Content uploaded by Satyabrata Aich
Author content
All content in this area was uploaded by Satyabrata Aich on Jun 09, 2020
Content may be subject to copyright.
May – June 2020
ISSN: 0193-4120 Page No. 16580 – 16586
16580
Published by: The Mattingley Publishing Co., Inc.
A CNN based Approach for the Detection of Brain
Tumor Using MRI Scans
Sobhangi Sarkar1, Avinash Kumar2, Sabyasachi Chakraborty3, Satyabrata Aich4, Jong-Seong Sim5,
Hee-Cheol Kim6
1,2School of Computer Engineering, KIIT University, Bhubaneswar, India
3Department of Computer Engineering, Inje University, Gimhae, South Korea
4,5Institute of Digital Anti-Aging Healthcare, Inje University, Gimhae, South Korea
6u-HARC, Inje University, Gimhae, South Korea
Article Info
Volume 83
Page Number: 16580 – 16586
Publication Issue:
May - June 2020
Article History
Article Received: 1May 2020
Revised: 11 May 2020
Accepted: 20 May 2020
Publication: 24May 2020
Abstract
In the present day, the rise in the number of diagnosis of brain tumour has reached an
enormous height. Gliomas and metastatic brain tumours are found to represent 30% of all
brain tumours that are diagnosed in human beings. Now with such an enormous number, it
is very important that a computer-aided detection system must be employed to diagnose the
brain tumour cases accurately and efficiently. Moreover, Magnetic resonance imaging
(MRI) has established itself to be one of the most effective tools for clinical diagnosis and
more specifically one of the most desired imaging modalities when it comes to the cohort of
the complete neuroimaging ecosystem. The proposed work leverages MRI scans (axial
slices) to detect the type of brain tumour. The dataset used in the study contained the data of
three most commonly diagnosed brain tumours namely, glioma, meningioma and pituitary
tumours. For the classification purpose a 2D Convolutional Neural Network (CNN) was
designed which propelled an overall accuracy of 91.3% and a recall of 88%, 81% and 99%
for the detection of meningioma, glioma and pituitary tumour respectively.
Keywords; brain tumor, cnn, imaging, mri, deep learning.
I. INTRODUCTION
Brain is considered as one of the most important and
cumbersome structure of the human body. It is
primarily the control center of the central nervous
system and is responsible for performing the daily
voluntary and involuntary activities in the human
body. The tumour is a fibrous mesh of unwanted
tissue growth inside our brain that proliferates in an
unconstrained way. It perpetrates interruption of the
normal function of brain cells and causes lethal
issues for the people suffering from it and can lead
to death if they are not detected early and accurately.
Tumours are primarily classified on two premises:
malignant or benign and the place of origin. The
benign form can be easily discriminated and have a
slow maturation rate and has well-defined borders.
Cancerous tumours are known as malignant [1].
These types of tumors are very pugnacious and
lethal in nature and are difficult to notice. And also,
these type of tumors affects other parts of the brain
and also spinal cord [2]. With 14,000 deaths every
year, Malignant tumors are known to be the most
dangerous tumors [3] For the detection of tumor the
doctors or physicians typically leverage Computed
Tomography (CT) or Magnetic Resonance Imaging
(MRI). MRI is one of the scanning technology that
provides high contrast scans and furnishes detailed
dynamics about the brain anatomy and the
aneurysms in the brain tissues [3]. The radiologist
consider MRI scans to be the most effective process
of scanning the brain tissues to fetch spatial
information about the brain and the specific tumor.
Our model presents the design of the automated
scheme that is designed to differentiate between
normal and abnormal MRI images and classify
May – June 2020
ISSN: 0193-4120 Page No. 16580 – 16586
16581
Published by: The Mattingley Publishing Co., Inc.
tumour whether they are meningioma tumour,
gliomatumour or pituitary tumour.
The paper is organized as follows. The recent work
and development that had been done are described
under section
In section III the methods and the methodology that
is involved in model development are discussed. In
section IV we have described our whole paper. And
lastly, the paper is concluded in section VI.
II. LITERATURE SURVEY
Many researchers have discussed the importance of
image processing in bio-medical in multiple ways
which include the processing of images of X-rays,
CT scans and MRI to detect the malformation and
irregularities in the human body. It has been
medically proven that MRI imaging is less harmful
than other imaging techniques used because it
prevents the body from the exposure of harmful
radiations but before analyzing any image, we need
to perform a complicated task of pre- processing the
images. Image pre-processing is a gradual process
where one step leads to another step. It involves step
like noise reduction, image enhancement, image
contrasting and when it comes down to medical
domain involving detection of brain tumour from an
MRI scan, removing the imprints of the skull from
the image becomes a prime procedure [4]. The next
step that would follow is converting the image into
grey scaled image, where the pixels of the image
only shows intensity without color but even after
this step the noise in the image is still present which
needs to be eradicated before further processing.
The noise in the image is removed by using filtering
techniques.
Filtering: It is a technique used to enhance a picture
by highlight some features while eliminating other
features that do not promote any information gain in
a particular study. It involves steps like noise
reduction, smoothing, sharpening and edge
reduction. The most frequently used filter is median
filter which is used for impulsive noise and speckle
noise and this technique has an edge over other
techniques because of it’sability to preserve the
edges of the image without cropping the signal. In
[2] and [3] researchers have used median filter to
remove the noise that are by default present in an
MRI but in some cases [5][6] Gaussian filters based
on the principle of convolution is also used to
minimze the noise in the image and blur the image
by blurring the edges and reducing the contrast. The
advantage of using a Gaussian filter is that it works
faster than other filters.
Segmentation: After successful removal of noise
from an image, the next process that follows is
segmentation in which an image is broken down into
pixels. This helps in easy analysis of the image to
deduce a meaningful observation from it. After the
grouping of pixels, each segment of pixel shares
some common features. The paper by Amina et al.
[5] uses K-means clustering for segmentation. In
this method images is segmented to multiple clusters
on basis of the nearest mean. Veer et al. [7] in her
research used thresholding for segmentation. In
thresholding if the value of intensity of a pixel is
significant than some predefined constant then it
turns the pixel to black and if the value of intensity
of pixel is less than the predefined value then it
makes the pixel white. An extension of thresholding
i.e. Otsu Binarization is used in the paper by Shil et
al. [8] in which a binary image is re-created from a
grayscale image by dividing into two classes namely
foreground and background. In the research by Rathi
et al. [10] she has used fuzzy C-means as the
clustering algorithm. In fuzzy C-means, data point is
allocated to cluster center on the basis of interspace
between the center of cluster and the data point
[11].Membership towards particular cluster
increases as closeness between data and cluster
center increases. The edge of using this algorithm is
that it gives accurate result for data set which are
overlapped. Another type of segmentation process
named watershed segmentation is used by [7] in
which different objects in an image is separated. The
algorithm which assumes pixels values as a local
May – June 2020
ISSN: 0193-4120 Page No. 16580 – 16586
16582
Published by: The Mattingley Publishing Co., Inc.
topography (elevation) is watershed algorithm. The
algorithm floods basins from the markers, until
basins attributed to different markers meet on
watershed lines. In most of the cases, we chose
markers as local minima of the image, from which
basins are flooded.
Post Segmentation: Till segmentation the motto of
the process remains same, but the further process
diversifies into different kind of classification i.e.
classifying the image as a normal or an abnormal
MRI and then the succeeding research progressed to
detect the size and location of abnormalities in MRI
(here, tumour). Then researchers added a fresh step
to the existing work by classifying the type of
tumour into two types namely Malignant (cancerous
tumour) and Benign (noncancerous tumour). Once
the segmentation phase is successfully carried out
many optimization techniques are applied to
improve the result obtained.
The first step towards this field of research is
detecting any kind of anomaly in a brain, [8] used
SVM classification for classifying the image into
two types normal MRI and abnormal MRI. [9]
detected the of tumour, skull, gray matter and white
matter using morphological operations which
compares the pixel value of input and output image
to get the size of the required part and using manual
segmentation she has calculated the area of the
tumour which gives the size of the tumour despite
the presence of other components of brain. For
classification, [5] used the algorithm of SVM family
which includes Linear, Cubic and Gaussian kernel
functions. This algorithm follows the principle of
drawing a hyper plane
by maximizing the margin between the classes by
using support. Three types of kernel functions are
used to improve the accuracy of result and using
SVM classifier he has identified the affected area as
well as training his model to predict the grade of the
tumour. [7] extended the further classification of
tumour into two types –primary tumour and
secondary tumour. Primary tumours are those
tumours which originate in brain. They are divided
into two subtypes: Malignant and Benign whereas
secondary tumours are those tumours which
originate in another part of the body and spreads to
brain eventually. For classification of primary
tumour she used Artificial Neural Network (ANN).
ANN is composed of many nodes. Each node takes
a single input performs an operation on it and passes
it another layer of nodes and at output layer each
node has a node value. It basically learns using
feedback. The advantage of using ANN over other
algorithms is that it can handle more variation than
any traditional algorithm.
III. METHODOLOGY
The paper discusses the method for detecting
abnormalities in the brain MRI images. To be more
specific on the use case, an automated system is
being developed that scans through the brain MRI
images to identify the tissue growth in the brain that
is the tumours. In the study a convolutional neural
network architecture is devised which accepts 2D
MRI image slices and determines the type of error
present. The deep learning model designed is trained
with three different types of brain tumours namely,
glioma, meningioma and pituitary tumours.
Figure 1 described below, shown the complete
procedure for the detection and classification of
brain tumour. Moreover, in the following segments,
each and every step of the procedure is been duly
explained.
May – June 2020
ISSN: 0193-4120 Page No. 16580 – 16586
16583
Published by: The Mattingley Publishing Co., Inc.
Figure 1: Flow of control of the complete
methodology for the classification of brain
tumour.
DATA ACQUISITION
We have obtained the data from [12] for the
development of our system. The brain T1-weighed
CE-MRI dataset was gathered from Nanfang
Hospital, Guangzhou, China, and General Hospital,
Tianjing Medical University, China, for the time
period of 5 years i.e 2005 to 2010. The set of the
data contains 3064 slices from 233 patients,
containing 708 meningiomas, 1426 gliomas, and
930 pituitary tumors. The images used have an in-
plane resolution of 512×512 with pixel size
0.49×0.49 mm2. The thickness of slice is 6 mm and
the gap between slice are 1 mm [13]. Figure 2
represents the data distribution of the 3 different
types of tumours. Also figure 3 (a, b and c) plots the
2D scans of the tumours of 3 different types.
Figure 2: Data distribution of the tumours
(a)
(b)
(c)
Figure 3: (a) Axial Scan of Gliomatumour (b)
Axial Scan of Meningioma Tumour (b) Coronal
scan of Pituitary tumour
DATA PREPROCESSING
Preprocessing mainly aims at enhancing the quality
of MR films and transform it in a form which is
suitable for further processing by computer vision
system. Moreover, Preprocessing also helps in
refining the images and improve some of the
parameters such as improving the visual aspect of
MR images, improve signal-noise ratio, cropping
some of the parts which is not required from the
background, making images smoother and retaining
the edges [5]. We used adaptive contrast
May – June 2020
ISSN: 0193-4120 Page No. 16580 – 16586
16584
Published by: The Mattingley Publishing Co., Inc.
enhancement based on altered sigmoid function to
improve the parameter of MR image such as signal-
to-noise ratio which gave more clarity of raw MR
images. Moreover, a Skull stripping activity was
also performed. The phenomenon of eliminating the
all nonbrain tissues from brain images is known as
Skull stripping. We can remove extra cerebral
tissues such as skin, fat and skull in the brain images
by the help of Skull stripping. Some of the popular
technique for skull stripping is skull strippig using
image contour, skull stripping based on histogram
analysis or a threshold value and skull stripping
based on morphological operation and segmentation.
MODEL DESIGN
In recent years it has been seen how supervised
learning has created a paradigm shift to solve some
most difficult problems. Mostly, after the advent of
deep learning technologies it is now a major
technology which is used in all the fields namely,
healthcare, finance, automated driving etc. [14].
Moreover, the MRI images that are often used by
the doctors and physicians for detection of neural
disorders needs very accurate and précised
examination. However, the examination of such
MRI images increases a huge overhead on the
physicians and doctors as it needs much expertise.
Therefore, leveraging a computer aided detection
system for the purpose seems to be very much
helpful as it increases the accuracy and efficiency of
the diagnosis.
In in the study a deep learning architecture is
developed by leveraging 2 dimensional
convolutional neural networks for the classification
of the type of tumour from MRI slices of the brain.
Table 1 shown below, plots the complete model
architecture of the developed convolutional neural
network model.
Table 1: CNN Model Architecture
Layer
Output Shape
Parameters
Conv_2D
(None, 512, 512, 64)
640
Conv_2D
(None, 510, 510, 64)
36928
Max_Pooling_2D
(None, 255, 255, 64)
0
Dropout
(None, 255, 255, 64)
0
Conv_2D
(None, 255, 255, 32)
18464
Conv_2D
(None, 253, 253, 32)
9248
Max_Pooling_2D
(None, 126, 126, 32)
0
Dropout
(None, 126, 126, 32)
0
Conv_2D
(None, 126, 126,16)
4624
Conv_2D
(None, 124, 124, 16)
2320
Max_Pooling_2D
(None, 62,62,16)
0
Dropout
(None, 62, 62, 16)
0
Flatten
(None, 61504)
0
Dense
(None, 512)
31490560
Dropout
(None, 512)
0
Dense
(None, 3)
1539
The layers used in the development of the
convolutional neural network architecture is been
defined below.
MODEL OPTIMIZATION AND HYPER-
PARAMETER TUNING
The convolutional neural network model that has
been developed in the work has used RMSprop
optimizer as it gives the best result on this type of
dataset. The RMSprop optimizer is alike the
gradient descent algorithm with momentum. The
RMSprop optimizer limits the oscillations in the
upright direction. Therefore, we can increase our
learning rate and our algorithm could take
substantial steps in the horizontal direction
converging quickly. The difference between
RMSprop and gradient descent is on how the
gradients are calculated. The gradients in RMS prop
are calculated on the basis of the running average
which is shown in equation 1.
(𝒘, 𝒕) = 𝜸𝒗(𝒘, 𝒕 − 𝟏) + (𝟏 − 𝜸)(𝛁𝑸𝒊(𝒘))𝟐 (1)
The parameters of the deep learning model that is
weights are being updates using the equation 2.
𝜼
𝒘≔𝒘 − 𝜵𝑸(𝒘) (2)
√𝒗(𝒘, 𝒕)
May – June 2020
ISSN: 0193-4120 Page No. 16580 – 16586
16585
Published by: The Mattingley Publishing Co., Inc.
RMSprop Bayesian Sequential Model-Based
Optimization (SMBO) has been utilized for knowing
the criterion of the optimization algorithm [15]. We
generally use Bayesian SMBO to minimze the
objective function by developing a probability
function which is based on the earlier evaluation
outcomes of the objective function.
The set of hyperparameters that we got by using
Bayesian SMBO is Learning Rate as 0.001, Rho as
0.9 and decay as 0.1
IV. RESULTS
The 2D convolutional neural network that was
developed in the work provided us with quite good
and effective results on the predictive power for the
different types of tumour. Also, the model
developed in the work prompted an average recall
and precision of 88% and 91% respectively for all
the types of tumours. Moreover the 10-fold cross
validation was performed on the complete dataset to
check for the generalizability of the model and it
was found that the model generalized pretty well
and showed a constant tendency towards the
precision and recall over the random data folds of
training and testing. Figure 4 below shows the
confusion matrix of the best obtained model. The
few miss predictions that have occurred in the data
set are because of the inertial noise that was
generated while the patient performed some
movement while on the scan machine.
Figure 4: Confusion Matrix
Figure 5.a and b, plotted below shows the Log Loss
and the accuracy of the best performing model. It
can be seen from the performance graph that the
model did not overfit and the validation data
maintained a proper generalizability with the trained
data.
Figure 5 (a) Log Loss plot (b) Accuracy plot.
V. CONCLUSION
The paper discussed and implemented a deep
learning architecture by leveraging 2D convolutional
neural networks for the classification of the different
types of brain tumor from MR image slices. The
development of such a system plays a huge role as
such systems are very much required for the
accurate and efficient diagnosis of such diseases and
health problems which are life threatening in nature.
The model developed in the study plotted an
accuracy of 91.3% and an overall precision and
recall of 91% and 88% respectively.
VI. REFERENCES
[1] Kapoor, L. and Thakur, S., 2017, January. A
survey on brain tumor detection using image
processing techniques. In 2017 7th International
May – June 2020
ISSN: 0193-4120 Page No. 16580 – 16586
16586
Published by: The Mattingley Publishing Co., Inc.
Conference on Cloud Computing, Data Science &
Engineering-Confluence (pp. 582-585). IEEE.
[2] Sobhaninia, Z., Rezaei, S., Noroozi, A., Ahmadi,
M., Zarrabi, H., Karimi, N., Emami, A. and
Samavi, S., 2018. Brain tumor segmentation using
deep learning by type specific sorting of images.
arXiv preprint arXiv:1809.07786.
[3] Havaei, M., Davy, A., Warde-Farley, D., Biard, A.,
Courville, A., Bengio, Y., Pal, C., Jodoin, P.M. and
Larochelle, H., 2017. Brain tumor segmentation
with deep neural networks. Medical image
analysis, 35, pp.18-31.
[4] Ari, A. and Hanbay, D., 2018. Deep learning based
brain tumor classification and detection system.
Turkish Journal of Electrical Engineering &
Computer Sciences, 26(5), pp.2275- 2286.
[5] Amin, J., Sharif, M., Yasmin, M. and Fernandes,
S.L., 2017. A distinctive approach in brain tumor
detection and classification using MRI. Pattern
Recognition Letters.
[6] Akkus, Z., Galimzianova, A., Hoogi, A., Rubin,
D.L. and Erickson, B.J., 2017. Deep learning for
brain MRI segmentation: state of the art and future
directions. Journal of digital imaging, 30(4),
pp.449-459.
[7] Handore, S.S.V., Deshpande, A. and Patil, P.M.,
2018, February. An Efficient Algorithm for
Segmentation and Classification of Brain Tumor.
In 2018 International Conference On Advances in
Communication and Computing Technology
(ICACCT) (pp. 83-88). IEEE.
[8] Shil, S.K., Polly, F.P., Hossain, M.A., Ifthekhar,
M.S., Uddin,M.N. and Jang, Y.M., 2017, October.
An improved brain tumor detection and
classification mechanism. In 2017 International
Conference on Information and Communication
Technology Convergence (ICTC) (pp. 54-57).
IEEE.
[9] John, S.A. and Thomas, C., An Approach for
Detection of the Components in Brain MRI Using
Vector Quantization and Morphological
Operations based Segmentation.
[10] Rathi, V.G.P. and Palani, S., 2015. Brain tumor
detection and classification using deep learning
classifier on MRI images. Research Journal of
Applied Sciences, Engineering and Technology,
10(2), pp.177-187.
[11] Choi, H.J., 2011. Synthesis of Multi-View Images
Based on a Convergence Camera Model. Journal
of information and communication convergence
engineering, 9(2), pp.197-200.
[12] Cheng, J. (2017) brain tumor dataset, Figshare,
[online] Available
at:https://figshare.com/articles/brain_tumor_data
set/1512427/5.
[13] Cheng, J., Huang, W., Cao, S., Yang, R., Yang,
W., Yun, Z., Wang, Z. and Feng, Q., 2015.
Enhanced performance of brain tumor
classification via tumor region augmentation and
partition. PloS one, 10(10), p.e0140381.
[14] Kim, H., Kim, J. and Jung, H. (2018)
“Convolutional Neural Network Based Image
Processing System,” Journal of information and
communication convergence engineering . Korean
Institute of Information and Telecommunications
Science, 16 (3), pp. 160–165. doi: 10.6109 /
JICCE.2018.16.3.160.
[15] Frazier, P.I., 2018. A tutorial on Bayesian
optimization. arXiv preprint arXiv:1807.02811.