BRAIN TUMOUR IDENTIFICATION USING

CONVOLUTIONAL NEURAL NETWORK

A project report submitted in partial fulfilment of the requirements for
the award of the degree of

BACHELOR OF TECHNOLOGY
IN
COMPUTER SCIENCE ENGINEERING
Submitted by

J. G. SIVA SAI (316126510085)

ROHITHA. K (316126510108)

S. DEEPIKA (316126510116)

M. N. SINDHURI (316126510098)

Under the guidance of

P. Naga Srinivasu
Assistant Professor

DEPARTMENT OF COMPUTER SCIENCE AND ENGINEERING

ANIL NEERUKONDA INSTITUTE OF TECHNOLOGY AND SCIENCES

(UGC AUTONOMOUS)
(Permanently Affiliated to AU, Approved by AICTE and Accredited by NBA & NAAC with ‘A’ Grade)

Sangivalasa, Bheemili Mandal, Visakhapatnam Dist. (A.P)

2019 – 2020

i
 DEPARTMENT OF COMPUTER SCIENCE AND ENGINEERING

ANIL NEERUKONDA INSTITUTE OF TECHNOLOGY AND SCIENCES

(UGC AUTONOMOUS)
(Permanently Affiliated to AU, Approved by AICTE and Accredited by NBA & NAAC with ‘A’ Grade)

Sangivalasa, Bheemili Mandal, Visakhapatnam Dist. (A.P)

2019 - 2020

BONAFIDE CERTIFICATE

This is to certify that the project report entitled “BRAIN TUMOUR IDENTIFICATION
USING CONVOLUTIONAL NEURAL NETWORK” submitted by J. G. SIVA SAI
(316126510085), ROHITHA. K (316126510108), S. DEEPIKA (316126510116), M. N.
SINDHURI (316126510098) in partial fulfilment of the requirements for the award of the
degree of Bachelor of Technology in Computer Science Engineering of Anil
Neerukonda Institute of Technology and Sciences (A), Visakhapatnam is a record of
bonafide work carried out under my guidance and supervision.

Project Guide Head of the Department

P. Naga Srinivasu Dr. R. Sivaranjani

Assistant Professor Professor
Department of CSE Department of CSE
ANITS ANITS

ii
 DECLARATION

We, J. G. SIVA SAI (316126510085), ROHITHA. K (316126510108), S.

DEEPIKA (316126510116), M. N. SINDHURI (316126510098) , of Fourth Year
B.Tech., in the Department of Computer Science and Engineering from ANITS,
Visakhapatnam, hereby declare that the project work entitled “BRAIN TUMOUR
IDENTIFICATION USING CONVOLUTIONAL NEURAL NETWORK” is carried
out by us and submitted in partial fulfilment of the requirements for the award of Bachelor
of Technology in Computer Science Engineering , under Anil Neerukonda Institute of
Technology & Sciences(A) during the academic year 2016-2020 and has not been
submitted to any other university for the award of any kind of degree.

J. G. SIVA SAI 316126510085

ROHITHA. K 316126510108

S. DEEPIKA 316126510116

M. N. SINDHURI 316126510098

iii
 ACKNOWLEDGEMENT

We would like to express our deep gratitude to our project guide, Mr. P. Naga
Srinivasu, Assistant Professor, Department of Computer Science and Engineering,
ANITS, for his guidance with unsurpassed knowledge and immense encouragement. We
are grateful to Dr. R. Sivaranjani, Head of the Department, Computer Science and
Engineering, for providing us with the required facilities for the completion of the project
work.

We are very much thankful to the Principal and Management, ANITS,

Sangivalasa, for their encouragement and cooperation to carry out this work.

We also thank our Project Coordinator Mrs. K. S. Deepthi, for her support and
encouragement. We express our thanks to all teaching faculty of Department of CSE,
whose suggestions during reviews helped us in accomplishment of our project. We would
like to thank Mrs. Udaya Lakshmi of the Department of CSE, for providing us the lab
resources in accomplishment of our project.

We would like to thank our parents, friends, and classmates for their encouragement
throughout our project period. At last but not the least, we thank everyone for supporting
us directly or indirectly in completing this project successfully.

J. G. SIVA SAI 316126510085

ROHITHA. K 316126510108

S. DEEPIKA 316126510116

M. N. SINDHURI 316126510098

iv
ABSTRACT:

The human brain is the major controller of the humanoid system. The abnormal
growth and division of cells in the brain lead to a brain tumor, and the further growth of
brain tumors leads to brain cancer. In the area of human health, Computer Vision plays a
significant role, which reduces the human judgment that gives accurate results. CT scans,
X-Ray, and MRI scans are the common imaging methods among magnetic resonance
imaging (MRI) that are the most reliable and secure. MRI detects every minute objects.
Our paper aims to focus on the use of different techniques for the discovery of brain cancer
using brain MRI. In this study, we performed pre-processing using the bilateral filter (BF)
for removal of the noises that are present in an MR image. This was followed by the binary
thresholding and Convolution Neural Network (CNN) segmentation techniques for reliable
detection of the tumor region. Training, testing, and validation datasets are used. Based on
our machine, we will predict whether the subject has a brain tumor or not. The resultant
outcomes will be examined through various performance examined metrics that include
accuracy, sensitivity, and specificity. It is desired that the proposed work would exhibit a
more exceptional performance over its counterparts.

KEYWORDS: Brain tumor, Magnetic resonance imaging, Adaptive Bilateral Filter,

Convolution Neural Network.

v
LIST OF FIGURES:

Figure No. Description of Figures Page No.

Fig 3.1 Location of tumors in eight different images 17
Fig 4.1 Module Division 19
Fig 5.1 Input Image 32
Fig 5.2 Gray-scale Image 33
Fig 5.3 Filtered Image 33
Fig 5.4 Image after edge detection 34
Fig 5.5 Thresholded and Binary Image 35
Fig 5.6 Morphological Image 35
Fig 5.7 Dilated Image 36
Fig 5.8 Training data history 37
Fig 5.9 Represents the performance analysis of CNN 45
Fig 5.10 Represents the performance of proposed CNN 46

LIST OF TABLES:

Table No. Description of Tables Page No.

Table 1 Represents the true positive, true negative, false 45
positive and false negative values of the
proposed approach for different set of images.
Table 2 Represents the Accuracy, Sensitivity, and 46
Specificity of the proposed approach for
different set of images.

vi
 TABLE OF CONTENTS

CHAPTER NO. TITLE PAGE NO.

Abstract v
Keywords v
List of figures vi
List of tables vi

1. INTRODUCTION 1
1.1 Brain anatomy 2
1.2 Motivation for the Work 3
1.3 Problem Statement 4
1.4 Scope 4
1.5 Organization of Thesis 5

2. LITERATURE SURVEY 6
2.1 Extraction of Brain Tumor Using Soft Computing Techniques 6
2.2 Improved Edge Detection for Brain Tumor Detection 6
2.3 Image Segmentation by Clustering Techniques 7
2.4 Image Segmentation by Using Morphological Operations 7
2.5 Brain Tumor Classification & Identification Using KNN Technique 7
2.6 Comparative Analysis of Thresholding & Edge Detection Techniques 7
2.7 Fusing Images with different focuses using SVM 8
2.8 Texture based tumor detection using Seeded Region Growing Method 8
2.9 Brain Tumor Detection Using Artificial Neural Networks 8
2.10 An Accurate Bayesian Method for Automatic Segmentation 9
2.11 MR Image Classification Using Adaboost for Brain Tumor Type 9
2.12 Brain MR Image Segmentation for Tumor Detection using ANN 9
2.13 Brain Image Segmentation using SVM 10
2.14 Segmentation based Fuzzy Partition Entropy & Genetic Algorithm 10

vii
 2.15 Segmentation and Classification of MRI Brain Tumor 11
2.16 Brain tumor grading based on Neural Networks and CNN 11
2.17 Brain Tumor Segmentation Using CNN in MR Images 11
2.18 Quantification of Brain Tumor and it’s Symmetric Analysis 12
2.19 Automatic segmentation using Convolutional Neural Networks 12
2.20 Segmentation of K – Means with advanced Dual Localization 13
2.21 Wavelet-based feature extraction for brain tumor diagnosis 13
2.22 Identification of tumor with feature extraction using DWT and PNN 13
2.23 Brain Tumor Detection Using Histogram Equalization and FSVM 14
2.24 Improved K-means algorithm in the design of RBF neural networks 14
2.25 Survey on Neural Networks Used for Medical Image Processing 15

3. CHALLENGES IN BRAIN TUMOR CLASSIFICATION 17

4. MODULE DIVISION 18
4.1 Image pre-processing and Image Enhancement 19
4.1.1 Image Pre-processing 19
4.1.1.1 Image Acquisition from Dataset 20
4.1.1.2 Convert image from One Colour Space to Another 20
4.1.1.3 Filters 20
4.1.2 Image Enhancement 21
4.1.2.1 Sobel Filter 21
4.2 Image Segmentation 22
4.2.1 Thresholding 23
4.2.2 Morphological Operations 24
4.3 Image Classification 25
4.3.1 Sequential 26
4.3.2 Convolution 26
4.3.3 Pooling 27
4.3.4 Flattening 27
4.3.5 Fully Connection 27

viii
5. EXPERIMENTAL ANALYSIS AND RESULTS 29
5.1 System Configuration 29
5.1.1 Software Requirements 29
5.1.2 Hardware Configuration 31
5.2 Sample Code and Results 32
5.3 Experimental Results 42
5.4 Performance Measures 44
5.5 Performance Evaluation 45

6. CONCLUSION AND FUTURE SCOPE 47

6.1 Conclusion 47
6.2 Future Scope 47

REFERNCES 48

ix
1 INTRODUCTION

Medical imaging is the technique and process of creating visual representations

of the interior of a body for clinical analysis and medical intervention, as well as visual
representation of the function of some organs or tissues. Medical imaging seeks to
reveal internal structures hidden by the skin and bones, as well as to diagnose and
treat disease. Medical imaging also establishes a database of
normal anatomy and physiology to make it possible to identify abnormalities.

The medical imaging processing refers to handling images by using the

computer. This processing includes many types of techniques and operations such as
image gaining, storage, presentation, and communication. This process pursues the
disorder identification and management. This process creates a data bank of the regular
structure and function of the organs to make it easy to recognize the anomalies. This
process includes both organic and radiological imaging which used electromagnetic
energies (X-rays and gamma), sonography, magnetic, scopes, and thermal and isotope
imaging. There are many other technologies used to record information about the
location and function of the body. Those techniques have many limitations compared
to those modulates which produce images.

An image processing technique is the usage of a computer to manipulate the

digital image. This technique has many benefits such as elasticity, adaptability, data
storing, and communication. With the growth of different image resizing techniques,
the images can be kept efficiently. This technique has many sets of rules to perform in
the images synchronously. The 2D and 3D images can be processed in multiple
dimensions.

1
1.1 BRAIN ANATOMY:

The brain tumor is one all the foremost common and, therefore, the deadliest
brain diseases that have affected and ruined several lives in the world. Cancer is a
disease in the brain in which cancer cells ascends in brain tissues. Conferring to a new
study on cancer, more than one lakh people are diagnosed with brain tumors every year
around the globe. Regardless of stable efforts to overcome the complications of brain
tumors, figures show unpleasing results for tumor patients. To contest this, scholars are
working on computer vision for a better understanding of the early stages of tumors and
how to overcome using advanced treatment options.

Magnetic resonance (MR) imaging and computed tomography (CT) scans of the
brain are the two most general tests to check the existence of a tumor and recognize its
position for progressive treatment decisions. These two scans are still used extensively
for their handiness, and the capability to yield high-definition images of pathological
tissues is more. At present, there are several other conducts offered for tumors, which
include surgery, therapies such as radiation therapy, and chemotherapy. The decision
for which treatment relies on the many factors such as size, kind, and grade of the tumor
present in the MR image. It’s conjointly chargeable for whether or not cancer has
reached the other portions of the body.

Precise sighting of the kind of brain abnormality is enormously needed for

treatment operations with a resolution to diminish diagnostic errors. The precision is
often makeshift utilizing computer-aided diagnosis (CAD) systems. The essential plan
of computer vision is to produce a reliable output, which is an associate estimation to
assist medical doctors in image understanding and to lessen image reading time. These
advancements increase the steadiness and correctness of medical diagnosis — however,
segmenting an MR image of the tumor and its area itself a very problematic job. The
occurrence of tumors in specific positions within the brain image without distinguishing
picture intensities is an additional issue that makes a computerized detection of brain
tumor and segmentation a problematic job.

2
1.2 MOTIVATION FOR THE WORK:

A brain tumor is defined as abnormal growth of cells within the brain or central
spinal canal. Some tumors can be cancerous thus they need to be detected and cured in
time. The exact cause of brain tumors is not clear and neither is exact set of symptoms
defined, thus, people may be suffering from it without realizing the danger. Primary
brain tumors can be either malignant (contain cancer cells) or benign (do not contain
cancer cells).

Brain tumor occurred when the cells were dividing and growing abnormally. It
is appearing to be a solid mass when it diagnosed with diagnostic medical imaging
techniques. There are two types of brain tumor which is primary brain tumor and
metastatic brain tumor. Primary brain tumor is the condition when the tumor is formed
in the brain and tended to stay there while the metastatic brain tumor is the tumor that
is formed elsewhere in the body and spread through the brain.

The symptom having of brain tumor depends on the location, size and type of
the tumor. It occurs when the tumor compressing the surrounding cells and gives out
pressure. Besides, it is also occurring when the tumor blocks the fluid that flows
throughout the brain. The common symptoms are having headache, nausea and
vomiting, and having problem in balancing and walking. Brain tumor can be detected
by the diagnostic imaging modalities such as CT scan and MRI. Both of the modalities
have advantages in detecting depending on the location type and the purpose of
examination needed. In this paper, we prefer to use the MRI images because it is easy
to examine and gives out accurate calcification and foreign mass location.

The MRI is the most regularly utilized strategy for imaging brain tumors and the
identification of its vicinity. The conventional technique for CT and MR image
classification and detection of tumor cells remains largely supported for the human
reviewing apart from different other methods. MR images are mainly used because there
are non-destructive and non-ionizing. MR imaging offers high-definition pictures that
are extensively utilized in discovering brain tumors. MRI has diverse schemes such as
flair, T1-weighted, T2-weighted images. There are many image processing techniques
such as pre-processing, segmentation of images, image improvements, feature
extraction, and classifiers.

3
1.3 PROBLEM STATEMENT:

Our study deals with automated brain tumor detection and classification.
Normally the anatomy of the brain is analyzed by MRI scans or CT scans. The aim of
the paper is tumor identification in brain MR images. The main reason for detection of
brain tumors is to provide aid to clinical diagnosis. The aim is to provide an algorithm
that guarantees the presence of a tumor by combining several procedures to provide a
foolproof method of tumor detection in MR brain images. The methods utilized are
filtering, erosion, dilation, threshold, and outlining of the tumor such as edge detection.

The focus of this project is MR brain images tumor extraction and its
representation in simpler form such that it is understandable by everyone. The objective
of this work is to bring some useful information in simpler form in front of the users,
especially for the medical staff treating the patient. The aim of this work is to define an
algorithm that will result in extracted image of the tumor from the MR brain image. The
resultant image will be able to provide information like size, dimension and position of
the tumor, and its boundary provides us with information related to the tumor that can
prove useful for various cases, which will provide a better base for the staff to decide
the curing procedure. Finally, we detect whether the given MR brain image has tumor
or not using Convolution Neural Network.

1.4 SCOPE:

Our aim is to develop an automated system for enhancement, segmentation and

classification of brain tumors. The system can be used by neurosurgeons and healthcare
specialists. The system incorporates image processing, pattern analysis, and computer
vision techniques and is expected to improve the sensitivity, specificity, and efficiency
of brain tumor screening. The primary goal of medical imaging projects is to extract
meaningful and accurate information from these images with the least error possible.
The proper combination and parameterization of the phases enables the development of
adjunct tools that can help on the early diagnosis or the monitoring of the tumor
identification and locations.

4
1.5 ORGANIZATION OF THESIS:
In this document, chapter 2 consists about literature survey. The literature survey
tells about the research done to work on the project. All the details about the papers,
websites on which the research work is done in order to work on the project is provided
in the literature survey. In chapter 4, we discuss about the various methodologies used
in the project. In chapter 5, the details about experimental analysis is discussed. The
experimental analysis includes sample code, result screenshots for a tested input image.
In the next chapter we give the conclusion about the project and also provide
information if the project can be implemented further or not. In the final chapter we
provide all the references for this project.

5
2. LITERATURE SURVEY

In Medical diagnosis, robustness and accuracy of the prediction algorithms are

very important, because the result is crucial for treatment of patients. There are many
popular classification and clustering algorithms used for prediction. The goal of
clustering a medical image is to simplify the representation of an image into a
meaningful image and make it easier to analyze. Several Clustering and Classification
algorithms are aimed at enhancing the prediction accuracy of diagnosis process in
detecting abnormalities.

In the literature survey we provide a brief summary of the different methods that
have been proposed for clustering over the period of 2002 to 2018.We have been though
25 papers each of which has a unique approach towards segmentation in some parameter
or the other. The summaries of each of the papers are provided below.

 A. Sivaramakrishnan And Dr. M. Karnan “A Novel Based Approach for

Extraction Of Brain Tumor In MRI Images Using Soft Computing
Techniques,” International Journal Of Advanced Research In Computer
And Communication Engineering, Vol. 2, Issue 4, April 2013.

A. Sivaramakrishnan et al. (2013) [1] projected an efficient and innovative discovery of

the brain tumor vicinity from an image that turned into finished using the Fuzzy C-
approach grouping algorithm and histogram equalization. The disintegration of images
is achieved by the usage of principal factor evaluation is done to reduce the extent of
the wavelet coefficient. The outcomes of the anticipated FCM clustering algorithm
accurately withdrawn tumor area from the MR images.

 Asra Aslam, Ekram Khan, M.M. Sufyan Beg, Improved Edge Detection
Algorithm for Brain Tumor Segmentation, Procedia Computer Science,
Volume 58,2015, Pp 430-437, ISSN 1877-0509.

M. M. Sufyan et al. [2] has presented a detection using enhanced edge technique for
brain-tumor segmentation that mainly relied on Sobel feature detection. Their presented
work associates the binary thresholding operation with the Sobel approach and
excavates diverse extents using a secure contour process. After the completion of that
process, cancer cells are extracted from the obtained picture using intensity values.

6
  B.Sathya and R.Manavalan, Image Segmentation by Clustering Methods:
Performance Analysis, International Journal of Computer Applications
(0975 – 8887) Volume 29– No.11, September 2011.

Sathya et al. (2011) [3], provided a different clustering algorithm such as K-means,
Improvised K-means, C-means, and improvised C-means algorithms. Their paper
presented an experimental analysis for massive dat=asets consisting of unique
photographs. They analyzed the discovered consequences using numerous parametric
tests.

 Devkota, B. & Alsadoon, Abeer & Prasad, P.W.C. & Singh, A.K. &
Elchouemi, A. (2018). Image Segmentation for Early Stage Brain Tumor
Detection using Mathematical Morphological Reconstruction. Procedia
Computer Science. 125. 115-123. 10.1016/j.procs.2017.12.017.

B. Devkota et al. [4] have proposed that a computer-aided detection (CAD) approach is
used to spot abnormal tissues via Morphological operations. Amongst all different
segmentation approaches existing, the morphological opening and closing operations
are preferred since it takes less processing time with the utmost efficiency in
withdrawing tumor areas with the least faults.

 K. Sudharani, T. C. Sarma and K. Satya Rasad, "Intelligent Brain Tumor

lesion classification and identification from MRI images using a K-NN
technique," 2015 International Conference on Control, Instrumentation,
Communication and Computational Technologies (ICCICCT),
Kumaracoil, 2015, pp. 777-780. DOI: 10.1109/ICCICCT.2015.7475384

K. Sudharani et al. [5] presented a K- nearest neighbor algorithm to the MR images to

identify and confine the hysterically full-fledged part within the abnormal tissues. The
proposed work is a sluggish methodology but produces exquisite effects. The accuracy
relies upon the sample training phase.

 Kaur, Jaskirat & Agrawal, Sunil & Renu, Vig. (2012). A Comparative
Analysis of Thresholding and Edge Detection Segmentation Techniques.
International Journal of Computer Applications.vol. 39.pp. 29-34.
10.5120/4898-7432.

7
Jaskirat Kaur et al. (2012) [6] defined a few clustering procedures for the segmentation
process and executed an assessment on distinctive styles for those techniques. Kaur
represented a scheme to measure selected clustering techniques based on their
steadiness in exceptional tenders. They also defined the diverse performance metric
tests, such as sensitivity, specificity, and accuracy.

 Li, Shutao, JT-Y. Kwok, IW-H. Tsang and Yaonan Wang. "Fusing images
with different focuses using support vector machines." IEEE Transactions
on neural networks 15, no. 6 (2004): 1555-1561.

J.T. Kwok et al. [7] delivered wavelet-based photograph fusion to easily cognizance at
the object with all focal lengths as several vision-related processing tasks can be carried
out more effortlessly when wholly substances within the images are bright. In their work
Kwok et al. investigated with different datasets, and results show that presented work
is extra correct as it does not get suffering from evenness at different activity stages
computations.

 M. Kumar and K. K. Mehta, "A Texture based Tumor detection and

automatic Segmentation using Seeded Region Growing Method,"
International Journal of Computer Technology and Applications, ISSN:
2229-6093, Vol. 2, Issue 4, PP. 855-859 August 2011.

Kumar and Mehta [8] proposed the texture-based technique in this paper. They
highlighted the effects of segmentation if the tumor tissue edges aren't shrill. The
performance of the proposed technology may get unwilling results due to those edges.
The texture evaluation and seeded region approach turned into executed inside the
MATLAB environment.

 Mahmoud, Dalia & Mohamed, Eltaher. (2012). Brain Tumor Detection

Using Artificial Neural Networks. Journal of Science and Technology. 13.
31-39.

Dalia Mahmoud et al. [9] presented a model using Artificial Neural Networks for tumor
detection in brain images. They implemented a computerized recognition system for
MR imaging the use of Artificial Neural Networks. That was observed that after the
Elman community was used during the recognition system, the period time and the

8
accuracy level were high, in comparison with other ANNs systems. This neural
community has a sigmoid characteristic which elevated the extent of accuracy of the
tumor segmentation.

 Marroquin J.L., Vemuri B.C., Botello S., Calderon F. (2002) An Accurate

and Efficient Bayesian Method for Automatic Segmentation of Brain MRI.
In: Heyden A., Sparr G., Nielsen M., Johansen P. (eds) Computer Vision
— ECCV 2002. ECCV 2002. Lecture Notes in Computer Science, vol 2353.
Springer, Berlin, Heidelberg.

L. Marroquin et al. [10] presented the automated 3d segmentation for brain MRI scans.
Using a separate parametric model in preference to a single multiplicative magnificence
will lessen the impact on the intensities of a grandeur. Brain atlas is hired to find non-
rigid conversion to map the usual brain. This transformation is further used to segment
the brain from nonbrain tissues, computing prior probabilities and finding automatic
initialization and finally applying the MPM-MAP algorithm to find out optimal
segmentation. Major findings from the study show that the MPM-MAP algorithm is
comparatively robust than EM in terms of errors while estimating the posterior
marginal. For optimal segmentation, the MPM-MAP algorithm involves only the
solution of linear systems and is therefore computationally efficient.

 Minz, Astina, and Chandrakant Mahobiya. “MR Image Classification

Using Adaboost for Brain Tumor Type.” 2017 IEEE 7th International
Advance Computing Conference (IACC) (2017): 701-705.

Astina minz et al. [11] implemented an operative automatic classification approach for
brain image that projected the usage of the AdaBoost gadget mastering algorithm. The
proposed system includes three main segments. Pre-processing has eradicated noises in
the datasets and converted images into grayscale. Median filtering and thresholding
segmentation are implemented in the pre-processed image.

 Monica Subashini.M, Sarat Kumar Sahoo, “Brain MR Image

Segmentation for TumorDetection using Artificial Neural Networks,”
International Journal of Engineering and Technology (IJET), Vol.5, No 2,
Apr-May 2013.

9
Monica Subashini and Sarat Kumar Sahoo [12] has suggested a technique for detecting
the tumor commencing the brain MR images. They also worked on different techniques,
which include pulse-coupled Neural Network and noise removal strategies for
reinforcing the mind MRI images and backpropagation network for classifying the brain
MRI images from tumor cells. They observed image enhancement and segmentation of
the usage of their proposed technique, and the backpropagation network helps in the
identification of a tumor in a brain MR image.

 S. Li, J.T. Kwok, I.W Tsang, and Y. Wang, ―Fusing Images with Different
Focuses using Support Vector Machines, Proceedings of the IEEE
transaction on Neural Networks, China, November 2007.

Li et al. [13] report that edge detection, image segmentation, and matching are not easy
to achieve in optical lenses that have long focal lengths. Previously, researchers have
proposed many techniques for this mechanism, one of which is wavelet-based image
fusion. The wavelet function can be improved by applying a discrete wavelet frame
transform (DWFT) and a support vector machine (SVM). In this paper, the authors
experimented with five sets of 256-level images. Experimental results show that this
technique is efficient and more accurate as it does not get affected by consistency
verification and activity level measurements. However, the paper is limited to only one
task related to fusion, and dynamic ranges are not considered during the calculation.

 H. Yu and J.L. Fan, ―Three-level Image Segmentation Based on

Maximum Fuzzy Partition Entropy of 2-D Histogram and Quantum
Genetic Algorithm, Advanced Intelligent Computing Theories, and
Applications. With Aspects of Artificial Intelligence. Lecture Notes in
Computer Science, Berlin, Heidelberg 2008.

Yu et al. [14] state that image segmentation is used for extracting meaningful objects
from an image. They propose segmenting an image into three parts, including dark, grey
and white. Z-function and s-function are used for the fuzzy division of the 2D
histogram. Afterward, QGA is used for finding a combination of 12 membership
parameters, which have a maximum value. This technique is used to enhance image
segmentation and the significance of their work is that three-level image segmentation
is used by following the maximum fuzzy partition of 2D Histograms. QGA is selected

10
for the optimal combination of parameters with the fuzzy partition entropy. The
proposed method of fuzzy partition entropy of 2D histogram generates better
performance than one-dimensional 3-level thresh holding method. Somehow, a large
number of possible combinations of 12 parameters in a multi-dimensional fuzzy
partition are used, and it is practically not feasible to compute each possible value;
therefore, QGA can be used to find the optimal combination.

 P.S. Mukambika, K Uma Rani, “Segmentation and Classification of MRI

Brain Tumor,” International Research Journal of Engineering and
Technology (IRJET), Vol.4, Issue 7, 2017, pp. 683 – 688, ISSN: 2395-0056

Mukambika et al. [15] proposed methodology for the subsequent stage’s classification
of the tumor, whether it is present or not. Their proposed work represents the
comparative study of strategies used for tumor identification from MR images, namely
the Level set approach and discrete wavelength transforms (DWT) and K-method
segmentation algorithms. After that phase, feature extraction is done followed SVM
classification.

 Pan, Yuehao & Huang, Weimin & Lin, Zhiping & Zhu, Wanzheng &
Zhou, Jiayin & Wong, Jocelyn & Ding, Zhongxiang. (2015). Brain tumor
grading based on Neural Networks and Convolutional Neural Networks.
Conference proceedings: Annual International Conference of the IEEE
Engineering in Medicine and Biology Society. IEEE Engineering in
Medicine and Biology Society. Conference. 2015. 699-702.
10.1109/EMBC.2015.7318458.

Yuehao Pan et al., [16] has used brain MRI pix for getting useful statistics for classifying
brain tumor. In their proposed method, they used Convolutional Neural Networks
(CNN) algorithms for developing a brain tumor detection system. The performance of
their CNN report is measured primarily based on sensitivity and specificity parameters,
which have stepped forward when in comparison to the Artificial Neural Networks
(ANN).

 S. Pereira, A. Pinto, V. Alves, and C. A. Silva, "Brain Tumor Segmentation

Using Convolutional Neural Networks in MRI Images," in IEEE
Transactions on Medical Imaging, vol. 35, no. 5, pp. 1240-1251, May 2016.

11
S. Pereira et al. [17] presented that magnetic resonance prevents physical segmentation
time in the medical areas. So, an automatic and reliable segmentation technique for
identifying abnormal tissues by using Convolutional Neural Network (CNN) had been
proposed in the research work. The massive three-dimensional and underlying
roughness amongst brain images makes the process of segmenting the image a severe
issue, so a robust methodology such as CNN is used.

 S. Roy And S. K. Bandyopadhyay, “Detection and Qualification Of Brain

Tumor From MRI Of Brain And Symmetric Analysis,” International
Journal Of Information And Communication Technology Research,
Volume 2 No.6, June 2012, Pp584-588

Roy et al. (2012) [18] calculated the tumor affected area for proportioned analysis. They
confirmed its software with numerous statistics groups with distinctive tumor sizes,
intensities, and location. They showed that their algorithm could robotically hit upon
and phase the brain tumor from the given photo. Image pre-processing consists of
fleeting that pictures to the filtering technique to remove distractors found in given
pictures. They first detect the tumor, segment it and then find out the area of tumor. One
of the important aspects is that after performing the quantitative analysis, we can
identify the status of an increase in the disease. They have suggested multi-step and
modular approach to solve the complex MRI segmentation problem. Tumor detection
is the first step in tumor segmentation. They have obtained good results in complex
situations. The authors claim that MRI segmentation is one of the essential tasks in the
medical area but boring and time-consuming if it is performed manually, so visually
study of MRI is more interesting and faster.

 Sankari Ali, and S. Vigneshwari. “Automatic tumor segmentation using

convolutional neural networks.” 2017 Third International Conference on
Science Technology Engineering & Management (ICONSTEM) (2017):
268-272.

A. Sankari and S. Vigneshwari [19] has proposed a Convolutional Neural Network

(CNN) segmentation, which principally based on the brain tumor classification method.
The proposed work used the non-linearity activation feature that's a leaky rectified linear
unit (LReLU). They primarily focused on necessary capabilities, which include mean

12
and entropy of the image and analyzed that the CNN algorithm is working higher for
representing the complicated and minute capabilities of brain tumor tissues present in
the MR Images.

 T.U Paul and S.K. Bandyopadhyay, ―Segmentation of Brain Tumor from

Brain MRI Images Reintroducing K – Means with advanced Dual
Localization +MethodTuhin, ‖ International Journal of Engineering
Research and Applications, Volume 3, Issue 1, June 2012, ISSN 2278-0882.

T.U Paul and S.K. Bandyopadhyay [20] has presented the brain segmentation that has
automated the use of the Dual Localization technique. In the initial phase, the skull
masks are generated for the brain MR images. The tumor areas are improvised using
the K-manner procedure. In the final step of their proposed work, they evaluated by its
dimensions such as length and breadth.

 Vaishali et al. (2015) Wavelet-based feature extraction for brain tumor

diagnosis—a survey. Int J Res Appl Sci Eng Technol (IJRASET) 3(V),
ISSN: 2321-9653

Vaishali [21] proposed a method that includes step by step procedure starting with
image pre-processing followed by extraction of useful objects and finally classification
of tumor region. Pre-processing is completed to enhance the image using eliminating
the noise via making use of Gaussian filters from the authentic ones. The next step is
feature extraction, in which a magnified image is used to extract the feature using a
symlet wavelet technique. The very last step is the classification of tumors by the use
of a Support vector machine (SVM).

 Varuna Shree, N., Kumar, T.N.R. Identification and classification of brain

tumor MRI images with feature extraction using DWT and probabilistic
neural network. Brain Inf. 5, 23–30 (2018) DOI:10.1007/s40708-017-0075-
5

Kumar and Varuna Shree [22] proposed work for the detection tumor region using
discrete wavelength transforms (DWT). This work consists of three phases, namely an
image enhancement using filtering technique, gray-level co-incidence matrix (GLCM)
feature extraction of tumor in addition to DWT based tumor location developing

13
segmentation. It is used to improve overall performance and reduce complexity. The
denoised accompanied by the aid of morphological filtering operations which put off
the noises that can be even shaped subsequent segmentation technique. The PNN
classifier is to use for classifying the abnormality, which is trained by different datasets,
and the accuracy is measured within the detection of tumor region of mind MR images.

 Vinotha, K., 2014. “Brain Tumor Detection and Classification Using

Histogram Equalization and Fuzzy Support Vector Machine Approach,”
International Journal of Engineering and Computer Science ISSN2319-
7242 3(5): 5823-5827.

K. Vinotha et al. [23] proposed brain tumor detection and the usage of the Histogram
Equalization (HE) and the Fuzzy Support Vector Machine (FSVM) classification
techniques. The brain MR image is pre-processed with histogram equalization and
segmented the apprehensive components from the photo primarily based on the MRF
algorithm for segmentation technique. MRF approach expanded the tumor
segmentation accuracy through which the overall performance of the proposed approach
changed into advanced.

 Sing, J.K. & Basu, D.K. & Nasipuri, Mita & Kundu, Megha. (2003).
Improved k-means algorithm in the design of RBF neural networks. 2. 841
- 845 Vol.2. 10.1109/TENCON.2003.1273297.

Sing et al. [24] propose a fuzzy adaptive RBI based neural network for MR brain image
segmentation. The hidden layer neuron of FARBF-NN neurons has been fuzzified to
reduce noise effect. Basu et al. assert that the medical image segmentation approach
involves a combination of texture and boundary information. The authors maintain that
geometric algebra can be used to obtain volumetric data representation using spheres,
nonrigid registration of spheres and real-time object tracking. Major contribution of the
proposed approach is that the use of the marching cube algorithm reduces the number
of primitives to model volumetric data and uses a lesser number of primitives for the
registration process, and thus makes the registration process faster. However, the study
has employed images obtained from CT scans, which has its own limitations like blurred
boundaries and similar grey levels between healthy and non-healthy tissues.

14
  Shi, Z., He, L., Suzuki, K., Nakamura, T., & Itoh, H. (2009). Survey on
Neural Networks Used for Medical Image Processing. International
Journal of computational science, 3(1), 86–100.

Shi et al. [25] employed neural networks for medical image processing, including the
key features of medical image pre-processing, segmentation, and object detection and
recognition. The study employed Hopfield and feedforward neural networks. The feed-
forward and Hopfield neural networks are simple to use and easy to implement. The
added advantage of Hopfield neural networks is that it does not require pre-experimental
knowledge. The time required to resolve image processing predicament is substantially
reduced by using a trained neural network

 Detection of Tumor in MRI Images Using Artificial Neural Networks

Automatic defects detection in MR images is very important in many diagnostic and

therapeutic applications. This work has introduced one automatic brain tumor detection
method to increase the accuracy and yield and decrease the diagnosis time. The goal is
classifying the tissues into two classes of normal and abnormal. MR images that have
been used here are MR images from normal and abnormal brain tissues. This method
uses from neural network to do this classification. The purpose of this project is to
classify the brain tissues to normal and abnormal classes automatically, which saves the
radiologist time, increases accuracy and yield of diagnosis.

 Survey on Brain Tumor Detection Techniques Using Magnetic Resonance

Images

The brain tumor is an abnormal growth of cells inside the skull which causes damage
to the other cells necessary for functioning human brain. Brain tumor detection is a
challenging task due to the complex structure of the human brain. MRI images
generated from MRI scanners using strong magnetic fields and radio waves to form
images of the body which helps for medical diagnosis. This paper gives an overview of
the various techniques used to detect the tumor in the human brain using MRI images.

15
  A Neural Network-based Method for Brain Abnormality Detection in MR
Images Using Gabor Wavelets

Nowadays, automatic defects detection in MR images is very important in many

diagnostic and therapeutic applications. This paper introduces a Novel automatic brain
tumor detection method that uses T1, T2_weighted and PD, MR images to determine
any abnormality in brain tissues. Here, it has been tried to give a clear description from
brain tissues using Gabor wavelets, energy, entropy, contrast and some other statistic
features such as mean, median, variance, correlation, values of maximum and minimum
intensity. It is used from a feature selection method to reduce the feature space too. this
method uses from neural network to do this classification. The purpose of this project
is to classify the brain tissues to normal and abnormal classes automatically, which
saves the radiologist time, increases accuracy and yield of diagnosis.

16
3. CHALLENGES IN TUMOR CLASSIFICATION

The identification of tumor is a very challenging task. The location, shape and
the structure of tumor varies significantly from patient to patient which makes the
segmentation a very challenging task. In the figure shown below, we have shown some
images of the same brain slice from different patients, which clearly reflect the variation
of the tumor. We can clearly see that the location of the tumor is different in all the 8
images/patients shown below. To make it worse, the shape and the intra-tumoral
structure is also different for all the eight patients/images. In fact, there can be more
than one region of the tumor as can be seen from the images below. This indeed reflects
the complexity of automatic segmentation.

Fig.3.1 Location of tumors in eight different images.

17
4. MODULE DIVISON

This provides the architecture of the system that would be developed by our
hands. It consists of six steps where the execution starts from taking an input image
from the data set followed by the image pre-processing, image enhancement, Image
segmentation using binary thresholding and the brain tumor classification using
Convolutional Neural Network. Finally, the output is observed after all the above-
mentioned steps are completed.

Each module is unique in its own way. Every step has its importance. This
architecture also includes a testing and training data set. The data set used is has been
downloaded from Kaggle which consists of nearly 2000 images that are used to test and
train the system. The input image is pre-processed by using the noise filter like Median
Filter and Bilateral Filter and the image is enhanced using the Sobel Filter. Then the
obtained image using segmented using binary thresholding and morphological
operations are performed on it. Finally, the image classification is done using
Convolutional Neural Network to predict whether the tumor is present or not.

18
 Fig 4.1 Module Division

MODULE 1: IMAGE PREPROCESSING AND IMAGE

ENHANCEMENT

4.1.1 IMAGE PREPROCESSING:

The Brain MRI image dataset has been downloaded from the Kaggle. The MRI
dataset consists of around 1900 MRI images, including normal, benign, and malignant.
These MRI images are taken as input to the primary step. The pre-processing is an
essential and initial step in improving the quality of the brain MRI Image. The critical
steps in pre-processing are the reduction of impulsive noises and image resizing. In the
initial phase, we convert the brain MRI image into its corresponding gray-scale image.
The removal of unwanted noise is done using the adaptive bilateral filtering technique
to remove the distorted noises that are present in the brain picture. This improves the
diagnosis and also increase the classification accuracy rate.

19
4.1.1.1 IMAGE ACQUISITION FROM DATASET:

In image processing, image acquisition is done by retrieving an image from

dataset for processing. It is the first step in the workflow sequence because, without an
image no processing is possible. The image that is acquired is completely unprocessed.
Here we process the image using the file path from the local device.

4.1.1.2 CONVERT THE IMAGE FROM ONE COLOR SPACE TO ANOTHER:

There are more than 150 color-space conversion methods available in OpenCV.
For color conversion, we use the function cv2.cvtColor(input_image, flag) where flag
determines the type of conversion. In our work, we convert the input image into the
gray-scale image.

4.1.1.3 FILTERS:

In image processing, filters are mainly used to suppress the high frequencies in
the image.

Median filter: It is a non-linear filtering technique used to remove noise from

the images. It is performed by sorting all the pixel values from the window into
numerical order and then replacing the pixel being considered with the median pixel
value. This filter removes the speckle noise and salt and pepper noise through ‘ON’ and
‘OFF’ of pixels by white and dark spots.

Bilateral filter: It is also a non-linear, noise-reducing smoothing filter for

images. It replaces the intensity of each pixel with a weighted average of intensity
values from nearby pixels. This weight is based on the Gaussian distribution. Bilateral
filtering smooth images while conserving edges utilizing a nonlinear grouping of
neighbouring image pixels. This filtering technique is simple, local, and concise. It
syndicates a grey level grounded on their likeness and the symmetrical nearness and
chooses near vales to farther values in both range and domain.

20
4.1.2 IMAGE ENHANCEMENT:

Image enhancement is a technique used to improve the image quality and

perceptibility by using computer-aided software. This technique includes both objective
and subjective enhancements. This technique includes points and local operations. The
local operations depend on the district input pixel values. Image enhancement has two
types: spatial and transform domain techniques. The spatial techniques work directly on
the pixel level, while the transform technique works on Fourier and later on the spatial
technique.

Edge detection is a segmentation technique that uses border recognition of strictly

linked objects or regions. This technique identifies the discontinuity of the objects. This
technique is used mainly in image study and to recognize the parts of the image where
a huge variation in intensity arises.

4.1.2.1 SOBEL FILTER:

The Sobel filter is used for edge detection. It works by calculating the gradient of
image intensity at each pixel within the image. It is widely used in image analysis to
help locate edges in images. Sobel operator is used for segmentation purpose. This
technique can be dependent on the central difference which tends toward the central
pixels on average. This technique can be expressed as 3 × 3 matric to the first derivative
of the Gaussian kernel. It combines smoothing and differentiation. For Sobel edge
detection the gradient of the image is calculated for each pixel position in the image.

1. We calculate two derivatives:

a. Horizontal changes: This is computed by convolving I with a

kernel Gx with odd size. For example, for a kernel size of 3, Gx would be
computed as:

Gx = [[-1 0 +1]

[-2 0 +2]

[-1 0 +1]]

21
 b. Vertical changes: This is computed by convolving I with a
kernel Gy with odd size. For example, for a kernel size of 3, Gy would be
computed as:

Gy = [[-1 -2 -1]

[ 0 0 0]

[+1 +2 +1]]

2. At each point of the image we calculate an approximation of the gradient in that

point by combining both results above:

G=(G2x+G2y)1/2

3. Although sometimes the following simpler equation is used:

G=|Gx|+|Gy|

After the completion of the pre-processing, the image will be free from the
noises, but we still need to enhance the image since the obtained image is smoothened,
edges may not be preserved, and the image will be dull. To overcome all these, we used
edge detection called Sobel filtering technique. The whole thing is done by calculating
the gradient of image intensities at each pixel within the image. It is widely used in
image analysis to help locate edges in images. It will also enhance the darker areas of
the image, slightly increase contrast and as sharp as possible.

MODULE 2: IMAGE SEGMENTATION USING BINARY

THRESHOLD

Image segmentation is a technique of segregating the image into many parts.

The basic aim of this segregation is to make the images easy to analyze and interpret
with preserving the quality. This technique is also used to trace the objects’ borders
within the images. This technique labels the pixels according to their intensity and
characteristics. Those parts represent the entire original image and acquire its
characteristics such as intensity and similarity. The image segmentation technique is
used to create contours of the body for clinical purposes. Segmentation is used in

22
machine perception, malignant disease analysis, tissue volumes, anatomical and
functional analyses, virtual reality visualization, and anomaly analysis, and object
definition and detection.

Segmentation methods has ability to detect or identify the abnormal portion

from the image which is useful for analyzing the size, volume, location, texture and
shape of the extracted image. MR image segmentation with the aid of preserving the
threshold information, which is convenient to identify the broken regions extra
precisely. It was a trendy surmise that the objects that are placed in close propinquity
might be sharing similar houses and characteristics.

4.2.1 THRESHOLDING:

Thresholding is the simplest method of image segmentation. It is a non-linear

operation that converts a grey-scale image into a binary image where the two levels are
assigned to pixels that are below or above the specified threshold value. In Open CV,
we use cv2.threshold() function:

cv2.threshold(src, thresh, maxval, type[dst])

This function applies fixed-level thresholding to a single-channel array. The

function is typically used to get a bi-level (binary) image out of a grayscale image for
removing a noise, that is, filtering out pixels with too small or too large values. “maxval”
is the set threshold value which compares with input values, when the input is greater
than the set threshold value it gives output is set maxval value and it is shown with white
color in gray images. when the input pixel intensity values are less than the set threshold,
its output is black color. There are several types of thresholding supported by the
function.

The function returns the computed threshold value and thresholder image.

1. src - input array (single-channel, 8-bit or 32-bit floating point). This is the
source image, which should be a grayscale image.

2. thresh - threshold value, and it is used to classify the pixel values.

23
 3. maxval - maximum value to use with the THRESH_BINARY and
THRESH_BINARY_INV thresholding types. It represents the value to be given if pixel
value is more than (sometimes less than) the threshold value.

4. type - thresholding type

 cv2.THRESH_BINARY
 cv2.THRESH_BINARY_INVY

4.2.2 MORPHOLOGICAL OPERATIONS:

Morphological operations apply a structuring element to an input image,

creating an output image of the same size. In a morphological operation, the value of
each pixel in the output image is based on a comparison of the corresponding pixel in
the input image with its neighbors.

The Morphological techniques are also used with segmentation techniques. The
morphological action is normally performed on binary images. It processes the
operations based on shape and it has a wide set of the image processing operation.
Erosion and Dilation are two methods of morphological operations which used in this
proposed work. We perform both Erosion and dilation operations used together.

Two main steps of the erosion and dilation morphological operation are opening
and closing. The first step is the opening of the MRI binary image. The main work of
opening operation is open up a gap which is present in between object and connect that
to a small collection of pixels. After setting of the bridge, the erosion again restored
with their actual size using dilation. If the binary image has been opened then the
subsequent opened same structured elements have not affected on that image. After
completing the opening operations next step is the closing operation. Based on the
closing operation while keeping the original region sizes, the erosion and dilation can
handle different hole in the image region. Dilation and Erosion are the basic
morphological operations. Dilation adds pixels to the boundaries of objects in an image,
while erosion removes pixels on object boundaries.

24
 Watershed Method: considers the gradient magnitude of an image as a
topographic surface where high gradient denotes peaks, while low gradient denotes
valleys. Start by filling every isolated valley with different coloured water. As the water
rises, water from different valleys will start to merge. To avoid that, barriers are built in
the locations where water merges. Continue the work of filling water and building
barriers until all the peaks are under water. Then the created barriers give the
segmentation result.

MODULE 3: BRAIN TUMOR IMAGE CLASSIFICATION USING

CONVOLUTIONAL NEURAL NETWORK

Classification is the best approaches for identification of images like any kind
of medical imaging. All classification algorithms are based on the prediction of image,
where one or more features and that each of these features belongs to one of several
classes.

An automatic and reliable classification method Convolutional Neural Network

(CNN) will be used since it is robust in structure which helps in identifying every minute
details. A Convolutional Neural Network (ConvNet/CNN) is a Deep Learning
algorithm which can take in an input image, assign importance to various
aspects/objects in the image and be able to differentiate one from the other. The pre-
processing required in a ConvNet is much lower as compared to other classification
algorithms. While in primitive methods filters are hand-engineered, with enough
training, ConvNet have the ability to learn these filters/characteristics.

A ConvNet is able to successfully capture the spatial and temporal dependencies

in an image through the application of relevant filters. The architecture performs a better
fitting to the image dataset due to the reduction in the number of parameters involved
and reusability of weights. In other words, the network can be trained to understand the
sophistication of the image better. The role of the ConvNet is to reduce the images into
a form which is easier to process, without losing features which are critical for getting
a good prediction.

For this step we need to import Keras and other packages that we’re going to use in
building the CNN. Import the following packages:

25
 Sequential is used to initialize the neural network.

 Convolution2D is used to make the convolutional network that deals with the
images.

 MaxPooling2D layer is used to add the pooling layers.

 Flatten is the function that converts the pooled feature map to a single column that
is passed to the fully connected layer.

 Dense adds the fully connected layer to the neural network.

4.3.1 SEQUENTIAL:

 To initialize the neural network, we create an object of the Sequential class.

 classifier = Sequential ()

4.3.2 CONVOLUTION:

 To add the convolution layer, we call the add function with the classifier object
and pass in Convolution2D with parameters. The first
argument feature_detectors which is the number of feature detectors that we
want to create. The second and third parameters are dimensions of the feature
detector matrix.
 We used 256 feature detectors for CNNs. The next parameter is input
shape which is the shape of the input image. The images will be converted into
this shape during pre-processing. If the image is black and white it will be
converted into a 2D array and if the image is coloured it will be converted into a
3D array.
 In this case, we’ll assume that we are working with coloured
images. Input_shape is passed in a tuple with the number of channels, which is 3
for a coloured image, and the dimensions of the 2D array in each channel. If you
are not using a GPU it’s advisable to use lower dimensions to reduce the
computation time. The final parameter is the activation function. Classifying
images is a nonlinear problem. So, we use the rectifier function to ensure that we
don’t have negative pixel values during computation. That’s how we achieve
non-linearity.

26
  classifier.add (Convolution2D (256, 3, 3, input_shape = (256, 256, 3),
activation=’relu’))

4.3.3 POOLING:

 The Pooling layer is responsible for reducing the spatial size of the convolved
feature. This is to decrease the computational power required to process the data
through dimensionality reduction. Furthermore, it is useful for extracting
dominant features which are rotational and positional invariant, thus maintaining
the process of effectively training of the model.
 There are two types of Pooling: Max Pooling and Average Pooling.
Max Pooling returns the maximum value from the portion of the image
covered by the Kernel. On the other hand, Average Pooling returns the average
of all the values from the portion of the image covered by the Kernel. Generally,
we use max pooling.
 In this step we reduce the size of the feature map. Generally, we create a pool
size of 2x2 for max pooling. This enables us to reduce the size of the feature map
while not losing important image information.
 classifier.add (MaxPooling2D (pool_size= (2,2)))

4.3.4 FLATTENING:

 In this step, all the pooled feature maps are taken and put into a single vector for
inputting it to the next layer.
 The Flatten function flattens all the feature maps into a single long column.
 classifier.add (Flatten ())

4.3.5 FULLY CONNECTION:

 The next step is to use the vector we obtained above as the input for the neural
network by using the Dense function in Keras. The first parameter
is output which is the number of nodes in the hidden layer. You can determine
the most appropriate number through experimentation. The higher the number of
dimensions the more computing resources you will need to fit the model. A
common practice is to pick the number of nodes in powers of two.
 classifier.add (Dense (output = 64))

27
 The next layer we have to add is the output layer. In this case, we’ll use
the sigmoid activation function since we expect a binary outcome. If we expected
more than two outcomes, we would use the SoftMax function.
 The output here is 1 since we just expect the predicted probabilities of the classes.
 classifier.add (Dense (output=1, activation=’sigmoid’))

28
5. EXPERIMENTAL ANALYSIS AND RESULTS

5.1 SYSTEM CONFIGURATION

5.1.1 SOFTWARE REQUIREMENTS

 Windows: Python 3.6.2 or above, PIP and NumPy 1.13.1

Python:

Python is an interpreted, high-level, general purpose programming language

created by Guido Van Rossum and first released in 1991, Python's design philosophy
emphasizes code Readability with its notable use of significant Whitespace. Its
language constructs and object-oriented approach aim to help programmers write
clear, logical code for small and large-scale projects. Python is dynamically typed
and garbage collected. It supports multiple programming paradigms, including
procedural, object-oriented, and functional programming.

PIP:

It is the package management system used to install and manage software

packages written in Python.

NumPy:

NumPy is a general-purpose array-processing package. It provides a high-

performance multidimensional array object, and tools for working with these arrays.
It is the fundamental package for scientific computing with Python. It contains various
features including these important ones:

 A powerful N-dimensional array object

 Sophisticated (broadcasting) functions

 Tools for integrating C/C++ and Fortran code

 Useful linear algebra, Fourier transform, and random number capabilities

29
Pandas:

Pandas is the most popular python library that is used for data analysis. It
provides highly optimized performance with back-end source code is purely written
in C or Python. We can analyze data in pandas with

1. Series
2. Data frames

Anaconda:

Anaconda is a free and open-source distribution of the Python and R

programming languages for scientific computing that aims to simplify package
management and deployment. Package versions are managed by the package
management system conda. The Anaconda distribution includes data-science
packages suitable for Windows, Linux, and macOS. Anaconda distribution comes
with 1,500 packages selected from PyPI as well as the conda package and virtual
environment manager. It also includes a GUI, Anaconda Navigator, as a graphical
alternative to the command-line interface (CLI).

Jupyter Notebook:

Anaconda distribution comes with 1,500 packages selected from PyPI as well
as the conda package and virtual environment manager. It also includes a
GUI, Anaconda Navigator, as a graphical alternative to the command line interface
(CLI). A Jupyter Notebook document is a JSON document, following a versioned
schema, and containing an ordered list of input/output cells which can contain code,
text mathematics, plots and rich media, usually ending with the “. ipynb" extension.

Tensor Flow:

Tensor flow is a free and open-source software library for dataflow and
differentiable programming across a range of tasks. It is a symbolic math library, and
is also used for machine learning applications such as neural networks. It is used for
both research and production at Google.

30
Keras:

Keras is an open-source neural-network library written in Python. It is capable

of running on top of TensorFlow, Microsoft Cognitive Toolkit, R, Theano, or Plaid
ML. Designed to enable fast experimentation with deep neural networks, it focuses
on being user-friendly, modular, and extensible. Keras contains numerous
implementations of commonly used neural-network building blocks such as
layers, objectives, activation functions, optimizers, and a host of tools to make
working with image and text data easier to simplify the coding necessary for writing
deep neural network code.

OpenCV:

OpenCV (Open source computer vision) is a library of programming functions

mainly aimed at real-time computer vision. Originally developed by Intel, it was later
supported by willow garage then Itseez (which was later acquired by Intel). The
library is cross platform and free for use under the open source BSD license. OpenCV
supports some models from deep learning frameworks like TensorFlow, Torch,
PyTorch (after converting to an ONNX model) and Caffe according to a defined list
of supported layers. It promotes Open Vision Capsules. which is a portable format,
compatible with all other formats.

5.1.2 HARDWARE CONFIGURATION

 Processor: Intel core i5 or above.

 64-bit, quad-core, 2.5 GHz minimum per core
 Ram: 4 GB or more
 Hard disk: 10 GB of available space or more.
 Display: Dual XGA (1024 x 768) or higher resolution monitors
 Operating system: Windows

31
5.2 SAMPLE CODE AND RESULTS

#SEGMENT
import cv2
# matplotlib is used for displaying images
import matplotlib.pyplot as plt
%matplotlib inline
# numpy is used for matrix manipulations
import numpy as np

path = r’D: \images\yes\Y32.jpg'

orig_img = cv2.imread(path,1) # 1 indicates color image
# OpenCV uses BGR while Matplotlib uses RGB format
# Display the color image with matplotlib
plt.imshow(cv2.cvtColor(orig_img, cv2.COLOR_BGR2RGB))
plt.axis('off')
plt.show()

Fig 5.1 Input Image

gray_img = cv2.cvtColor( orig_img, cv2.COLOR_BGR2GRAY )

plt.imshow(gray_img,cmap='gray')
plt.axis('off')
plt.show()

32
Fig 5.2 Gray-scale Image

# To remove salt and pepper noise #Using 5*5 kernel

filtered = cv2.bilateralFilter(gray_img, 5,10,2.5)
plt.imshow(filtered,cmap='gray')
plt.axis('off')
plt.show()

Fig 5.3 Filtered Image

# 3*3 Sobel Filters

Gx= np.array([[-1, 0, 1], [-2, 0, 2], [-1, 0, 1]])
print ("Gx \n", Gx)
Gy = np.array([[-1, -2, -1], [0, 0, 0], [1, 2, 1]])
print ("Gy \n", Gy)
img_sobelx = cv2.Sobel(filtered,cv2.CV_8U,1,0,ksize=3)

33
 img_sobely = cv2.Sobel(filtered,cv2.CV_8U,0,1,ksize=3)
#del f = Gx + Gy
# Adding mask to the image
img_sobel = img_sobelx + img_sobely+gray_img
plt.imshow(img_sobel,cmap='gray')
plt.axis('off')
plt.show()

Fig 5.4 Image after edge detection

# Set threshold and maxValue

threshold = 50
maxValue = 255
# Threshold the pixel values
th, thresh = cv2.threshold(img_sobel, threshold, maxValue,
cv2.THRESH_BINARY)
plt.imshow(thresh,cmap='gray')
plt.axis('off')
plt.show()

34
Fig 5.5 Thresholded and Binary Image

# To remove any small white noises in the image using morphological opening.
kernel = np.ones((3,3),np.uint8)
opening = cv2.morphologyEx(thresh, cv2.MORPH_OPEN, kernel, iterations = 2)
plt.imshow(opening,cmap='gray')
plt.axis('off')
plt.show()

Fig 5.6 Morphological Image

# White region shows sure foreground area
sure_bg = cv2.dilate(opening,kernel,iterations=3)
dist_transform = cv2.distanceTransform(opening, cv2.DIST_L2, 5)
ret, sure_fg = cv2.threshold(dist_transform,0.7*dist_transform.max(),255,0)
plt.imshow(sure_fg,cmap='gray')
plt.axis('off')
plt.show()

35
Fig 5.7 Dilated Image

#TRAINING DATA
import tensorflow as tf
from tensorflow.keras.datasets import cifar10
from tensorflow.keras.preprocessing.image import ImageDataGenerator
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense, Dropout, Activation, Flatten
from tensorflow.keras.layers import Conv2D, MaxPooling2D

import pickle
pickle_in = open("XX.pickle","rb")
X = pickle.load(pickle_in)
pickle_in = open("YY.pickle","rb")
y = pickle.load(pickle_in)

X = X/255.0
model = Sequential()

model.add(Conv2D(256, (3, 3), input_shape=X.shape[1:]))

model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))

36
model.add(Conv2D(256, (3, 3)))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))

model.add(Flatten())

model.add(Dense(64))
model.add(Dense(1))
model.add(Activation('sigmoid'))

model.compile(loss='binary_crossentropy', optimizer='adam',metrics=['accuracy'])
model.fit(X, y, batch_size=164, epochs=10, validation_split=0.3)

Fig 5.8 Training data history

37
#DATALOAD

import numpy as np

import matplotlib.pyplot as plt

import os

import cv2

from tqdm import tqdm

DATADIR = "D:\dataset1"

CATEGORIES = ["no", "yes"]

for category in CATEGORIES:

path = os.path.join(DATADIR,category)

for img in os.listdir(path):

img_array = cv2.imread(os.path.join(path,img) ,cv2.IMREAD_GRAYSCALE)

# convert to array

plt.imshow(img_array, cmap='gray') # graph it

plt.show() # display!

break # we just want one for now so break

break #...and one more!

print(img_array)

print(img_array.shape)

38
 IMG_SIZE = 100

new_array = cv2.resize(img_array, (IMG_SIZE, IMG_SIZE))

plt.imshow(new_array, cmap='gray')

plt.show()

training_data = []

def create_training_data():

for category in CATEGORIES:

path = os.path.join(DATADIR,category) # create path

class_num = CATEGORIES.index(category) # get the classification (0 or a 1).

for img in tqdm(os.listdir(path)): # iterate over each image

try:

img_array = cv2.imread(os.path.join(path,img)
,cv2.IMREAD_GRAYSCALE) # convert to array

new_array = cv2.resize(img_array, (IMG_SIZE, IMG_SIZE)) # resize to

normalize data size

training_data.append([new_array, class_num]) # add this to our

training_data

except Exception as e: # in the interest in keeping the output clean...

pass

39
create_training_data()

print(len(training_data))

import random

random.shuffle(training_data)

for sample in training_data[:10]:

print(sample[1])

X = []

y = []

for features,label in training_data:

X.append(features)

y.append(label)

print(X[0].reshape(-1, IMG_SIZE, IMG_SIZE, 1))

X = np.array(X).reshape(-1, IMG_SIZE, IMG_SIZE, 1)

import pickle

pickle_out = open("XX.pickle","wb")

pickle.dump(X, pickle_out)

pickle_out.close()

40
pickle_out = open("YY.pickle","wb")

pickle.dump(y, pickle_out)

pickle_out.close()

41
5.3 EXPERIMENTAL RESULTS:

Sample Input:

Predicted Output: Yes

Observed Output:

42
Sample Input:

Predicted Output: No

Observed Output:

43
5.4 PERFORMANCE MEASURES:

The proposed algorithm has been assessed through various performance

evaluation metrics that include True Positive, True Negative the former one that
designates how many times does the proposed algorithm is able to correctly recognize
the damaged region as damaged region and the later one designates how many times
does the proposed algorithm correctly identified non-damaged region as non-damaged
region. And the False Positive (FN) and False Negative (FN) the former one designates
how many times does the proposed algorithm fails to recognize the damaged region
correctly, and the later represents how many times does the proposed algorithm fails to
identify the non-tumors region as non-tumors regions. Basing on values of TP, TN, FP,
and FN, the values of Accuracy, Specificity and sensitivity are calculated of the
proposed algorithm.

𝑻𝑷 + 𝑻𝑵
𝑨𝒄𝒄𝒖𝒓𝒂𝒄𝒚 =
𝑻𝑷 + 𝑻𝑵 + 𝑭𝑷 + 𝑭𝑵

𝑻𝑵
𝑺𝒑𝒆𝒄𝒊𝒇𝒊𝒄𝒊𝒕𝒚 =
𝑻𝑵 + 𝑭𝑷

𝑻𝑵
𝑺𝒆𝒏𝒔𝒊𝒕𝒊𝒗𝒊𝒕𝒚 =
𝑻𝑷+𝑭𝑵

44
5.5 PERFORMANCE EVALUTION:

On experimentation, it was observed that the proposed methodology seems to

be outperformed when compared to all different set of images. Among all the images,
the proposed Convolutional Neural Network (CNN) based approach seems too much
better in terms of quality of the output in 128 *128 images when compared to its other
sized images which are represented in table and charts.

TABLE 1 Represents the true positive, true negative, false positive and false negative
values of the proposed approach for different set of images.

False False
Different set of True True
Positive Negative
Images Positive (%) Negative (%)
(%) (%)
128 * 128 Images 83.7 84.5 16.3 15.5
256 * 256 Images 82.4 84.1 17.6 15.9
512 * 512 Images 82.1 83.7 17.9 16.3

Performance Analysis Chart

False Negative (%)

False Positive (%)

True Negative (%)

True Positive (%)

0 10 20 30 40 50 60 70 80 90

512 * 512 Images 256 * 256 Images 128 * 128 Images

Fig. 5.9 Represents the performance analysis of CNN

It is observed from table 2 upon performing proposed segmentation technique

for different set of images that have the ability to recognize the isolated region from the
MR images that are used to analyze the shape and size of the denoised image. We have
used Convolutional Neural Network (CNN) for segmentation, and the output of our
proposed work is pleased with better accuracy, sensitivity, and computational time.

45
 TABLE 2 Represents the Accuracy, Sensitivity, and Specificity of the proposed
approach for different set of images.

Different set of
Accuracy (%) Sensitivity (%) Specificity (%)
Images

128 * 128 Images 84.1 84.4 83.8

256 * 256 Images 83.3 83.4 82.7
512 * 512 Images 82.9 83.3 82.4

Accuracy, Sensitivity and Specificity of the Proposed

approach.
85
84.5
84
83.5
83
82.5
82
81.5
81
Accuracy (%) Sensitivity (%) Specificity (%)

128 * 128 Images 256 * 256 Images 512 * 512 Images

Fig. 5.10 Represents the performance of proposed CNN

46
6. CONCLUSION AND FUTURE SCOPE

6.1 CONCLUSION:

We proposed a computerized method for the segmentation and identification of

a brain tumor using the Convolution Neural Network. The input MR images are read
from the local device using the file path and converted into grayscale images. These
images are pre-processed using an adaptive bilateral filtering technique for the
elimination of noises that are present inside the original image. The binary thresholding
is applied to the denoised image, and Convolution Neural Network segmentation is
applied, which helps in figuring out the tumor region in the MR images. The proposed
model had obtained an accuracy of 84% and yields promising results without any errors
and much less computational time.

6.2 FUTURE SCOPE:

It is observed on extermination that the proposed approach needs a vast training

set for better accurate results; in the field of medical image processing, the gathering of
medical data is a tedious job, and, in few cases, the datasets might not be available. In
all such cases, the proposed algorithm must be robust enough for accurate recognition
of tumor regions from MR Images. The proposed approach can be further improvised
through in cooperating weakly trained algorithms that can identify the abnormalities
with a minimum training data and also self-learning algorithms would aid in enhancing
the accuracy of the algorithm and reduce the computational time.

47
REFERENCES

[1] A. Sivaramakrishnan And Dr.M.Karnan “A Novel Based Approach For

Extraction Of Brain Tumor In MRI Images Using Soft Computing Techniques,”
International Journal Of Advanced Research In Computer And Communication
Engineering, Vol. 2, Issue 4, April 2013.

[2] Asra Aslam, Ekram Khan, M.M. Sufyan Beg, Improved Edge Detection
Algorithm for Brain Tumor Segmentation, Procedia Computer Science, Volume
58,2015, Pp 430-437, ISSN 1877-0509.

[3] B.Sathya and R.Manavalan, Image Segmentation by Clustering Methods:

Performance Analysis, International Journal of Computer Applications (0975 – 8887)
Volume 29– No.11, September 2011.

[4] Devkota, B. & Alsadoon, Abeer & Prasad, P.W.C. & Singh, A.K. & Elchouemi,
A.. (2018). Image Segmentation for Early Stage Brain Tumor Detection using
Mathematical Morphological Reconstruction. Procedia Computer Science. 125. 115-
123. 10.1016/j.procs.2017.12.017.

[5] K. Sudharani, T. C. Sarma and K. Satya Rasad, "Intelligent Brain Tumor lesion
classification and identification from MRI images using k-NN technique," 2015
International Conference on Control, Instrumentation, Communication and
Computational Technologies (ICCICCT), Kumaracoil, 2015, pp. 777-780. DOI:
10.1109/ICCICCT.2015.7475384

[6] Kaur, Jaskirat & Agrawal, Sunil & Renu, Vig. (2012). A Comparative Analysis
of Thresholding and Edge Detection Segmentation Techniques. International Journal of
Computer Applications.vol. 39.pp. 29-34. 10.5120/4898-7432.

[7] Li, Shutao, JT-Y. Kwok, IW-H. Tsang and Yaonan Wang. "Fusing images with
different focuses using support vector machines." IEEE Transactions on neural
networks 15, no. 6 (2004): 1555-1561.

[8] M. Kumar and K. K. Mehta, "A Texture based Tumor detection and automatic
Segmentation using Seeded Region Growing Method," International Journal of

48
Computer Technology and Applications, ISSN: 2229-6093, Vol. 2, Issue 4, PP. 855-
859 August 2011.

[9] Mahmoud, Dalia & Mohamed, Eltaher. (2012). Brain Tumor Detection Using
Artificial Neural Networks. Journal of Science and Technology. 13. 31-39.

[10] Marroquin J.L., Vemuri B.C., Botello S., Calderon F. (2002) An Accurate and
Efficient Bayesian Method for Automatic Segmentation of Brain MRI. In: Heyden A.,
Sparr G., Nielsen M., Johansen P. (eds) Computer Vision — ECCV 2002. ECCV 2002.
Lecture Notes in Computer Science, vol 2353. Springer, Berlin, Heidelberg.

[11] Minz, Astina, and Chandrakant Mahobiya. “MR Image Classification Using
Adaboost for Brain Tumor Type.” 2017 IEEE 7th International Advance Computing
Conference (IACC) (2017): 701-705.

[12] Monica Subashini.M, Sarat Kumar Sahoo, “Brain MR Image Segmentation for
TumorDetection using Artificial Neural Networks,” International Journal of
Engineering and Technology (IJET), Vol.5, No 2, Apr-May 2013.

[13] S. Li, J.T. Kwok, I.W Tsang, and Y. Wang, ―Fusing Images with Different
Focuses using Support Vector Machines, Proceedings of the IEEE transaction on Neural
Networks, China, November 2007.

[14] H. Yu and J.L. Fan, ―Three-level Image Segmentation Based on Maximum Fuzzy
Partition Entropy of 2-D Histogram and Quantum Genetic Algorithm, Advanced
Intelligent Computing Theories, and Applications. With Aspects of Artificial
Intelligence. Lecture Notes in Computer Science, Berlin, Heidelberg 2008.

[15] P.S. Mukambika, K Uma Rani, “Segmentation and Classification of MRI Brain
Tumor,” International Research Journal of Engineering and Technology (IRJET),
Vol.4, Issue 7, 2017, pp. 683 – 688, ISSN: 2395-0056

[16] Pan, Yuehao & Huang, Weimin & Lin, Zhiping & Zhu, Wanzheng & Zhou,
Jiayin & Wong, Jocelyn & Ding, Zhongxiang. (2015). Brain tumor grading based on
Neural Networks and Convolutional Neural Networks. Conference
proceedings: Annual International Conference of the IEEE Engineering in Medicine

49
and Biology Society. IEEE Engineering in Medicine and Biology Society. Conference.
2015. 699-702. 10.1109/EMBC.2015.7318458.

[17] S. Pereira, A. Pinto, V. Alves, and C. A. Silva, "Brain Tumor Segmentation

Using Convolutional Neural Networks in MRI Images," in IEEE Transactions on
Medical Imaging, vol. 35, no. 5, pp. 1240-1251, May 2016.

[18] S. Roy And S.K.Bandyopadhyay, “Detection And Qualification Of Brain Tumor

From MRI Of Brain And Symmetric Analysis,” International Journal Of Information
And Communication Technology Research, Volume 2 No.6, June 2012, Pp584-588

[19] Sankari, Ali, and S. Vigneshwari. “Automatic tumor segmentation using

convolutional neural networks.” 2017 Third International Conference on Science
Technology Engineering & Management (ICONSTEM) (2017): 268-272.

[20] T.U Paul and S.K. Bandyopadhyay, ―Segmentation of Brain Tumor from Brain
MRI Images Reintroducing K – Means with advanced Dual Localization
MethodTuhin,‖ International Journal of Engineering Research and Applications,
Volume 3, Issue 1, June 2012, ISSN 2278-0882.

[21] Vaishali et al. (2015) Wavelet-based feature extraction for brain tumor
diagnosis—a survey. Int J Res Appl Sci Eng Technol (IJRASET) 3(V), ISSN: 2321-
9653

[22] Varuna Shree, N., Kumar, T.N.R. Identification and classification of brain tumor
MRI images with feature extraction using DWT and probabilistic neural network. Brain
Inf. 5, 23–30 (2018) doi:10.1007/s40708-017-0075-5

[23] Vinotha, K., 2014. “Brain Tumor Detection and Classification Using Histogram
Equalization and Fuzzy Support Vector Machine Approach,” International Journal of
Engineering and Computer Science ISSN2319-7242 3(5): 5823-5827.

[24] Sing, J.K. & Basu, D.K. & Nasipuri, Mita & Kundu, Megha. (2003). Improved
k-means algorithm in the design of RBF neural networks. 2. 841 - 845 Vol.2.
10.1109/TENCON.2003.1273297.

50
 [25] Shi, Z., He, L., Suzuki, K., Nakamura, T., & Itoh, H. (2009). Survey on Neural
Networks Used for Medical Image Processing. International Journal of computational
science, 3(1), 86–100.

[26] P. Naga Srinivasu, G. Srinivas, T Srinivas Rao, (2016). ‘An Automated Brain
MRI image segmentation using a Generic Algorithm and TLBO.’ International Journal
of Control Theory and Applications, Vol: 9(32).

[27] P. Naga Srinivasu, T. Srinivasa Rao, Valentina Emilia Balas. (2020). A

systematic approach for identification of tumor regions in the human brain through
HARIS algorithm, Deep Learning Techniques for Biomedical and Health Informatics,
Academic Press. Pages 97-118. https://doi.org/10.1016/B978-0-12-819061-6.00004-5.

51
 Proceedings of the International Conference on Inventive Computing and Informatics (ICICI 2017)
IEEE Xplore Compliant - Part Number: CFP17L34-ART, ISBN: 978-1-5386-4031-9

A Study on Brain Tumor Segmentation Using

Convolution Neural Network

Anil Singh Parihar

Department of Information Technology
Delhi Technological University
New Delhi, India
Email: [email protected]

Abstract— Computer vision is playing important role in the data. Neurosurgeons normally require accurate depiction of the
field of human health care. This role is growing day by day. The tumor before diagnosing the tumor and performing the surgery.
application of computer vision techniques in health care has one Doctors/ specialists perform manual brain tumor segmentation,
of the aim to reduce human judgement in diagnosis. Thus, but its time-consuming. Hence, to design and develop
human error in judgement may be reduced. Brain related automated brain tumor segmentation methods, various
diagnosis demands at most care and a minute error in judgment researchers have been working continually. Gliomas are the
may be disastrous. This makes medical imaging very important most usual and vigorous brain tumors, which in their highest
field. Various imaging methods like CT Scans, X-Ray, and MRI grade can lead to a short life expectancy [3]. Thus, the key
are available but MRI is the most reliable and safe. Even the
stage to boost the life quality of patients is planning the
smallest aberrances in the human body can be identified using
treatment. The automatic segmentation of Magnetic Resonance
imaging techniques. More preferred contrast information about
brain tissues is provided by Magnetic Resonance imaging (MRI). brain images can be enacted through a convolutional neural
Image segmentation is an important problem in medical imaging, network (CNN). The medical image analysis field has made
which involves separating the tumor and organisms out of the CNNs very popular. On comparison to old machine learning
medical data. Machine learning (ML) has gained enormous methods, there is no requirement of hand-crafted features in
application with innovation in hardware requirements for CNNs which are a requisite for classification. Instead,
computing. Convolutional neural networks (CNN) is one of the convolution kernel sets are learned that are specially trained for
most effective techniques in ML. CNN has find applications in the problem of classification. The old machine learning
almost every field of research. CNN also find effective methods have been using kernels like Gaussian or Haar-like for
applications in brain MRI segmentation. In this paper, we getting the occurrence detail, but CNNs pick up the kernel set
present a study on CNN based MRI segmentation. based on the training data as provided. Like this, the
information relevant to the task can be automatically extracted
Keywords— Brain tumor, Segmentation, medical, imaging, by the system [4].
Convolution, MRI
II. CONVOLUTIONAL NEURAL NETWORK
I. INTRODUCTION
Convolution is performed on the artificial neural networks
An important role is played by Medical imaging in current (ANN) that gives us the Convolutional neural network [4].
medical research and clinical practice. Tumor detection mainly Neurons having weights and biases, which can be learned,
involves MRI, Computed tomography (CT) and ultrasound forms a CNN. Three main layers used in building up the
images [1]. Precise extraction of tumor is necessary and convolutional neural network architecture are- Convolutional
advantageous. When we have huge number of sequence of layer, pooling layer, and fully connected layer. A convolutional
images, the manual observation of tumors is time-consuming, neural network (CNN) is so named as it contains one or more
and frequently, the depiction quality depends on the operators. convolutional layers. In the input images, certain local features
So, for clinical analysis, medical image segmentation has are detected through convolutional layers [5]. There is a
gained much attention and different segmentation methods connection between every node of a convolutional layer and a
have been proposed. Imaging techniques allow medical subset of neurons which are connected spatially. This helps in
practitioners and researchers, even before performing invasive detecting the local forms (structures) in the channels of input
surgery, to assess activities and disorders in the human brain. image. The weights on the connections are shared among the
The technique magnetic resonance imaging (MRI) is of good convolutional layer's nodes, to search for the similar local trait
quality among other techniques as it gives better contrast in the input channels. Every shared weight set is known as a
details about the brain tissues from a variety of image kernel (convolution kernel). Across the input images, the local
sequences [2]. The neurosurgeons and medical scientists have features (whose strength is visible in the feature map) to be
great opportunity with the increase in brain MR image data. detected are learned by convolutional layer having kernels. A
The diagnosis using computer aids and treatment can be pooling layer is the next layer after convolution layer in CNN
administered through the analysis and processing of image whose main purpose is decreasing the size of the representation

978-1-5386-4031-9/17/$31.00 ©2017 IEEE 198

 Proceedings of the International Conference on Inventive Computing and Informatics (ICICI 2017)
IEEE Xplore Compliant - Part Number: CFP17L34-ART, ISBN: 978-1-5386-4031-9

spatially, which further helps reducing the parameter numbers IL={pc1,ip10,ip20,....,ip90,pc2}for every sequence are learned out
and, network’s computational complexity, and also helps in of the training data set to perform normalization. In this, pc1
controlling the overfitting [6]. To reduce the feature maps' size, and pc2 are taken for each MRI sequence and at the lth
the max-pooling layer is used, which selects the maximum percentile, intensity is represented as ipl. After using this
feature response among the local neighborhoods that can be training data, next step in the intensity normalization method is
overlapping or non-overlapping. This is done by disposing the transforming the original intensities linearly among two
precise location of maximum responses. CNNs usually have a landmarks into the learned landmarks. Like this, every
number of convolutional and pooling layers. A number of fully sequence can have the similar histogram across various
connected (FC) layers will follow other layers as mentioned. subjects. After the normalization of the MRI images, mean
Finally, CNN contains a Softmax layer, or regression layer, value of intensity and standard deviation are calculated. These
that generates the wanted outputs. Back-propagation algorithm are calculated over all the training patches which are taken for
are used to train the CNN’s, in the similar way as multilayer each image sequence. Then, to have mean to be zero and
perceptions are trained. Further, the following cost function is variance to be unit, the patches on each image sequence are
minimized which is done with respect to the weights W that are normalized.
unknown.
2) Convolutional Neural Network
X
1 The convolutional layers are applied to convolve an image
L 
X
¦ ln(p(y
i
i
X i )) (1)
(or signal) with kernels to result into the feature maps. Thus,
the previous layer is connected though the kernel weights to an
Where number of training images are represented by |X|, element in the feature map. During the training phase, the
the ith training image represents Xi and yi is the label, and the kernels' weights are adjusted to improve the input
probability by which Xi is correctly classified is given by characteristics by using backpropagation method. Since all the
p(yi|Xi). The weights in lst convolutional layer are Wlt, at the units of the similar feature maps share the kernels, the
iteration t, and the cost over a small batch of size N is convolutional layers will be having lesser weights for training
represented by L’, then the updated weights in the next iteration
than the FC layers which are dense, and thus making CNN less
can be calculated by using:
susceptible to overfitting and easy to train. The similar feature
«tN / x ¼» can be observed irrespective of the location, making it
Jt J¬
translation invariance, as over all the image, the similar kernel
wL is convolved. The information of the neighborhood is extracted
Vi t 1 PVi t  J tD i (2)
wWi by using kernels. On each neural unit’s output, a non-linear
activation function is implemented.
Wi t 1 Wi t  Vi t 1
where αl is the lth layer learning rate, μ is the momentum Architecture of CNN: This architecture aims at a
which specify the earlier updated weight in the current segmentation method which is reliable. But, the brain tumors
iteration, and γ specifies the scheduling rate. show a great variability in the intra-tumoral structures, thus
making the problem of segmentation very challenging [9].
III. BRAIN TUMOR SEGMENTATION USING CNN CNN is designed here to solve this complex problem and
making the transformation of intensity normalization, tuned for
In this work, we presented study of two CNN based brain each grade of tumor. The tumor can be classified as Low Grade
MRI segmentation: Brain Tumor Segmentation Using Gliomas (LGG) and High Grade Gliomas (HGG) [3]. The
Convolutional Neural Networks in MRI Images (BTCNN) [7] HGG's architecture is deeper than LGG's as going deeper did
and Automatic Segmentation of MR Brain Images with a not give better results. More layers need to be included with the
Convolutional Neural Network (ASCNN) [8] weights to go deeper, which may result in the increase of
overfitting. This arises a need for Dropout within the LGG.
A. BTCNN While dropout is present in HGG, because the database that is
This work represents the method for segmentation of brain used for the purpose of assessment, was having more HGG
tumor [8] in MRI images with the use of convolutional than cases of LGG. Moreover, the HGG and LGG have
neural network. In this paper, the author uses the deep different appearance and patterns. For segmentation, a precise
architectures of CNN with convolutional kernels, which are perception about the location is needed. Pooling gives
small, for segmentation of gliomas in MRI images. Fig.2 affirmative effect so as to remove the irrelevant details.
presents the method of segmenting brain tumor using CNN. Although, it can have a bad effect by removing the relevant
Pre-processing, classification using CNN and post- details. The feature maps need to be stuffed just before
processing are the three main stages in the process of convolution. This is done to make the feature maps in output to
segmentation. have the similar dimensions. LReLU (Leaky Rectifier Linear
Unit) [10] is the activation function used in all layers carrying
1) Pre-Processing: The bias field distortion can alter the weights, having a single exception of the last layer which uses
MRI images which can vary the intensity of similar tissues in softmax. The activation function is given:
the complete image. Thus, the intensity normalization method
is applied so as to form the similar ranges of intensity and f (x) max(0, x)  D min(0, x) (3)
contrast, across various patients. A set of intensity landmarks

978-1-5386-4031-9/17/$31.00 ©2017 IEEE 199

 Proceedings of the International Conference on Inventive Computing and Informatics (ICICI 2017)
IEEE Xplore Compliant - Part Number: CFP17L34-ART, ISBN: 978-1-5386-4031-9

Where α is leakiness parameter. Dropout has been used only in by getting the responses which are highest, using the max-
the fully connecter layers. The main responsibility of an pooling. As the patch dimensions gets decremented in each
activation function is transforming the data, non-linearly. layer, the number of kernels that are trained gets incremented.
For each and every input patch size, separate fully connected
Training of CNN: For the CNN to be trained, the loss function (FC) layers are used, just after convolution layers. Further,
need to be minimized. The loss function is given as: these layers are then connected to a softmax output layer for
H ¦¦ c j , k log(cˆ j , k ) (4) performing the final classification. Due to their speed in
j k training CNNs, for every node rectified linear units (ReLU) are
used. Drop-out is used on the fully connected (FC) layers so as
where c is target and ĉ is its probabilistic prediction. to decrease the consequence of overfitting on the training data
set [13].
3) Post-Processing
In this step, the clusters which are small are to be classified
as tumor. Then, the Volumetric constrains are imposed by IV. COMPARATIVE ANALYSIS
removing the clusters obtained in the segmentation by the
In BTCNN [7], CNN is used for segmenting the brain tumor
CNN that are lesser than a preset threshold.
in MRI images, in which initial stage involves intensity
B. ASCNN normalization of acquisitions from multiple scanners, of MRI
The paper ASCNN [11] presents a method, which is images. In this paper, brain tumor segmentation is done on the
based on a multiple scale CNN, for the automatic basis of grades of tumors whether they are HGG or LGG i.e.
segmentation of magnetic resonance images of brain into a high grade or low grade. Also, the prospect of deep
number of classes. To acquire accurate segmentation details architectures is investigated through the small kernels. This is
and spatial consistency, the multi-scale approach is used. The done by the comparison of deep CNN with shallow
proposed method, unlike the previous work for brain tumor architectures. Shallow architectures were found to have lower
segmentation using CNN, permits to eliminate the explicitly functioning, even though a large number of feature maps were
defined spatial features. The multiple patch and kernel sizes used [14] [15]. Ultimately, it was confirmed that in the
are used together in this method. The method need to learn the training of CNN, the activation function LReLU was of great
multi-scale features which determine both the intensity and importance than ReLU.
spatial characteristics. The other approaches of multi-scale
While in ASCNN [11], automatic segmentation of magnetic
CNNs, as compared using the multiple patch and kernel sizes,
resonance brain images has been assessed on manually
provide the multiple scale features by using the feature maps
segmented preterm neonatal and adult images. The evaluation
as extra input for the fully connected (FC) layer. This is done
was done for images acquired at distinct ages and using
directly after the first convolution layer.
distinct protocols for acquisition. Also, different patch sizes
Here, the neonatal images have N being set to 9 with are used and each patch has an eye on a different aspect of
classes of tissue being 8 and background. N is the number of segmentation problem. The patch which is smaller permits for
output classes. Similarly, the ageing adult images have N being comprehensive analysis but it misses the spatial consistency.
set to 8 with classes of tissue being 7 and background, and to In contrast, the patch which is larger results in smoother
the young adult images having N being set to 7 with 6 classes segmentation, but it misses the smaller details.
of tissue and background. Max-pooling is executed only for the The evaluation parameters used in [7] are Dice Similarity
two greatest sizes of patches, after the third convolution layer. Coefficient (DSC), Positive Predictive Value (PPV) and
Sensitivity. While in [11], there are two parameters used for
Moreover, the images of developing neonates can also be
evaluation, Dice coefficient and mean surface distance. These
segmented using the same procedure [12]. This can be done at
are calculated between the manual and automatic segmentation.
distinct ages. Also, it is applicable to coronal and axial images.
These parameters have been compared on the basis of different
CNN will be used to classify, each voxel in the image, into the
classes i.e. class is based on grade or different ages.
classes of brain tissue. Image patches provide information
about each voxel that where the voxel that is important is in the
center. For using the multiple scale information for every V. CONCLUSION
voxel, patch sizes in multiple number are used. The larger The methods of Brain Tumor Segmentation using
scales can easily recognize the voxel located in the image. Convolutional Neural Network (CNN) in MR Images has been
While detailed information is provided by smaller scales i.e. for surveyed. Image segmentation is an important problem in
the local neighborhood of a voxel. For every patch size, the medical imaging which involves segmenting organisms and
kernel sizes that are trained are not same. For the bigger tumors from medical data. The methods and their results
patches, the bigger kernel sizes are used. A distinct network havebeen studied for brain tumor segmentation using CNN.
branch is used for every patch sizes and the layer which is only First Method involves brain tumor segmentation that tackles
shared is the output layer. For every patch and kernel size, the the diversity in multiple sites’, multiple scanner MRI images,
weights and biases are optimized specifically. As multiple using the method of intensity normalization which shows that it
number of convolutional layers are used, the output images are is relevant for good segmentation using MRI images Also, it
subsampled after every convolution. The subsampling is done compares the results of parameters for different grades of

978-1-5386-4031-9/17/$31.00 ©2017 IEEE 200

 Proceedings of the International Conference on Inventive Computing and Informatics (ICICI 2017)
IEEE Xplore Compliant - Part Number: CFP17L34-ART, ISBN: 978-1-5386-4031-9

tumor classes i.e. high grade tumor or low grade tumor. Second Training or Fine Tuning?", IEEE TRANSACTIONS ON MEDICAL
Method shows segmentation results which are accurate for IMAGING, VOL. 35, NO. 5, MAY 2016.
images obtained at distinct ages and using distinct acquisition [5] A.S. Razavian, H. Azizpour, J. Sullivan, and S. Carlsson, “CNN features
off-the-shelf: An astounding baseline for recognition,” in Proc. IEEE
protocols. It also compared certain parameters for different Conf. Comput. Vis. Pattern Recognit. Workshops, 2014, pp. 512–519.
classes. [6] Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,” Nature, vol. 521,
no. 7553, pp. 436–444, 2015.
REFERENCES [7] Sérgio Pereira, Adriano Pinto,et al, "Brain Tumor Segmentation
Using Convolutional Neural Networks in MRI Images", IEEE
TRANSACTIONS ON MEDICAL IMAGING, VOL. 35, NO. 5, MAY
2016.
[1] Silvia Franchini, Antonio Gentile, Filippo Sorbello, Giorgio Vassallo, [8] Liang Lin, Wei Yang,et al, "Inference With Collaborative Model for
and Salvatore Vitabile, "ConformalALU: A Conformal Geometric Interactive Tumor Segmentation in Medical Image Sequences", IEEE
Algebra Coprocessor for Medical Image Processing", IEEE TRANSACTIONS ON CYBERNETICS, VOL. 46, NO. 12,
TRANSACTIONS ON COMPUTERS, VOL. 64, NO. 4, APRIL 2015. DECEMBER 2016.
[2] Shang-Ling Jui, Shichen Zhang, Weilun Xiong, Fangxiaoqi Yu, [9] Bjoern H. Menze*, Andras Jakab,et al,"The Multimodal Brain Tumor
Mingjian Fu, and Dongmei Wang, "Brain MRI Tumor Segmentation Image Segmentation Benchmark (BRATS)", IEEE TRANSACTIONS
with 3D Intracranial Structure Deformation Features", IEEE Intelligent ON MEDICAL IMAGING, VOL. 34, NO. 10, OCTOBER 2015.
Systems,V.31 n.2, p.66-76, March 2016. [10] A. L. Maas, A. Y. Hannun, and A. Y. Ng, “Rectifier nonlinearities
[3] S.Bauer et al., “ A Survey of MRI-based medical image analysis for improve neural network acoustic models,” in Proc. ICML, 2013, vol. 30.
brain tumor studies”, Phys. Med. Biol., vol. 58, no. 13, pp. 97-129, [11] Pim Moeskops, Max A. Viergever et al, "Automatic Segmentation of
2013. MR Brain Images With a Convolutional Neural Network", IEEE
[4] Nima Tajbakhsh, Jae Y. Shin, Suryakanth R. Gurudu, R. Todd Hurst, et TRANSACTIONS ON MEDICAL IMAGING, VOL. 35, NO. 5, MAY
al, "Convolutional Neural Networks for Medical Image Analysis: Full 2016.

978-1-5386-4031-9/17/$31.00 ©2017 IEEE 201

 Available online at www.sciencedirect.com

ScienceDirect
Future Computing and Informatics Journal 3 (2018) 68e71
http://www.journals.elsevier.com/future-computing-and-informatics-journal/

Classification using deep learning neural networks for brain tumors

Heba Mohsen a,*, El-Sayed A. El-Dahshan b,c, El-Sayed M. El-Horbaty d, Abdel-Badeeh M. Salem d
a
Faculty of Computers and Information Technology, Future University, Cairo, Egypt
b
Egyptian E-Learning University, Giza, Egypt
c
Faculty of Science, Ain Shams University, Cairo, Egypt
d
Faculty of Computer and Information Sciences, Ain Shams University, Cairo, Egypt
Received 26 October 2017; accepted 5 December 2017
Available online 24 December 2017

Abstract

Deep Learning is a new machine learning field that gained a lot of interest over the past few years. It was widely applied to several ap-
plications and proven to be a powerful machine learning tool for many of the complex problems. In this paper we used Deep Neural Network
classifier which is one of the DL architectures for classifying a dataset of 66 brain MRIs into 4 classes e.g. normal, glioblastoma, sarcoma and
metastatic bronchogenic carcinoma tumors. The classifier was combined with the discrete wavelet transform (DWT) the powerful feature
extraction tool and principal components analysis (PCA) and the evaluation of the performance was quite good over all the performance
measures.
Copyright © 2017 Faculty of Computers and Information Technology, Future University in Egypt. Production and hosting by Elsevier B.V. This
is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).

Keywords: Machine learning; Deep learning; Deep neural network; Discrete wavelet transform; Principle component analysis; Fuzzy c-means; Magnetic resonance
images

1. Introduction called primary malignant tumor or to be originated elsewhere

in the body and spread to the brain which called secondary
Brain is one of the most complex organs in the human body malignant tumor [3e5].
that works with billions of cells. A brain tumor arise when Brain magnetic resonance imaging (MRI) is one of the best
there is uncontrolled division of cells forming an abnormal imaging techniques that researchers relied on for detecting the
group of cells around or inside the brain. That group of cells brain tumors and modeling of the tumor progression in both
can affect the normal functionality of the brain activity and the detection and the treatment phases. MRI images have a big
destroy the healthy cells [1,2]. Brain tumors classified to impact in the automatic medical image analysis field for its
benign or low-grade (grade I and II) and malignant tumors or ability to provide a lot of information about the brain structure
high-grade (grade III and IV). Benign tumors are non- and abnormalities within the brain tissues due to the high
progressive (non-cancerous) so considered to be less aggres- resolution of the images [3,6e8]. In fact, Researchers pre-
sive, they originated in the brain and grows slowly; also it sented different automated approaches for brain tumors
cannot spread to anywhere else in the body. However, ma- detection and type classification using brain MRI images since
lignant tumors are cancerous and grow rapidly with undefined it became possible to scan and load medical images to the
boundaries. They can be originated in the brain itself which computer. However, Support Vector Machine (SVM) and
Neural Networks (NN) are the widely used approaches for
* Corresponding author. their good performance over the last few decades [9]. But
E-mail address: [email protected] (H. Mohsen).
Peer review under responsibility of Faculty of Computers and Information
recently, deep learning (DL) models set an exciting trend in
Technology, Future University in Egypt. machine learning as the deep architecture can efficiently

https://doi.org/10.1016/j.fcij.2017.12.001
2314-7288/Copyright © 2017 Faculty of Computers and Information Technology, Future University in Egypt. Production and hosting by Elsevier B.V. This is an
open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
 H. Mohsen et al. / Future Computing and Informatics Journal 3 (2018) 68e71 69

represent complex relationships without requiring a huge

number of nodes like in the shallow architectures e.g. SVM
and K-nearest neighbor (KNN). For that reason, they grew
rapidly to become the state of the art in different health
informatics areas such as bioinformatics, medical informatics
and medical image analysis [7,9,10].
The contribution of this paper is applying the deep learning
concept to perform an automated brain tumors classification
using brain MRI images and measure its performance. The
proposed methodology aims to differentiate between normal
brain and some types of brain tumors such as glioblastoma,
sarcoma and metastatic bronchogenic carcinoma tumors using
brain MRI images. The proposed methodology uses a set of
features extracted by the discrete wavelet transform (DWT) Fig. 1. DNNs architecture.
feature extraction technique from the segmented brain MRI im-
ages, to train the DNN classifier for brain tumors classification.
The structure of this paper is organized as follows: in input flows from the input layer to the output layer through
Section 2 there's an overview on the deep learning concept and number of hidden layers which are more than two layers [13].
architecture, section 3 described the steps of the proposed Fig. 1 illustrates the typical architecture for DNNs where Ni is
methodology, section 4 presents the experimental results and the input layer contains of neurons for the input features, No is
discussion and the conclusion and future work is given in the output layer contains neurons for the output classes and
section 5. Nh,l are the hidden layers.

2. Overview on deep learning 3. Methodology

Deep learning (DL) is a subfield of machine learning based Our proposed methodology based on the DNN learning
on learning multiple levels of representations by making a architecture for classification where the classifier is identifying
hierarchy of features where the higher levels are defined from the brain tumors in brain MRIs.
the lower levels and the same lower level features can help in The proposed methodology for classifying the brain tumors
defining many higher level features [11]. DL structure extends in brain MRIs is as follows:
the traditional neural networks (NN) by adding more hidden
layers to the network architecture between the input and Step 1: Brain MRIs Dataset acquisition
output layers to model more complex and nonlinear relation- Step 2: Image segmentation using Fuzzy C-means
ships. This concept gained the researchers interest in the recent Step 3: Feature extraction using discrete wavelet transform
years for its good performance to become the best solution in (DWT) and reduction using Principle component
many problems in medical image analysis applications such as analysis (PCA) technique
image denoising, segmentation, registration and classification Step 4: Classification using DNN
[7,10e13].
There are various DL architectures, convolutional neural
networks (CNN) is a common used architecture in recent years 3.1. Data acquisition
that can perform complex operations using convolution filters
[7,9,10]. A typical CNN architecture is a sequence of feed- According to the World Health Organization (WHO) clas-
forward layers implementing convolutional filters and pooling sification system to identify brain tumors, there are more than
layers, after the last pooling layer CNN adopts several fully- 120 types of brain tumors which differ in origin, location, size,
connected layers that work on converting the 2D feature characteristics of the tumor tissues [15,16]. In this paper we were
maps of the previous layers into 1D vector for classification concerning with three types of malignant tumors which are:
[10]. Even though the CNN architecture has an advantage of
doesn't require a feature extraction process before being  Glioblastoma: primary malignant brain tumors that are
applied but training a CNN from scratch is a time consuming classified as Grade IV and developed from star-shaped
and difficult as it needs a very large labeled dataset for cells, called astrocytes that support nerve cells. It usually
building and training before the model is ready for classifi- starts in the cerebrum.
cation which is not always available. Moreover the hardware  Sarcoma: has different grades that vary from grade I to
requirements for processing the large number of filters for the grade IV and it arises in the connective tissues like blood
large size of images e.g. 256
256 [7,10,14]. vessels.
Deep Neural Network (DNN) is another DL architecture  Metastatic bronchogenic carcinoma: secondary malignant
that is widely used for classification or regression with success brain tumors that was spread to the brain from broncho-
in many areas. It's a typical feedforward network which the genic carcinoma lung tumor.
70 H. Mohsen et al. / Future Computing and Informatics Journal 3 (2018) 68e71

3.3. Feature extraction and reduction

After segmenting the Brain MR images into 5 sections

features of the segmented tumor is extracted using discrete
wavelet transform (DWT). DWT has the advantage of
extracting the most relevant features at different directions
and scales as they provide localized time-frequency infor-
mation of a signal using cascaded filter banks of high-pass
Fig. 2. Brain MRIs dataset sample. and low-pass filters to extract features in a hierarchy
manner [18].
Fig. 4 shows a 2-levels DWT decomposition of an image
The dataset consists of 66 real human brain MRIs with 22 where the functions h(n) and g(n) represent the coefficients
normal and 44 abnormal images which are glioblastoma, of the high-pass and low-pass filters, respectively. As a
sarcoma and metastatic bronchogenic carcinoma tumors result, there are four sub-band (LL, LH, HH, HL) images at
collected from Harvard Medical School website (http://med. each level. The LL subband can be regarded as the approx-
harvard.edu/AANLIB/) [15]. All the brain MRIs was in axial imation component of the image, while the LH, HL, HH
plane, T2-weighted and 256
256 pixel. A sample of the subbands can be regarded as the detailed components of the
dataset is illustrated in Fig. 2. image [18,19].
Our methodology utilizes a 3-levels decomposition of Haar
3.2. Image segmentation wavelet which was also used in our previous work [18] to
extract 32
32 ¼ 1024 features for each brain MRI. Although
Image segmentation is the non-trivial task of separating the this number is not so big compared to the number of feature
different normal brain tissues such as gray matter (GM), white maps resulted by the convolution filters of CNNs but we used
matter (WM) and cerebrospinal fluid (CSF) and the skull from the principal components analysis (PCA) [18] to approximate
the tumor tissues in brain MR images [17] as the resulted the original extracted features with lower dimensional feature
segmented tumor part only would be used in the next steps. In vectors.
this work we used the Fuzzy C-means clustering technique to
segment the image into 5 sections as it had good results in our 3.4. Classification
previous work and also for comparison purposes [18]. Fig. 3
shows the results of segmenting a sample image using Fuzzy After the features are extracted and selected, the classifi-
C-means. cation step using DNN is performed on the resulted feature
vector. Classification is performed by using 7-fold cross vali-
dation technique for building and training the DNN of 7 hid-
den layers structure. Also for evaluating the performance of
the selected classifier, we employed other machine learning
classification algorithms from WEKA [20] using the same
criteria. The selected classification algorithms are KNN with
K ¼ 1 and 3, Linear discriminant analysis (LDA) and from our
previous work [18] SMO-SVM.

4. Experimental results and discussion

Fig. 3. A sample image segmented using FCM.

The experiment took place using two tools:

Fig. 4. 2-levels DWT decomposition of an image.

 H. Mohsen et al. / Future Computing and Informatics Journal 3 (2018) 68e71 71

Table 1 large size images (256
256). In addition using the DNN
Performance of DNN, KNN K ¼ 1 and 3, LDA and SMO classifiers. classifier shows high accuracy compared to traditional clas-
Algorithm Classification Recall Precision F-Measure AUC (ROC) sifiers. The good results achieved using the DWT could be
rate employed with the CNN in the future and compare the
DNN 96.97% 0.97 0.97 0.97 0.984 results.
KNN K ¼ 1 95.45% 0.955 0.956 0.955 0.967
KNN K ¼ 3 86.36% 0.864 0.892 0.866 0.954
LDA 95.45% 0.955 0.957 0.955 0.983
References
a
SMO 93.94% 0.939 0.941 0.963 0.939
a [1] Kavitha AR, Chitra L, kanaga R. Brain tumor segmentation using genetic
Previous work ref [18].
algorithm with SVM classifier. Int J Adv Res Electr Electron Instrum Eng
2016;5(3):1468e71.
1 [2] Logeswari T, Karnan M. An improved implementation of brain tumor
detection using segmentation based on hierarchical self organizing map.
Int J Comput Theory Eng 2010;2(4):591e5.
0.95 [3] Khambhata Kruti G, Panchal Sandip R. Multiclass classification of brain
tumor in MR images. Int J Innov Res Comput Commun Eng 2016;4(5):
8982e92.
0.9 [4] Kaur G, Rani J. MRI brain tumor segmentation methods-a review. Int J
Comput Eng Technol (IJCET) 2016;6(3):760e4.
[5] Das V, Rajan J. Techniques for MRI brain tumor detection: a survey. Int J
0.85 Res Comput Appl Inf Tech 2016;4(3):53e6.
[6] Zacharaki EI, Wang S, Chawla S, Soo Yoo D, Wolf R, Melhem ER, et al.
Classification of brain tumor type and grade using MRI texture and shape
0.8 in a machine learning scheme. Magn Reson Med 2009;62:1609e18.
Classificaon Recall Precision F-Measure ROC [7] Litjens G, Kooi T, Bejnordi BE, Setio AA, Ciompi F, Ghafoorian M,
rate et al. A survey on deep learning in medical image analysis. Med Image
Anal 2017;42:60e88.
[8] Singh L, Chetty G, Sharma D. A novel machine learning approach for
DNN KNN K=1 KNN K=3 LDA SMO detecting the brain abnormalities from MRI structural images. In: IAPR
international conference on pattern recognition in bioinformatics. Berlin
Fig. 5. Comparison graph for the performance of DNN, KNN K ¼ 1 and 3, Heidelberg: Springer; 2012. p. 94e105.
LDA and SMO classifiers. [9] Pan Y, Huang W, Lin Z, Zhu W, Zhou J, Wong J, et al. Brain tumor
grading based on neural networks and convolutional neural networks. In:
Engineering in medicine and biology society (EMBC), 37th annual in-
 We prepared the brain MRI dataset and performed the first ternational conference of the IEEE; 2015. p. 699e702.
three steps of the methodology using MATLAB R2015a [10] Ravì D, Wong C, Deligianni F, Berthelot M, Andreu-Perez J, Lo B, et al.
 Weka 3.9 tool was used for performing the classifications Deep learning for health informatics. IEEE J Biomed Health Inf 2017;
and the evaluation of the selected classifiers. 21(1):4e21.
[11] Tharani S, Yamini C. Classification using convolutional neural network
for heart and diabetics datasets. Int J Adv Res Comp Commun Eng 2016;
The evaluation of the performance for the proposed meth- 5(12):417e22.
odology was measured in terms of average classification rate, [12] Le QVA. Tutorial on deep learning - Part 1: nonlinear classi- fiers and the
average recall, average precision, average F-Measure and backpropagation algorithm. 2015. http://robotics.stanford.edu/~quocle/
average area under the ROC curve (AUC) of all the four tutorial1.pdf.
[13] Anuse A, Vyas V. A novel training algorithm for convolutional neural
classes (normal, glioblastoma, sarcoma and metastatic bron- network. Contr Intell Syst 2016;2(3):221e34.
chogenic carcinoma tumors) and compared to the performance [14] Ahmed KB, Hall LO, Goldgof DB, Liu R, Gatenby RA. Fine-tuning
of other classifiers in the same terms. convolutional deep features for MRI based brain tumor classification. In:
As seen from Table 1 and the chart in Fig. 5, the DNN Medical Imaging 2017: Computer-Aided Diagnosis, Vol. 10134. Inter-
classifier gave good results combined with the DWT feature national Society for Optics and Photonics; 2017 Mar 3. 101342E.
[15] Suhag Sonu, Saini Lalit Mohan. Automatic brain tumor detection and
extraction tool in all the performance measures over all other classification using svm classifier. In: Proceedings of ISER 2nd inter-
classifiers. national conference, Singapore; July 2015. p. 55e9.
[16] http://braintumor.org/brain-tumor-information/understanding-brain-tumors/
5. Conclusion and future work tumor-types/.
[17] Gordillo N, Montseny E, Sobrevilla P. State of the art survey on MRI
brain tumor segmentation. Magn Reson Imag 2013;31(8):1426e38.
In this paper we proposed an efficient methodology which [18] Mohsen, H., El-Dahshan, E.A., El-Horbaty, E.M., Salem, A.M. Brain
combines the discrete wavelet transform (DWT) with the tumor type classification based on support vector machine in magnetic
Deep Neural Network (DNN) to classify the brain MRIs into resonance images, Annals Of “Dunarea De Jos” University Of Galati,
Normal and 3 types of malignant brain tumors: glioblastoma, Mathematics, Physics, Theoretical mechanics, Fascicle II, Year IX (XL),
sarcoma and metastatic bronchogenic carcinoma. The new No. 1; 2017.
[19] Ahmad M, Hassan M, Shafi I, Osman A. Classification of tumors in
methodology architecture resemble the convolutional neural human brain MRI using wavelet and support vector machine. IOSR J
networks (CNN) architecture but requires less hardware Comput Eng 2012;8(2):25e31.
specifications and takes a convenient time of processing for [20] http://www.cs.waikato.ac.nz/ml/weka/.
 AN AUTOMATED SEGMENTATION OF BRAIN
MR IMAGE THROUGH FUZZY RECURRENT
NEURAL NETWORK
Jalluri Gnana Siva Sai1, P Naga Srinivasu2, Munjila Naga Sindhuri3,
Rohitha Kola4, Sreesailam Deepika5

1
Dept. of CSE, Anil Neerukonda Institute of Technologies and Sciences,
Visakhapatnam, India
[email protected],
2
Dept. of CSE, Anil Neerukonda Institute of Technologies and Sciences,
Visakhapatnam, India
[email protected],
3
Dept. of CSE, Anil Neerukonda Institute of Technologies and Sciences,
Visakhapatnam, India
[email protected],
4
Dept. of CSE, Anil Neerukonda Institute of Technologies and Sciences,
Visakhapatnam, India
[email protected],
5
Dept. of CSE, Anil Neerukonda Institute of Technologies and Sciences,
Visakhapatnam, India
[email protected]

Abstract--The human brain is the major controller of the humanoid

system. The abnormal growth and division of cells in the brain lead to a brain
tumor, and the further growth of brain tumors leads to brain cancer. In the
area of human health, Computer Vision plays a significant role, which
reduces the human judgment that gives accurate results. CT scans, X-Ray,
and MRI scans are the common imaging methods among magnetic resonance
imaging (MRI) that are the most reliable and secure. MRI detects every
minute objects. Our paper aims to focus on the use of different techniques for
the discovery of brain cancer using brain MRI. In this study, we performed
pre-processing using the adaptive bilateral filter (ABF) for removal of the
noises that are present in an MR image. This was followed by the binary
thresholding and Fuzzy Recurrent Neural Network (FR-Net) segmentation
techniques for reliable detection of the tumor region. Training, testing, and
validation datasets are used. Based on our machine, we will predict whether
the subject has a brain tumor or not. The resultant outcomes will be
examined through various performance examined metrics that include
accuracy, sensitivity, and specificity. It is desired that the proposed work
would exhibit a more exceptional performance over its counterparts.
 2

Keywords: Brain tumor, Magnetic resonance imaging, Adaptive Bilateral

Filter, Fuzzy Recurrent neural network.

1 Introduction

The brain tumor is one all the foremost common and, therefore, the deadliest
brain diseases that have affected and ruined several lives in the world. Cancer is a
disease in the brain in which cancer cells ascends in brain tissues. Conferring to a
new study on cancer, more than one lakh people are diagnosed with brain tumors
every year around the globe. Regardless of stable efforts to overcome the
complications of brain tumors, figures show unpleasing results for tumor patients.
To contest this, scholars are working on computer vision for a better
understanding of the early stages of tumors and how to overcome using advanced
treatment options.

Magnetic resonance (MR) imaging and computed tomography (CT)

scans of the brain are the two most general tests to check the existence of a tumor
and recognize its position for progressive treatment decisions. These two scans are
still used extensively for their handiness, and the capability to yield high-
definition images of pathological tissues is more. At present, there are several
other conducts offered for tumors, which include surgery, therapies such as
radiation therapy, and chemotherapy. The decision for which treatment relies on
the many factors such as size, kind, and grade of the tumor present in the MR
image. It’s conjointly chargeable for whether or not cancer has reached the other
portions of the body.

Precise sighting of the kind of brain abnormality is enormously needed

for treatment operations with a resolution to diminish diagnostic errors. The
precision is often makeshift utilizing computer-aided diagnosis (CAD) systems.
The essential plan of computer vision is to produce a reliable output, which is an
associate estimation to assist medical doctors in image understanding and to lessen
image reading time. These advancements increase the steadiness and correctness
of medical diagnosis — however, segmenting an MR image of the tumor and its
area itself a very problematic job. The occurrence of tumors in specific positions
within the brain image without distinguishing picture intensities is an additional
issue that makes a computerized detection of brain tumor and segmentation a
problematic job.

The MRI is the most regularly utilized strategy for imaging brain tumors
and the identification of its vicinity. The conventional technique for CT and MR
image classification and detection of tumor cells remains largely supported for the
human reviewing apart from different other methods. MR images are mainly used
because there are non-destructive and non-ionizing. MR imaging offers high-
 3

definition pictures that are extensively utilized in discovering brain tumors. MRI
has diverse schemes such as flair, T1-weighted, T2-weighted images. There are
many image processing techniques such as pre-processing, segmentation of
images, image improvements, feature extraction, and classifiers.

Sathya et al. (2011) [3], provided a different clustering algorithm such as K-

means, Improvised K-means, C-means, and improvised C-means algorithms.
Their paper presented an experimental analysis for massive datasets consisting of
unique photographs. They analyzed the discovered consequences using numerous
parametric tests.

Roy et al. (2012) [18] calculated the tumor affected area for proportioned
analysis. They confirmed its software with numerous statistics groups with
distinctive tumor sizes, intensities, and location. They showed that their algorithm
could robotically hit upon and phase the brain tumor from the given photo. Image
pre-processing consists of fleeting that pictures to the filtering technique to
remove distractors found in given pictures.

Jaskirat Kaur et al. (2012) [6] defined a few clustering procedures for the
segmentation process and executed an assessment on distinctive styles for those
techniques. Kaur represented a scheme to measure selected clustering techniques
based on their steadiness in exceptional tenders. They also defined the diverse
performance metric tests, such as sensitivity, specificity, and accuracy.

A. Sivaramakrishnan et al. (2013) [1] projected an efficient and innovative

discovery of the brain tumor vicinity from an image that turned into finished using
the Fuzzy C-approach grouping algorithm and histogram equalization. The
disintegration of images is achieved by the usage of principal factor evaluation is
done to reduce the extent of the wavelet coefficient. The outcomes of the
anticipated FCM clustering algorithm accurately withdrawn tumor area from the
MR images.

K. Sudharani et al. [5] presented a K- nearest neighbour algorithm to the MR

images to identify and confine the hysterically full-fledged part within the
abnormal tissues. The proposed work is a sluggish methodology but produces
exquisite effects. The accuracy relies upon the sample training phase.

L. Marroquin et al. [10] presented the automated 3d segmentation for brain

MRI scans. Using a separate parametric model in preference to a single
multiplicative magnificence will lessen the impact on the intensities of a grandeur.
Brain atlas is hired to find non-rigid conversion to map the usual brain. This
conversion is later cast-off segment the tumor region.

J.T. Kwok et al. [7] delivered wavelet-based photograph fusion to easily

cognizance at the object with all focal lengths as several vision-related processing
tasks can be carried out more effortlessly when wholly substances within the
 4

images are bright. In their work Kwok et al. investigated with different datasets,
and results show that presented work is extra correct as it does not get suffering
from evenness at different activity stages computations.

Mukambika et al. [15] proposed methodology for the subsequent stage’s

classification of the tumor, whether it is present or not. Their proposed work
represents the comparative study of strategies used for tumor identification from
MR images, namely the Level set approach and discrete wavelength transforms
(DWT) and K-method segmentation algorithms. After that phase, feature
extraction is done followed SVM classification.

Astina minz et al. [11] implemented an operative automatic classification

approach for brain image that projected the usage of the AdaBoost gadget
mastering algorithm. The proposed system includes three main segments. Pre-
processing has eradicated noises in the datasets and converted images into
grayscale. Median filtering and thresholding segmentation are implemented in the
pre-processed image.

Kumar and Mehta [8] proposed the texture-based technique in this paper. They
highlighted the effects of segmentation if the tumor tissue edges aren't shrill. The
performance of the proposed technology may get unwilling results due to those
edges. The texture evaluation and seeded region approach turned into executed
inside the MATLAB environment.

S. Pereira et al. [17] presented that magnetic resonance prevents physical

segmentation time in the medical areas. So, an automatic and reliable
segmentation technique for identifying abnormal tissues by using Convolutional
Neural Network (CNN) had been proposed in the research work. The massive
three-dimensional and underlying roughness amongst brain images makes the
process of segmenting the image a severe issue, so a robust methodology such as
CNN is used.

B. Devkota et al. [4] have proposed that a computer-aided detection (CAD)

approach is used to spot abnormal tissues via Morphological operations. Amongst
all different segmentation approaches existing, the morphological opening and
closing operations are preferred since it takes less processing time with the utmost
efficiency in withdrawing tumor areas with the least faults.

K. Vinotha et al. [23] proposed brain tumor detection and the usage of the
Histogram Equalization (HE) and the Fuzzy Support Vector Machine (FSVM)
classification techniques. The brain MR image is pre-processed with histogram
equalization and segmented the apprehensive components from the photo
primarily based on the MRF algorithm for segmentation technique. MRF approach
expanded the tumor segmentation accuracy through which the overall
performance of the proposed approach changed into advanced.
 5

Dalia Mahmoud et al. [9] presented a model using Artificial Neural Networks
for tumor detection in brain images. They implemented a computerized
recognition system for MR imaging the use of Artificial Neural Networks. That
was observed that after the Elman community was used during the recognition
system, the period time and the accuracy level were high, in comparison with
other ANNs systems. This neural community has a sigmoid characteristic which
elevated the extent of accuracy of the tumor segmentation.

M. M. Sufyan et al. [2] has presented a detection using enhanced edge

technique for brain-tumor segmentation that mainly relied on Sobel feature
detection. Their presented work associates the binary thresholding operation with
the Sobel approach and excavates diverse extents using a secure contour process.
After the completion of that process, cancer cells are extracted from the obtained
picture using intensity values.

Kumar and Varuna Shree [22] proposed work for the detection tumor region
using discrete wavelength transforms (DWT). This work consists of three phases,
namely an image enhancement using filtering technique, gray-level co-incidence
matrix (GLCM) feature extraction of tumor in addition to DWT based tumor
location developing segmentation. It is used to improve overall performance and
reduce complexity. The denoised accompanied by the aid of morphological
filtering operations which put off the noises that can be even shaped subsequent
segmentation technique. The PNN classifier is to use for classifying the
abnormality, which is trained by different datasets, and the accuracy is measured
within the detection of tumor region of mind MR images.

T.U Paul and S.K. Bandyopadhyay [20] has presented the brain segmentation
that has automated the use of the Dual Localization technique. In the initial phase,
the skull masks are generated for the brain MR images. The tumor areas are
improvised using the K-manner procedure. In the final step of their proposed
work, they evaluated by its dimensions such as length and breadth.

Vaishali [21] proposed a method that includes step by step procedure starting
with image pre-processing followed by extraction of useful objects and finally
classification of tumor region. Pre-processing is completed to enhance the image
using eliminating the noise via making use of Gaussian filters from the authentic
ones. The next step is feature extraction, in which a magnified image is used to
extract the feature using a symlet wavelet technique. The very last step is the
classification of tumors by the use of a Support vector machine (SVM).

Yuehao Pan et al., [16] has used brain MRI pix for getting useful statistics for
classifying brain tumor. In their proposed method, they used Convolutional Neural
Networks (CNN) algorithms for developing a brain tumor detection system. The
performance of their CNN report is measured primarily based on sensitivity and
specificity parameters, which have stepped forward when in comparison to the
Artificial Neural Networks (ANN).
 6

Monica Subashini and Sarat Kumar Sahoo [12] has suggested a technique for
detecting the tumor commencing the brain MR images. They also worked on
different techniques, which include pulse-coupled Neural Network and noise
removal strategies for reinforcing the mind MRI images and backpropagation
network for classifying the brain MRI images from tumor cells. They observed
image enhancement and segmentation of the usage of their proposed technique,
and the backpropagation network helps in the identification of a tumor in a brain
MR image.

A. Sankari and S. Vigneshwari [19] has proposed a Convolutional Neural

Network (CNN) segmentation, which principally based on the brain tumor
classification method. The proposed work used the non-linearity activation feature
that's a leaky rectified linear unit (LReLU). They primarily focused on necessary
capabilities, which include mean and entropy of the image and analyzed that the
CNN algorithm is working higher for representing the complicated and minute
capabilities of brain tumor tissues present in the MR Images.

2 Objective of Paper

A tumor is a mass of tissues that grow on body parts out of control of

regular forces, which regulates the growth. Tumors can, directly and indirectly,
damage all the healthy brain cells. Tumors can be of dissimilar sizes, positions,
and locations. They can also overlap their intensities over healthy tissues. Tumors
may be primary and may not be primary too. The most common primary tumors
related to the brain are Gliomas and Malignant. Identification of brain tumors in
MRI scans during earlier stages are essential in medical diagnosis. So, the core
objective of our project is to develop an automated segmentation of brain MR
image for a tumor detection system that can detect a tumor from magnetic
resonance brain images. In our paper, we proposed a tumor detection system that
presented good accuracy and less computational time when compared to its
contour parts.

The proposed system begins with reading the MRI image from the
dataset, and then image pre-processing is done by using filtering techniques such
as adaptive bilateral for removal of noise pixels present in the original brain tumor
image. After applying filters, the image gets smoothened, and edges cannot be
detected. To overcome this, we use an edge detection technique. So, the brain
MRI image enhancement is done using the Sobel filter. The brain MRI
segmentation is done with the support of Fuzzy Recurrent Neural Networks based
on the concept of the feed-forward framework. After the segmentation stage,
various features are extracted using shape and texture primarily based on the brain
 7

MR image. After feature extraction, brain tumor MRI image classification is done
using Convolutional Neural Network, and classification accuracy is measured.

3 Preprocessing Using Adaptive Bilateral Filter And Image

Enhancement

The Brain MRI image dataset has been downloaded from the Kaggle.
The MRI dataset consists of around 1900 MRI images, including normal, benign,
and malignant. These MRI images are taken as input to the primary step. The pre-
processing is an essential and initial step in improving the quality of the brain MRI
Image. The critical steps in pre-processing are the reduction of impulsive noises
and image resizing. In the initial phase, we convert the brain MRI image into its
corresponding gray-scale image. The removal of unwanted noise is done using the
adaptive bilateral filtering technique to remove the distorted noises that are present
in the brain picture. This improves the diagnosis and also increase the
classification accuracy rate.

Bilateral filtering smooth images while conserving edges utilizing a

nonlinear grouping of neighbouring image pixels. This filtering technique is
simple, local, and concise. It syndicates a grey level grounded on their likeness
and the symmetrical nearness and chooses near vales to farther values in both
range and domain. So, we proposed a new smoothing and sharpening algorithm,
Adaptive Bilateral Filter.

MR image segmentation with the aid of preserving the threshold

information, which is convenient to identify the broken regions extra precisely. It
was a trendy surmise that the objects that are placed in close propinquity might be
sharing similar houses and characteristics. Here for noise removal, the intensity is
taken into consideration as the top characteristic. So, when the pixel is captivated
via the noise, then such pixel will showcase an indigent correlation with the
neighboring pixels in such a case that precise pixel is smoothened by using
thinking about all eight neighboring pixels.

The inceptive threshold value is evaluated and is designated by a variable

Th between two pixels Im and In. The pixel Im is assigned to the respective
segments most effective if the pixel Im is keeping a darkish insensitive than the
evaluated threshold. When the intensity value of In is brighter than the pixel is
allocated to the adjacent segment, which is calculated by means of the objective
function.
 8

Fig. 1. Represents the architecture diagram of FR-Net

The initial threshold value is approximated by considering the intensities

Px and Py of pixels Im and In by using the following equation.

𝑃𝑥 +𝑃𝑦
Th= 𝑊 × ∑255
𝑖=0 2

Where W is the Laplacian second-order differential equation that spread

pixel over the segments of the MR image by isotropic derivative. The value of W
is derived as

ꝺ2 𝑤 ꝺ2 𝑤
W = ꝺI + ꝺI𝑛
𝑚

After the completion of the pre-processing, the image will be free from
the noises, but we still need to enhance the image since the obtained image is
smoothened, edges may not be preserved, and the image will be dull. To overcome
all these, we use edge detection called Sobel filtering technique. The whole thing
is done by calculating the gradient of image intensities at each pixel within the
image. It is widely used in image analysis to help locate edges in images. It
combines Gaussian smoothing and differentiation. It will also enhance the darker
areas of the image, slightly increase contrast and as sharp as possible.
 9

4 Fuzzy Recurrent Neural Network (FR-Net) based Image

Segmentation

In this paper, we have proposed a different and robust automated

segmentation mechanism for the segmentation of the brain MR image through
Recurrent Neural Network based on the concept of the feed-forward framework
that works based on the numerous image pyramids, that works almost like a Level
set based deformable Model. In the proposed approach at each phase, the pace of
the curves is being assessed with q quite lesser network features that would
consume minimal training set when compared to that of the conventional
Convolutional neural network-based approach which needs a rigorous training of
the data for better optimal results, and the execution time of the traditional
approach is more when compared to that of the proposed FR-Net approach.

The essences of the Recurrent Neural Network have further improved

through the fuzzifying the resultant outcome that would yield an optimal
segmentation of the MR image when compared to that of the traditional MR
segmentation approaches. The proposed approach could be better understood by
the formulations stated below. Let is assumed that the input image sequences that
are denoted by 𝐼𝑠 = {𝐼𝑠 |𝑡𝑓 = 1,2,3 … , 𝑚𝑎𝑥𝑡 } that are concerning the
predetermined time frames 𝑡𝑓 and the resultant prognosticated label sequence is
represented through the variable 𝐿 = {𝐿𝑠 |𝑡𝑓 = 1,2,3 … 𝑚𝑎𝑥𝑡 }. The proposed
algorithm would work concerning the Fuzzy Membership evaluation and the RNN
through Fully Convolutional Neural networks.

Recurrent Neural Networks is an approach that is being customized to

accept the outcomes of the previous inputs and outputs by storing the patterns in
the network. The fully convolutional sublayer of the over the input image I and the
key features that are assumed as 𝑘𝑓 .The U-net structure is being used in the
process of identification of the elements for the automated segmentation of the
MR image. The pixel-wise feature map 𝑓𝑚 at every given point is being assumed
through various units, Convolutional Long Short-Term Memory(C-LSTM) that is
being capable of analyzing the composed as following

𝑥𝑠 = 𝜎(𝑖𝑠 ∗ 𝑝𝑚𝑖 + 𝑧𝑠−1 ∗ 𝑝𝑧𝑖 + 𝑏𝑥 )

𝑦𝑠 = 𝜎(𝑖𝑠 ∗ 𝑝𝑚𝑦 + 𝑧𝑠−1 ∗ 𝑝𝑧𝑦 + 𝑏𝑦 )
𝑢𝑠 = 𝑢𝑠−1 ⨀𝑦𝑠 + 𝑥𝑠 ⨀𝑡𝑎𝑛 𝑧(𝑖𝑠 ∗ 𝑝𝑚𝑐 + 𝑧𝑠−1 ∗ 𝑝𝑧𝑐 + 𝑏𝑢 )
𝑣𝑠 = 𝜎(𝑖𝑠 ∗ 𝑝𝑚𝑜 + 𝑧𝑠−1 ∗ 𝑝𝑧0 + 𝑏0 )
𝑝𝑠 = 𝑣𝑠 ⊙ tan 𝑧 (𝑢𝑠 )

In the above statements, the symbol * represents convolution and

represents the variable wise multiplication, and the symbol represents the sigmoid
function, and the variables are the list of input where the variable (i) denotes the
input gate, (f) represents the forget gate, (c) signifies memory cell and (o)
 10

represents the output gate, and the variable p designates the convolutional kernel
and b designates the bias at each gate. And the variable represents the input vector
map and designate the output vector space. The output variable concerning the
time stamp that demonstrates both the input at the current moment and the output
at the previous state denoted by and that are being used for the process of
prediction. However, from figure 1, we can better understand that the proposed
approach would be using both the foreground and background information for the
operation of prediction.

5 Experimental Results and Discussion

The proposed approach has been practically implemented through the

real-time dataset that is available in the Brain Web, the images of sizes 256 x 256
and 512 x 512 has been used for practical implementation of the algorithm, And
the proposed algorithm has been assessed through various performance evaluation
metrics that include True Positive, True Negative the former one that designates
how many times does the proposed algorithm is able to correctly recognize the
damaged region as damaged region and the later one designates how many times
does the proposed algorithm correctly identified non-damaged region as non-
damaged region. And the False Positive (FN) and False Negative (FN) the former
one designates how many times does the proposed algorithm fails to recognize the
damaged region correctly, and the later represents how many times does the
proposed algorithm fails to identify the non-tumors region as non-tumors regions.
Basing on values of TP, TN, FP, and FN, the values of Sensitivity and Specificity
of the proposed algorithm.

On experimentation, it was observed that the proposed methodology

seems to be outperformed when compared to that of Convolutional Neural
Network (CNN), Heuristic approach for real-time segmentation (HARIS)
proposed by P Naga Srinivasu et al. [14] and Genetic Algorithm with Teacher
Learner Based Optimization (TLBO) as stated by P Naga Srinivasu et al. [13].
Among all the aforementioned approaches, the proposed Fuzzy Recurrent Neural
Network (FR-Net) based approach seems too much better in terms of quality of
the output and optimized computational time when compared to its counterparts
presented in table 3.

TABLE 1 Represents the performance evaluation of an adaptive bilateral filter for a

256*256 size image

Noise Variance PSNR MSE RMSE IQI

9 33.3 29.0 5.3 0.562

 11

8 35.1 26.2 5.1 0.599

7 36.4 23.1 4.8 0.659

5 38.9 20.4 4.5 0.701

4 40.6 17.0 4.2 0.742

3 41.9 14.8 3.9 0.785

1 44.2 11.9 3.5 0.863

TABLE 2 Represents the performance evaluation of an adaptive bilateral filter for 512*512
size image.

Noise Variance PSNR MSE RMSE IQI

9 35.4 27.1 5.2 0.411

8 37.7 23.4 4.8 0.472

7 40.1 21.2 4.5 0.501

5 45.7 16.9 4.1 0.572

4 48.2 14.1 3.7 0.606

3 49.3 12.2 3.4 0.676

1 50.6 10.5 3.2 0.704

It is observed from the above table 1 and table 2 upon performing the pre-
processing for the noise removal using the proposed filtering technique so the
quality of image has been improved that would be having a substantial influence
on the eminence of the outcome, we have used Adaptive Bilateral filter for
restoring the image by preserving the edge related information, and the outcome
of the proposed approach seems to be more pleasing with better IQI and PSNR
value. Moreover, the proposed noise removal technique also displayed minimal
values for both MSE and RMSE. It is observed on the practical implementation
that the proposed algorithm has exhibited better performance over the smaller
image when compared to the smaller size image. Moreover, the computational
time is considerably high for larger size MR images.
 12

Fig. 2. Represents the performance of Adaptive Bilateral Filter in image de-noising.

TABLE 3 Represents the Accuracy, Sensitivity, and Specificity of the proposed

approach in comparison with its counterparts.

Segment
True True False False Computat
ation Accura Sensitiv Specific
Positive Negativ Positive Negativ ional
Techniq cy (%) ity (%) ity (%)
(%) e (%) (%) e (%) Time(s)
ues

GA with
71.9 76.4 28.1 23.6 74.2 75.2 73.1 1.7
TLBO

CNN 78.3 82.1 21.7 17.9 80.2 81.4 79.3 2.7

HARIS 83.7 84.5 16.3 15.5 84.1 84.4 83.8 2.6

FR-NET 86.4 89.2 13.6 10.8 87.8 88.8 86.7 2.3

 13

Fig. 3. Represents the performance of proposed Fuzzy RNN

TABLE 4 Represents the precise location and the approximated evaluation of tumor

Tumor Whole Enhanced

Core (TC) Tumor Tumor
(WT) (ET)
T1-Weighted MR Image
9.236 3.468 8.312
of size 256 X 256 (Without
  
Adaptive Bilateral Filter)
0.981 0.561 0.769
T1-Weighted MR Image
8.436 4.642 7.924
of size 256 X 256 (With
  
Adaptive Bilateral Filter)
0.689 0.881 0.489
T1-Weighted MR Image
9.102 4.432 7.282
of size 512 X 512 (Without
  
Adaptive Bilateral Filter)
0.819 0.598 0.533
T1-Weighted MR Image
8.272 5.121 6.963
of size 512 X 512 (With
  
Adaptive Bilateral Filter)
0.622 0.932 0.416
 14

It is observed in the practical implementation of the proposed approach the

tumor region is being correctly identified, and the proposed method can recognize
the growth of the region of the tumor with almost a negligible difference with the
actual ground facts. It has been observed for both the 256 x 256 and 512 x 512
size MR images, and it seems that the proposed approach has outperformed for a
smaller size image over the larger size one, In table 4, the first column represents
the actual core that is at the center of the tumor and the whole tumor that also
considers the partially effected tissues based in the texture value and the last
column represents the enhances tumor from the previous examination.
 15

Fig. 4. Experimental results of Fuzzy RNN for MR image of size 256 x 256.
 16

Fig. 5. Experimental results of Fuzzy RNN for MR image of size 512 x 512

The above resultant outcomes presented in figure 4 and figure 5 the image from
left to right represents in a step by step process of the proposed algorithm. The
first image represents the original image, which is read from the dataset. The
second image is the denoised image, which is carried out by applying the adaptive
bilateral filter. The third image is the enhanced image obtained after binary
thresholding. The fourth image is an image that is by means of after the Fuzzy
Recurrent Neural Network (FR-Net). The final image represents the tumor
location from the given segmented image.

It is observed from table 3 upon performing various segmentation

techniques that have the ability to recognize the isolated region from the MR
images that are used to analyze the shape and size of the denoised image. We have
used Fuzzy Recurrent Neural Networks for segmentation, and the output of our
proposed work is pleased with better accuracy, sensitivity, and computational
time.

6 Conclusion

We proposed a computerized method for the segmentation and

identification of a brain tumor using the Fuzzy Recurrent Neural Network. The
input MR images are read from the local device using the file path and converted
into grayscale images. These images are pre-processed using an adaptive bilateral
filtering technique for the elimination of noises that are present inside the original
image. The binary thresholding is applied to the denoised image, and Fuzzy
Recurrent Neural Network segmentation is applied, which helps in figuring out the
tumor region in the MR images. The proposed model had obtained an accuracy of
75% and yields promising results without any errors and much less computational
time.
 17

7 Future Scope

It is observed on extermination that the proposed approach needs a vast

training set for better accurate results; in the field of medical image processing, the
gathering of medical data is a tedious job, and, in few cases, the datasets might not
be available. In all such cases, the proposed algorithm must be robust enough for
accurate recognition of tumor regions from MR Images. The proposed approach
can be further improvised through in cooperating weakly trained algorithms that
can identify the abnormalities with a minimum training data and also self-learning
algorithms would aid in enhancing the accuracy of the algorithm and reduce the
computational time.

8 References

[1] A.Sivaramakrishnan And Dr.M.Karnan “A Novel Based Approach For

Extraction Of Brain Tumor In Mri Images Using Soft Computing Techniques,”
International Journal Of Advanced Research In Computer And Communication
Engineering, Vol. 2, Issue 4, April 2013.

[2] Asra Aslam, Ekram Khan, M.M. Sufyan Beg, Improved Edge Detection
Algorithm for Brain Tumor Segmentation, Procedia Computer Science, Volume
58,2015, Pp 430-437, ISSN 1877-0509.

[3] B.Sathya and R.Manavalan, Image Segmentation by Clustering Methods:

Performance Analysis, International Journal of Computer Applications (0975 –
8887) Volume 29– No.11, September 2011.

[4] Devkota, B. & Alsadoon, Abeer & Prasad, P.W.C. & Singh, A.K. & Elchouemi,
A.. (2018). Image Segmentation for Early Stage Brain Tumor Detection using
Mathematical Morphological Reconstruction. Procedia Computer Science. 125.
115-123. 10.1016/j.procs.2017.12.017.

[5] K. Sudharani, T. C. Sarma and K. Satya Rasad, "Intelligent Brain Tumor

lesion classification and identification from MRI images using k-NN technique,"
2015 International Conference on Control, Instrumentation, Communication and
Computational Technologies (ICCICCT), Kumaracoil, 2015, pp. 777-780. DOI:
10.1109/ICCICCT.2015.7475384

[6] Kaur, Jaskirat & Agrawal, Sunil & Renu, Vig. (2012). A Comparative Analysis
of Thresholding and Edge Detection Segmentation Techniques. International
Journal of Computer Applications.vol. 39.pp. 29-34. 10.5120/4898-7432.
 18

[7] Li, Shutao, JT-Y. Kwok, IW-H. Tsang and Yaonan Wang. "Fusing images with
different focuses using support vector machines." IEEE Transactions on neural
networks 15, no. 6 (2004): 1555-1561.

[8] M. Kumar and K. K. Mehta, "A Texture based Tumor detection and
automatic Segmentation using Seeded Region Growing Method," International
Journal of Computer Technology and Applications, ISSN: 2229-6093, Vol. 2, Issue
4, PP. 855-859 August 2011.

[9] Mahmoud, Dalia & Mohamed, Eltaher. (2012). Brain Tumor Detection
Using Artificial Neural Networks. Journal of Science and Technology. 13. 31-39.

[10] Marroquin J.L., Vemuri B.C., Botello S., Calderon F. (2002) An Accurate
and Efficient Bayesian Method for Automatic Segmentation of Brain MRI. In:
Heyden A., Sparr G., Nielsen M., Johansen P. (eds) Computer Vision — ECCV
2002. ECCV 2002. Lecture Notes in Computer Science, vol 2353. Springer, Berlin,
Heidelberg.

[11] Minz, Astina, and Chandrakant Mahobiya. “MR Image Classification Using
Adaboost for Brain Tumor Type.” 2017 IEEE 7th International Advance Computing
Conference (IACC) (2017): 701-705.

[12] Monica Subashini.M, Sarat Kumar Sahoo, “Brain MR Image Segmentation

for TumorDetection using Artificial Neural Networks,” International Journal of
Engineering and Technology (IJET), Vol.5, No 2, Apr-May 2013.

[13] P. Naga Srinivasu, G. Srinivas, T Srinivas Rao, (2016). ‘An Automated Brain
MRI image segmentation using a Generic Algorithm and TLBO.’ International
Journal of Control Theory and Applications, Vol: 9(32).

[14] P. Naga Srinivasu, T. Srinivasa Rao, Valentina Emilia Balas. (2020). A

systematic approach for identification of tumor regions in the human brain
through HARIS algorithm, Deep Learning Techniques for Biomedical and Health
Informatics, Academic Press.Pages 97-118. https://doi.org/10.1016/B978-0-12-
819061-6.00004-5.

[15] P.S. Mukambika, K Uma Rani, “Segmentation and Classification of MRI

Brain Tumor,” International Research Journal of Engineering and Technology
(IRJET), Vol.4, Issue 7, 2017, pp. 683 – 688, ISSN: 2395-0056

[16] Pan, Yuehao & Huang, Weimin & Lin, Zhiping & Zhu, Wanzheng & Zhou,
Jiayin & Wong, Jocelyn & Ding, Zhongxiang. (2015). Brain tumor grading based on
Neural Networks and Convolutional Neural Networks. Conference proceedings:
 19

Annual International Conference of the IEEE Engineering in Medicine and Biology

Society. IEEE Engineering in Medicine and Biology Society. Conference. 2015.
699-702. 10.1109/EMBC.2015.7318458.

[17] S. Pereira, A. Pinto, V. Alves, and C. A. Silva, "Brain Tumor Segmentation

Using Convolutional Neural Networks in MRI Images," in IEEE Transactions on
Medical Imaging, vol. 35, no. 5, pp. 1240-1251, May 2016.

[18] S. Roy And S.K.Bandyopadhyay, “Detection And Qualification Of Brain

Tumor From MRI Of Brain And Symmetric Analysis,” International Journal Of
Information And Communication Technology Research, Volume 2 No.6, June
2012, Pp584-588

[19] Sankari, Ali, and S. Vigneshwari. “Automatic tumor segmentation using

convolutional neural networks.” 2017 Third International Conference on Science
Technology Engineering & Management (ICONSTEM) (2017): 268-272.

[20] T.U Paul and S.K. Bandyopadhyay, ―Segmentation of Brain Tumor from
Brain MRI Images Reintroducing K – Means with advanced Dual Localization
MethodTuhin,‖ International Journal of Engineering Research and Applications,
Volume 3, Issue 1, June 2012, ISSN 2278-0882.

[21] Vaishali et al. (2015) Wavelet-based feature extraction for brain tumor
diagnosis—a survey. Int J Res Appl Sci Eng Technol (IJRASET) 3(V), ISSN: 2321-
9653

[22] Varuna Shree, N., Kumar, T.N.R. Identification and classification of brain
tumor MRI images with feature extraction using DWT and probabilistic neural
network. Brain Inf. 5, 23–30 (2018) doi:10.1007/s40708-017-0075-5

[23] Vinotha, K., 2014. “Brain Tumor Detection and Classification Using
Histogram Equalization and Fuzzy Support Vector Machine Approach,”
International Journal of Engineering and Computer Science ISSN2319-7242 3(5):
5823-5827.