice guide is shaped toward the finish of the
clamor recognition stage, with 0s showing the places of uncorrupted pixels, and 1s for
those tainted ones.

To achieve this evenhanded, every one of the pixels inside a pre-characterized

window that middle on the considered pixel will be assembled into three bunches;
consequently, two limits b1and b2 are needed to be resolved. For every pixel being
thought of, if 0≤xi,j ≤b1, the pixel will be relegated to the lower-power group; in any
case, to the medium-force bunch for b1<xi,j≤b2 or to the focused energy group for
b2<xi,j≤255. Clearly, if the middle pixel being viewed as falls onto the center group,
it will be treated as uncorrupted, since its power esteem is neither generally low nor
moderately high. Else, all things considered, the pixel has been defiled by motivation
clamor. Unmistakably, the precision of bunching results (subsequently, the exactness
of commotion identification) at last relies upon how exact the recognized limits b1and
b2 are.

The limit discriminative cycle comprises of two emphasis, in which the subsequent
emphasis might be summoned restrictively. In the principal emphasis, a broadened
neighborhood window with a size of 21x21 (exactly decided) is utilized to analyze
whether the considered pixel is an uncorrupted one.

5
Clamor is any undesired data that sullies a picture. Motivation commotion is an
extraordinary sort of clamor, which have a wide range of roots. The Salt and Pepper
type drive clamor is ordinarily brought about by breaking down of the pixel
components in the camera sensors, broken memory areas, or timing mistakes in the
digitization interaction. For the pictures debased by Salt and Pepper clamor, the
boisterous pixels can take just the most extreme or the base qualities in the powerful
reach.

An important non linear filter that will preserve the edges and remove impulse
noise is standard median filter Median (SMF). It replaces every pixel by its median
value from its neighborhood and often removes desirable details in the image.
Specialized median filter such as weighted median filter (WMF) were proposed to
improve the performance of median filter by giving more weight to some selected
pixel in the filtering window. But they still implemented uniformly across the images
without considering the current pixel whether is noisy or noise free. Therefore, a
noise-detection process to discriminate between uncorrupted pixels and corrupted
pixels is highly desirable. Some of decision based algorithms, such as Adaptive
Median filter (AMF), These algorithms first detect the noisy pixels and remove it by
applying either standard median filter or its variants. These filters are effective in
removing low to medium density impulse noise only.

1. Less effective in removing Gaussian or random-intensity noise. The median

filter can remove noise only if the noisy pixels occupy less than one half of the
neighborhood area.

2. Repeating will remove noise but at the expense of detail (posterizationoccurs )

where pixel brightness values are leveled across regions "group of pixels
having similar brightness values
3. High computational cost (for sorting N pixels, the temporal complexity is
O(N·logN), When the median filter must be carried out in real time, the
software implementation in general-purpose processors does not usually give
good results and FPGAsare a good alternative
4. Some median algorithms are not good for real time processing.