International Journal of Engineering Applied Sciences and Teclogy, 2021

Vol. 6, Issue 7, ISSN No. 2455-2143, Pages 227-238

Published Online November 2021 in IJEAST (http://www.ijeast.com)

BRAIN TUMOR CLASSIFICATION USING

DEEP LEARNING
Himank Dave, Nikhil Kant
BTech in Electronics and Telecommunication
Mukesh Patel School of Technology Management & Engineering, Mumbai, India

Nishank Dave, Divya Ghorui

BE in Electronics and Telecommunication
St. Francis Institute of Technology
Mumbai, India

Malignant tumors are cancerous which can be further

Abstract - The early detection of the tumor plays an categorized as primary and secondary tumors. Primary
important role in the recovery of the patient. In our tumors have starting points within the brain and secondary
proposed model, we have collected MRI scans as it helps tumors can extend from a location to another location.
with the information about the blood supply inside the benign tumors are non-cancerous and can also be
brain. Thus, for the recognition of anomaly, for categorized as adenomas and fibromas, benign tumors and
examining the increasing of the ailment, and for the malignant tumors are also known as low-grade and high-
diagnosis, we prepared a data set consisting of various grade tumors. The detection of brain tumors is done by
MRI images. We then focused on removing unwanted analyzing x-rays MIRs and ultrasound etc. The most
noise and image enhancement. The image common type of primary brain tumor seen in adults is
characteristics can be enhanced by using image pre- gliomas. Pituitary tumors are considered benign and
processing techniques. The image enhancement depends develop in the pituitary gland. They Are less likely to be
upon different factors like computational time, sprayed but can cause serious health hazards because they
computational cost, quality of the uncorrupted image, are present in sensitive areas of the brain. Meningioma
and the techniques used for noise elimination. We have tumors arise from meninges, which are layers of tissues that
made use of various filters for the image pre-processing. protect the spinal cord and the brain. These tumors are
In our next step, image segmentation, an image is considered benign because they have a slow growth and are
divided into several regions. We have implemented also less likely to spread.
different types of segmentation techniques including Image processing and deep learning algorithms are
active contours snakes, fuzzy C means, and region- mainly used to analyze medical images and to identify and
derived triple thresholding. We have further classify the brain tumor into subcategories, and help
implemented two hybrid segmentation models and used professionals with the diagnoses.
computer-aided detection techniques. Post-processing In this paper we present a six-step model to classify and
of the data is done using artificial bee colony detect brain tumors with high efficiency, in the first stage
optimization and watershed filtering and extraction. We we have made use of various filters for the image pre-
then classify two images into tumor and non-tumor processing. In our next step, image segmentation, an image
category using the VGG-16 CNN model. The features of is divided into several regions. We have implemented
the segmented images were further classified into different types of segmentation techniques including active
various types of tumors, including Glioma tumor, contours snakes, fuzzy C means, and region-derived triple
Meningioma tumor, Pituitary tumor, and no tumor thresholding. We have further implemented two hybrid
using one-hot encoding. This approach was further segmentation models and used computer-aided detection
validated using synthetic and real MR image dataset techniques. Post-processing of the data is done using
from Kaggle (name of data set), to detect and classify artificial bee colony optimization and watershed filtering
different types of tumor. and extraction. We then classify two images into tumor and
non-tumor category using the VGG-16 CNN model. The
features of the segmented images were further classified
I. INTRODUCTION into various types of tumors, including Glioma tumor,
A brain tumor is an aggregation of unusual self in the brain Meningioma tumor, Pituitary tumor, and no tumor using
tissue. Brain tumors are classified as malignant and benign. one-hot encoding. Further, this approach was validated

227
 International Journal of Engineering Applied Sciences and Teclogy, 2021
Vol. 6, Issue 7, ISSN No. 2455-2143, Pages 227-238
Published Online November 2021 in IJEAST (http://www.ijeast.com)

using synthetic and real MR image dataset from Kaggle block is individually equalized. So in a small area, a
(name of data set), to detect and classify different types of histogram would confine to a small region.
tumor. However, there are chances of noises being amplified. To
This paper is divided into 6 sections. Section 2 gives us a avoid this problem, we apply contrast limiting. The pixel
literature survey conducted on different methods for a better values above the specified contrast limit are clipped and
understanding of the project. Section 3 describes the then uniformly distributed to other bins before filtering.
methods utilized in this paper. Section 4 provides the An image contains a lot of features like edges, brightness
experimental results over MR images. The conclusions are or contrast, and color. Blurring is a method to process this
given in section 5. Section 6 gives details about the future image. An image looks more sharp and detailed if we can
scope of the paper. perceive all the objects and their shapes correctly. In
Section 2: Methodology blurring, we reduce the content of the edges to transition
from one color to another color very smoothly.
Gaussian filters are fast processing and more effective
II. DATA PREPARATION AND PRE-PROCESSING while smoothening is applied on the images. It might not
Data preparation is a data mining technique that involves preserve image brightness and not particularly effective at
the transformation of raw data to a format that is more removing salt and pepper noise. On the other hand, Median
interpretable and makes the images more suitable for Filters maintain the sharpness of the picture boundaries and
further processing. While preparing the images, we effective in non-linear smoothing. However median filter
lacks in distinguish fine details from the noise. In Bilateral
implemented the technique of crop normalization on MRI
images. This is a technique to find the extreme points in filtering, most of the fine texture is filtered away, and yet
contours. We determined the extreme contours in the input all contours are as crisp as in the original image.
image and outlined the skull. We removed the black
redundant background, without damaging any pixel quality. III. IMAGE SEGMENTATION
The advantage of doing Crop Normalization is that we get
Classically, image segmentation is defined as the
a clear region of interest for further processing and we
partitioning of an image into non-overlapping regions,
remove tags from the MRI without damaging any detail
which are homogeneous with respect to some
from the MRI.
characteristics such as intensity or texture. We have used
The output was as follows:
the following techniques for segmentation of the tumor
We have used Tw1 and Tw2 mixed MRI images of brain
from the MRI images:
tumour. The input dataset contains images with a different
dimension and with different aspect ratios. Therefore, the
A. Active Contours Snakes
images in the dataset are resized to a pre-set format, since
the pre-trained models used require the images to be Active Contours Snakes is a segmentation using curves
200×200 dimensions. and not pixels. Users can define a curve, which will then
FILTERS: evolve.
We have used a variety of filters to smooth out the MRI The evolution of the curve should satisfy two aspects. It
Images. To improve the contrast of the image, we used a should adapt to image edges and it has to be smoothly
Histogram Equalisation filter. An image can have values varying.
confined in a specific range. A brighter image may have all To satisfy this evolution we need to have a measure. This
pixels confined to high values and a dull image may have is parameterized by the energy of the curve. The objective
all pixels confined to low values. However, a good image to satisfy the two aspects is measured by minimizing an
will have pixels from all regions of the image. Hence energy parameter E(c), where c represents the curve under
through adjustment, the intensities can be better distributed evolution.
on the histogram. This method spreads out the most E(c)= Einternal(c) + Eexternal(c)
frequent intensity values, resulting in areas associated with The curve has two aspects. The first is internal to the
lower contrast to gain a higher contrast. Histogram curve which corresponds to the smoothly varying aspect of
equalization normally improves the contrast of the image the curve. This parameter regularizes the curve and makes
Histogram equalization is responsible for increasing the it smooth. It depends on the shape of the curve as we want
global contrast of the image but is more effective only in to minimize the distortions. The second is the external to
images that have close contrast values. the curve that looks for edges. This parameter refers to the
Histogram is not confined to a particular region, so we direction towards which the curve should evolve. External
can lose information due to over brightness. Therefore, to parameter depends on the image intensity.
solve this problem, adaptive histogram equalization is used. Internal Energy:
The image is divided into small 8x8 size blocks and each 1
Einternal(c) = ∫ (𝛼 ∥ 𝑐′(𝑠) ∥2 + 𝛽 ∥ 𝑐′′(𝑠) ∥2 ) 𝑑𝑠
0

228
 International Journal of Engineering Applied Sciences and Teclogy, 2021
Vol. 6, Issue 7, ISSN No. 2455-2143, Pages 227-238
Published Online November 2021 in IJEAST (http://www.ijeast.com)

c'(s) and c''(s) are the first and the second derivative of the The second-order derivative ensures that the curve is not
curve respectively. too 'bendy'. If there are irregularities or oscillations in the
α and β are the two hyper parameters which we have curve, then the c'' value will be high. Hence low second-
taken as 0.06 and 0.6 respectively. These values refer to the order derivative prevents points of the curve from
weights we want to locate to the first and second-order oscillating.
derivatives of the curve. We want to minimize the internal energy to as low as
ds indicates that the curve is divided into small segments. possible and to achieve that we should ensure a curve that
The limit 0 to 1 means we are converting the entire segment is smooth with no oscillations.
of the curve. The first-order derivative is anisotropic, and the second-
We define a region of interest i.e. the tumor outlining. As order derivative is isotropic.
1
it is a discrete system, the curve is divided into various Eexternal(c) = ∫0 −∥ ∇𝐼(𝑐(𝑠)) ∥2 𝑑𝑠 =
segments. We define a starting point. 0 to 1 integral limits 1
𝜕𝐼(𝑥(𝑠), 𝑦(𝑠)) 2 𝜕𝐼(𝑥(𝑠), 𝑦(𝑠)) 2
traverse the curve from the start to the endpoint. The small ∫ − [( ) + ( ) ]𝑑𝑠
0 𝜕𝑥 𝜕𝑦
separation is ds. Each of the location division has an x and
y coordinate which is represented by s. Here, I represents the intensity at locations of x(s) and y(s).
The first-order derivative is to keep the curve not too It is obvious that when we have a large edge the intensity
'stretchy'. If the two-division points selected are far apart, difference is also high. Applying a negative sign to a high
then c' value will be very high. Higher value means less
smoothness. Hence, to make it smoother the points chosen
should be closer. Low c' means we are taking closer
samples.

Fig. 1. Output of crop normalization

Fig. 2.

229
 International Journal of Engineering Applied Sciences and Teclogy, 2021
Vol. 6, Issue 7, ISSN No. 2455-2143, Pages 227-238
Published Online November 2021 in IJEAST (http://www.ijeast.com)

Fig. 3.

magnitude will ensure that the energy is minimum. As the snake Cluster center initialization is done randomly. Any four
moves towards lower energies, it will be attracted towards the points will be randomly selected and be chosen as the center
large edges. of the four clusters we have defined. The Euclidean distance
from each sample point to the center of the cluster is
B. Fuzzy C means calculated using the formula:
Di,j=∑ml=1(xi,l-cj,l)2, 1≤i≤n, 1≤j≤k
There exists a set S={x1,x2.....xn} and any chosen point
xi either belongs or does not belong to this set. Here i and j refer to the sample point and the center
It is an indicator function to determine this possibility. If cluster respectively.
Is is 0, then xi does not belong to S and if Is is 1, then xi Membership function can be defined as the degree to
does belong to S. The set is divided into n number of data which a sample point belongs to a particular cluster. Thus
points, k number of clusters, and m number of numerical points nearer to the cluster center will have a higher degree
attributes. of membership than those towards the edges or further.
Fuzzy set mainly deals with 1's and 0's and all the values in Membership function can be calculated using the following
between them. Fuzzy is unlike crisp sets where membership equation:
values can only take values 0 and 1.
In our project, we have initialized four clusters and Ui,j=1/(∑kl=1(Di,j/Di,j)2/(m-1)) ; 1≤i≤n, 1≤j≤k
chosen a fuzziness factor of 2.

Fig. 4.

230
 International Journal of Engineering Applied Sciences and Teclogy, 2021
Vol. 6, Issue 7, ISSN No. 2455-2143, Pages 227-238
Published Online November 2021 in IJEAST (http://www.ijeast.com)

Fig. 5.

simplicity and computational speed Thresholding occupies

The fraction distance from any point to the cluster center is a central position in image segmentation. In Region Derived
raised to the inverse fizziness parameter, making a fuzzy Triple Thresholding, pixels are clustered into four regions
measurement. This is divided by the sum of all fraction based on their intensities. It uses the process of
distances to ensure that the sum of all memberships is 1. thresholding, with two local thresholds and one global
The centroid for each cluster is further recomputed. This is threshold. There is automatic generation of three thresholds
done by taking the mean of all the sample points in the based on intrinsic characteristics of the image. The global
cluster and taking this value as the new centroid and is threshold is obtained through neighborhood comparison of
represented by the equation: localized regions spanning through the image. The global
Ci,j=(∑ni=1(ui,j)mxi,l)/(∑ki=1(ui,j)m) ; 1≤j≤k, 1≤l≤m threshold forms two sets of pixels, from which the local
The objective function is the product of summation of thresholds are obtained. In our project, we have modified
square of Euclidean distance and membership function this approach by fashioning it with Contrast Limited
raised to the power of a user-defined exponent value. Adaptive Histogram Equalization. At first, we run the
The main objective is to minimize objective function given image through CLAHE filters, and then from the output of
by the equation: the image, we calculate one global and two local thresholds.
Ja(X,C,U)=∑kj=1∑ni=1(ui,j)aDi,j) ; 1≤a<∞ Regions of size 5x5 were considered over the entire image
which was shifted by one pixel horizontally. Expectation
C. Region Derived Triple Thresholding value of pixel intensities for each region calculated:
weighted average of all possible results is calculated by
In the region-based segmentation, the image is taking each potential variable and multiplying it by its
segmented by grouping pixels into distinct regions based on probability of occurrence in that region. This is done till the
some common characteristics like texture, contour, whole image is covered.
intensity level, etc. Thresholding is a method of region- Later we calculate the Average of maximum expectation
based image segmentation, where the grouping of pixels is and minimum expectation value, which gives the global
done based on the difference between the pixel intensities threshold value of the image. Global threshold t divides all
and the threshold intensity, due to its the pixels into

Fig. 6.
two sets- S1 with pixel intensities less than or equal to t and
S2 consists of all pixel intensities that are more than t. Mean
of both sets computed: two local threshold values obtained.
231
 International Journal of Engineering Applied Sciences and Teclogy, 2021
Vol. 6, Issue 7, ISSN No. 2455-2143, Pages 227-238
Published Online November 2021 in IJEAST (http://www.ijeast.com)

Then the Image is segmented into 4 sets using the three B. Hybrid Snake’s-Enhanced FCM Method
thresholds obtained.
The hybrid Snake’s-Enhanced FCM is Edge Based
g(x,y)= a, when 0 ≤ f(x,y) ≤ m1
Snakes segmentation normalized Enhanced FCM (EnFCM)
=b, when m1 < f(x,y) ≤ thres1
feature extraction which combines Snake’s Tumor
=c, when thres1 < f(x,y) ≤ m2
Segmentation method with EnFCM clustering which is an
=d, when f(x,y)> m2
improved version of traditional Fuzzy C-Means introduced
It is seen that even though the proposed algorithm
by Szilagyi et al. with a filter introduced inside the cyclical
clusters pixels into four regions only, it successfully detects
optimization problem, which led to better image quality.
all of the regions of interest (ROI) specified in the images.
By selecting appropriate values of the intensity levels a, b,
c, and d, different areas of the image can be highlighted and
C. Hybrid Triple Thresholding- Modified FCM Method
object identification can be done. It can be inferred that
region-based triple thresholding fusioned with CLAHE is a The hybrid triple thresholding-Modified FCM or bias-
simple yet efficient thresholding method for the corrected FCM is Region Derived Triple Thresholding
segmentation of grayscale images. normalized Modified FCM(MFCM) feature extraction
which combines Triple Thresholding Segmentation into
IV. FEATURE EXTRACTION
four regions based on intrinsic image characteristics with
MFCM clustering which is modified object function of
A. Hybrid Segmentation standard Fuzzy C-Means algorithm introduced by Ahmed
et al. This compensates for intensity inhomogeneities and
Automatic segmentation of the abnormality from the MR
imaging modalities is a huge concern for the radiologist. In the labelling of a pixel takes the labels in its immediate
the past various distinguished approaches and methods have neighborhood into account. This regularisation results in
been proposed based on different approaches like edge piecewise-homogeneous labelling.
detection, region growing, fuzzy logics, and clustering for
fast and effective segmentation. Edge-based snakes used for
the extraction of the tumor region from brain MR image D. Computer-Aided Detection:
gives satisfactory results but due to weak gradient Features of the brain tumor have been extracted to further
magnitude, the algorithm works effectively well only for improve the efficiency of the classification algorithms in
high contrast images and degrades performance with low detecting the type of tumor. The more processing of MRI
contrast images. The various clustering-based segmentation images along with feature extraction is done, the higher
approaches like k-means, fuzzy clustering works well for accuracy will be shown by the classification models.
the extraction of tumors but as the algorithm was based on To detect & extract features of brain tumors, a model is
randomness so whenever the algorithm runs gives different constructed by combining Convolutional Neural Networks
outputs. To overcome we suggest using Hybrid with MRI detection technology using Computer-aided
Segmentation, a Technique integrating two or more detection technology.
techniques which is efficiently giving better results than the The main function of this model is to segment and recognize
segmentation algorithms working alone. A combination of brain tumors and use a convolutional layer to perform
both thresholding and fuzzy clustering techniques will be convolution operations for improving recognition
considered because thresholding finds out the uniform efficiency & rate and combine artificially selected features
regions for fuzzy clustering, which makes the clustering with machine learning features. Also, feature fusion is used
performance better. In this article, we have proposed two by this mechanism to further improve the diagnostic results.
Hybrid Segmentation techniques for user-defined and
automated segmentation.

Fig. 7.

232
 International Journal of Engineering Applied Sciences and Teclogy, 2021
Vol. 6, Issue 7, ISSN No. 2455-2143, Pages 227-238
Published Online November 2021 in IJEAST (http://www.ijeast.com)

Fig. 8.

Fig. 9.

Fig. 10.

V. POST-PROCESSING Artificial Bee Colony and its output is provided to the

We propose a Post-Processing Stage which consists of watershed algorithm. The term watershed refers to a ridge
Artificial Bee Colony Optimization [4] and Watershed [5] that divides areas drained by different river systems. A
of the image is retrieved to extract the tumor. MRI images catchment basin is the geographical area draining into a
always contain a significant amount of noise caused by the river or reservoir. Watershed Algorithm causes averse
segmentation but due to our novel approach of Hybrid
operator, equipment, and the environment. This leads to
Segmentation of Triple Thresholding and modified FCM
serious inaccuracies in the segmentation as any changes in
pixels’ intensity such as noise, significantly affect the with Artificial Bee Colony [6] we get better output at the
clustering results. One way to improve the accuracy of the end of post-processing, with help of spatial domain filtering
results is to use the optimization method. Optimization is a we calculate a gradient image to mark borders of skull
technique used to seek values for a set of parameters that region and tumor region, then we calculate marker image
maximize or minimize objective functions subject to certain and finally develop a tumor developed image which
constraints. Artificial Bee Colony algorithm is a novel highlights tumor region and is then feed to classifiers. Post
optimization algorithm inspired by the natural behavior of Processing removes noise from filtering, operations, and
honey bees in their search process for the best food sources. clustering is done on image and therefore increases output
Inspired by Karaboga’s algorithm we implemented efficiency. Now the extracted image is sent to the classifier.

Fig. 11.
233
 International Journal of Engineering Applied Sciences and Teclogy, 2021
Vol. 6, Issue 7, ISSN No. 2455-2143, Pages 227-238
Published Online November 2021 in IJEAST (http://www.ijeast.com)

value using the NumPy argmax() function and then using

VI. CLASSIFICATION the integer value in a reverse lookup table of character
values to integers.
VGG16 is a convolution neural network (CNN) While classifying the brain tumor, four different classes
architecture that was used to win The ImageNet Large Scale were made for predicting the exact type of tumor. The
Visual Recognition Challenge. It is considered to be one of tumors are labelled as glioma, meningioma, pituitary & no
the most excellent vision model architecture to date. A tumor. Integer coding was used for converting these labels
unique thing about the VGG16 is that instead of having a to integers.
large number of hyper-parameters they focused on having The kind of tumor that gets recognized is marked with
convolution layers of 3x3 filter with a stride 1 and used the value 1, while the rest three columns are marked with zero.
same padding and max-pool layers of 2x2 filter of stride 2. In a dataset, if the images uploaded are {Pituitary,
It follows this arrangement of convolution and max pool Glioma, Meningioma, Meningioma, Pituitary,
layers consistently throughout the whole architecture. In the Meningioma, Glioma, No tumor}, then the following would
end, it has 2 fully connected layers followed by a SoftMax be the One Hot Encoding representation: (Table 2)
for output. The 16 in VGG16 refers to it has 16 layers that On running the code for training the dataset containing
have weights. This network is pretty large and it has about MRI images of the 4 types of tumors using Keras, it was
138 million parameters. observed that the most common primary brain malignancies
In our project, we have used the model for brain tumor are gliomas. Our aim was to diagnosis the type of tumor
image classification. The input dimensions of the model are efficiently using deep learning methods for predicting the
(224,224,3) which is the size of the image that is 224x224. survival rate and, the most suitable treatment planning &
A total of 204 images were used to train the model which identification of risk factors if any, for the patients in time.
may seem less but are more than enough for achieving great To fulfil this aim, accurate & robust tumor segmentation &
accuracy. To attain better metrics and compensate for the classification was required.
loss we reduced the epochs to 25 from 50 and trained them Gliomas are categorized into glioblastoma & lower-
each for 5 steps instead of 8. The results were significantly grade glioma, the former being the most fatal type. Gliomas
better than the previous model. The training accuracy is contain various heterogeneous histological sub-regions,
97.14% while the testing accuracy is 80.85%. The epoch vs including peritumoral edema, a necrotic core, an enhancing
accuracy graph depicts that it is an optimal model which & a non-enhancing tumor core. Magnetic resonance
neither overfits nor underfits. imaging (MRI) is used in radiology very often to portray the
phenotype and intrinsic heterogeneity of gliomas, since
One Hot Encoding multimodal MRI scans, such as T1-weighted, contrast-
Categorical data consists of information that is classified enhanced T1-weighted, T2-weighted, and FLAIR images,
as groups. It is used to group information with similar provide complementary profiles for different sub-regions of
characteristics, which may not have a defined logical order. gliomas.
The parameters in a categorical dataset are labelled Meningioma tumors are often quite vascular without a
according to their features & characteristics. To make blood-brain barrier, whereas gliomas have varying degrees
categorical data more expressive, the approach of One Hot of vascularity and blood-brain barrier disruption depending
Encoding has been used. Both, the input & output variables on grade & characteristics.
which are categories, are converted to numbers. A pituitary tumor is an abnormal growth of cells in the
This approach is termed as One Hot Encoding because, pituitary gland, which is the main hormone-producing
in its representation, each row has only one feature with a gland in the body. About the size of a pea, the pituitary
value of 1 and the rest of the features have a value of 0. To gland is located in the center of the brain behind the nose
convert the categorical data to binary values, the categorical and eyes. A tumor in the pituitary gland may affect the
values need to be mapped to integer values. Following this, balance of hormones in the body and affect a person's
each integer value is represented as a binary vector which overall health.
consists of all 0 values except the index of the integer, A good fit curve was obtained for the model loss graph,
which is marked with value 1. thus showing that the learning algorithm was efficient as it
The categories are first converted into numbers. This is neither overfit nor underfit the model. The training &
done by Integer encoding. A mapping of all possible input validation loss decreased to a point of stability with a certain
values is created from labelled values to integer values. This minimal generalization gap value between the two final loss
mapping is then used to encode the input categorical data. values.
The integer encoding is then converted to One Hot To test the algorithm on a random MRI image, an image
Encoding. Encoding is done one integer character at a time. from the given data set was uploaded on the model. The
Next, the index of a specific character is marked with value classification algorithm worked well on the MRI image, as
1 while the rest are marked with value 0. This has been done the accuracy
by locating the index of in the binary vector with the largest
234
 International Journal of Engineering Applied Sciences and Teclogy, 2021
Vol. 6, Issue 7, ISSN No. 2455-2143, Pages 227-238
Published Online November 2021 in IJEAST (http://www.ijeast.com)

Fig. 12. Epoch VS Accuracy Graph of VGG 16 Model

Fig. 13. Epoch VS Loss Graph of VGG 16 Model

TABLE I
CASE VALUES FOR VARIOUS TYPES OF TUMORS

Tumour type Class

Glioma tumour 1
Meningioma 2
Pituitary tumour 3
No tumour 4

TABLE II
BINARY OUTPUTS FOR VARIOUS TYPES OF TUMORS MENTIONED IN THE EXAMPLE SET

Glioma (1) Meningioma (2) Pituitary (3) No Tumour (4)

0 0 1 0
1 0 0 0
0 1 0 0
0 1 0 0
0 0 1 0
0 1 0 0
1 0 0 0
0 0 0 1

235
 International Journal of Engineering Applied Sciences and Teclogy, 2021
Vol. 6, Issue 7, ISSN No. 2455-2143, Pages 227-238
Published Online November 2021 in IJEAST (http://www.ijeast.com)

Fig. 14. One hot encoding: Model loss graph for Epoch VS Loss Perfect fit observed for epoch: 19 and batch size: 32

patients suffering from various kinds of brain tumor to cure

maximum cases.
obtained for prediction of the tumour was 99.959%. The A web application can be deployed for implementing the
type of the tumour was Pituitary. entire deep learning model so that the dataset containing
The loss of the model is lower on the training dataset than MRI images of the patients can be processed in a
the testing dataset. sophisticated way to get the required results following the
As a good fit was obtained for the model, if the training pipelined algorithm of MRI image Pre-processing-
was continued then the model would have overfitted. Segmentation- Feature Extraction- Classification.
Once designed & deployed with all the required features,
VII. CONCLUSION the web application can be used by the medical
Pre-processing, segmentation and Post Processing Stage professionals in healthcare
plays a vital role in increasing efficiency of detecting and institutes for coming to conclusion about the tumor & its
classifying tumor at deployment of CNN classifier. type and treating several patients at a time.
Artificial Bee Colony Optimization solves the Fuzzy
Clustering Problem and therefore its application in Bio- IX. REFERENCES
Medical Projects is very crucial as it helps in increasing [1] L. Szilagyi, Z. Benyol, S$. M. Szilagy and H. S.
efficiency. Hybrid Segmentation is better than using a Adam, "MR Brain Image Segmentation Using an Enhanced
single segmentation technique as it’s a combination of the Fuzzy (:-Means Algorithm," Proceedings of the 25" Annual
pros of two techniques, giving better segmentation results International Conference of the IEEE EMBS Cancun,
and eases work at the extraction stage. Our proposed Mexico September 17-21, 2003.
algorithm reduces manual segmentation and improves [2] Zhili Chen and Reyer Zwigeelaar, "A Modified Fuzzy
classification. The loss of the model is lower on the training C-Means Algorithm for Breast Tissue Density
dataset than the testing dataset. Segmentation in Mammograms," Proceedings of the 10th
As a good fit was obtained for the model, if the training was IEEE International Conference on Information Technology
continued then the model would have overfitted. and Applications in Biomedicine, Corfu, Greece, 17,
January, 2011.
VIII. FUTURE SCOPE [3] WEIGUANG WANG , FANLONG BU , ZIYILIN
As it is known, brain tumor is a very common disease and AND SHUANGQING ZHAI, “Learning Methods of
millions of people succumb to it every year. The earlier it is Convolutional Neural Network Combined With Image
diagnosed, the higher are the chances for the patients to Feature Extraction in Brain Tumor Detection,” School of
survive. This research primarily deals with the fast Traditional Chinese Medicine, Beijing University of
diagnosis of a large number of patients suffering from brain Chinese Medicine, Beijing 100029, China, VOLUME B,
tumor. The deep learning algorithms used in this research 2020, Digital Object Identifier 10.1 109/ACCESS
deal with quick diagnosis of the brain tumor and its type for .2020.3016282.
thousands of patients at a time, so that medical professionals [4] G.S8.Gopika, DrJ.Shanthini and Dr.§.Karthik,
can provide speedy treatment more efficiently to all the "HYBRID APPROACH POR THE BRAIN TUMORS
DETECTION & SEGMENTATION USING ARTIFICIAL

236
 International Journal of Engineering Applied Sciences and Teclogy, 2021
Vol. 6, Issue 7, ISSN No. 2455-2143, Pages 227-238
Published Online November 2021 in IJEAST (http://www.ijeast.com)

BEE COLONY OPTIMIZATION WITH FCM," Gulyas, "3D Deep Learning on Medical Images: A
International Conference on Soft-computing and Network Review", Sensors, vol.20, 5097, 2020.
Security (IC- SNS), 2018. doi:10.3390/s20185097
[5] Najlae Idrissi, Beni Mellal, Fatima Ezzahra Ajmi and [11] Justin Ker, Satya P. Singh, Yeqi Bai, Jai Rao,
Beni Mellal, “A Hybrid Segmentation Approach for Brain Tchoyoson Lim and Lipo P. Wang, “Image Thresholding
Tumor Extraction and Detection.” ICMCS' 14, Moroccan Improves 3-Dimensional Convolutional Neural Network
imperial legendary city, Marrakech, April 14-16, 2014. Diagnosis of Different Acute Brain Hemorrhages on
[6] Mutasem K. Alsmadi, “MRI Brain Segmentation Computed Tomography Scans,” Sensors, 2019, 19, 2167,
Using a Hybrid Artifical Bee Colony Algorithm with doi: 10.3390/s19092167
Fuzzy- C Mean Algorithm,“ Journal of Applied Sciences [12] Satya P. Singh, Lipo P. Wang, Sukrit Gupta, Balazs
15(1): 100-109. 2015 ISSN 1812-5654 / DOI: 10 Gulyas, and Parasuraman Padmanabhan, “Shallow 3D
3923/jas.2015.100.109, Asian Network for Scientific CNN for Detecting Acute Brain Hemorrhage from Medical
Information. Imaging Sensors,” IEEE Sensors Journal, 2020.
[7] Ciresan, Dan; Meier, Ueli; Schmidhuber, Jurgen, [13] Justin Ker, Lipo P. Wang, Jai Rao, and Tchoyoson
“Mult- column deep neural networks for image Lim, “Deep Learning Applications in Medical Image
classification,” 2012 IEEE Conference on Computer Vision Analysis”, IEEE Access, vol.6, pp. 9375-9389, 2018.
and Pattern Recognition, pp. 3642-3649. [14] Ross Girshick;, Fast R-CNN, Proceedings of the
[8] Justin Ker, Yeqi Bai, Hwei Yee Lee, Jai Rao, and IEEE International Conference on Computer Vision
Lipo P. Wang, “Automated brain histology classification (ICCV), 2015, pp. 1440-1448
using machine learning,” Journal of Clinical Neuroscience,
vol.66, pp. 239-245, 2019.
[9] Yadav, 8.S., Jadhav, 8.M. Deep convolutional neural
network based medical image classification for disease
diagnosis. J Big Data 6, 113 (2019).
[10] Satya P. Singh, Lipo P. Wang, Sukrit Gupta,
Haveesh Goli, Parasuraman Padmanabhan, and Balazs

Fig. 15. Output for an MRI image tested with the One Hot Encoding model

237
 International Journal of Engineering Applied Sciences and Teclogy, 2021
Vol. 6, Issue 7, ISSN No. 2455-2143, Pages 227-238
Published Online November 2021 in IJEAST (http://www.ijeast.com)

Fig. 16. One hot encoding: Model loss graph for Epoch VS Loss. Underfit curve observed for epoch: 4 and batch size 32

Fig. 17. One hot encoding: Model loss graph for Epoch VS Loss. Overfit curve observed for epoch: 30 and batch size: 25

238