Content uploaded by Aryan Methil
Author content
All content in this area was uploaded by Aryan Methil on Apr 30, 2021
Content may be subject to copyright.
Brain Tumor Detection Using Deep Learning and
Image Processing
Aryan Sagar Methil
Narsee Monjee Institute of
Management Studies
Mumbai, India
aryanmethil@gmail.com
Abstract — Brain Tumor Detection is one of the most
difficult tasks in medical image processing. The detection task
is difficult to perform because there is a lot of diversity in the
images as brain tumors come in different shapes and textures.
Brain tumors arise from different types of cells and the cells
can suggest things like the nature, severity, and rarity of the
tumor. Tumors can occur in different locations and the
location of tumors can suggest something about the type of
cells causing the tumor which can aid further diagnosis. The
task of brain tumor detection can become aggravating by the
problems which are present in almost all digital images eg.
illumination problems. Tumor and non-tumor images can have
overlapping image intensities which makes it difficult for any
model to make good predictions from raw images. This paper
proposes a novel method to detect brain tumors from various
brain images by first carrying out different image
preprocessing methods ie. Histogram equalization and opening
which was followed by a convolutional neural network. The
paper also discusses other image preprocessing techniques
apart from the ones that are finalized for training and their
impact on our dataset. The experimental study was carried on
a dataset with different tumor shapes, sizes, textures, and
locations. Convolutional Neural Network (CNN) was employed
for the task of classification. In our work, CNN achieved a
recall of 98.55% on the training set, 99.73% on the validation
set which is very compelling.
Keywords— Brain Tumor Detection, Computer-aided
Diagnosis, Computer Vision, Convolutional Neural Networks,
Deep Learning, Image Processing, Transfer Learning.
I. INTRODUCTION
A Brain Tumor is a mass of tissue in which the cells
multiply uncontrollably. It arises from different cells - both
in the brain and outside. Primary tumors are the ones that
originate in the brain itself whereas secondary tumors are
the ones that metastasize to different parts of the body.
Tumors can have different origins and based on the cells or
the origin obtained from different types of tumors. For
example, gangliogliomas are tumors that include neoplastic
neurons and are mostly grade I or low-grade tumors which
indicates that the tumor is well-differentiated and has slow
growth [16]. Another example is meningioma which
originates from the meninges (The set of 3 membranes
covering the brain and spinal cord) and can be grade I, II, or
III and it is slow-growing [16]. Symptoms of a brain tumor
include headache which can be acute and persistent,
muscular disorders, dizziness, cognitive disorders, etc.
Treatment of the same includes chemotherapy, radiotherapy,
tomotherapy, and surgery (craniotomy). Although brain
tumor is comparatively infrequent ie. 1.4% new cases per
year [17], in developed countries, fatalities due to brain
tumors have increased over the past few decades. CNS
tumor cases in India range from 5 to 10 per 100,000
population with an increasing trend and it accounts for 2%
of malignancies[11][12][13].
In recent times, Computer-aided diagnosis of diseases is
gaining interest and is helping doctors take swift decisions.
One such approach is using Convolutional Neural Networks
(CNN) to learn the spatial and temporal features from the
given dataset which are necessary to identify the disease. A
Convolutional Neural Network is a special type of neural
network which specializes in handling image datasets. The
very fundamental principle of this neural network is
performing a convolution operation between the kernel and
the image to extract the features. All neural networks learn
by the iterative updation of the weights matrix. Here, it is
required to find the optimal kernel values for all layers of
the CNN model. Hence, the kernel values themselves act as
the weights of this model and the optimal values of the
kernel are gradually learned through backpropagation and
gradient descent. Backpropagation is the computation of
derivatives of the loss function with respect to the weights
and biases in a backward fashion. Gradient descent is the
periodic updation of the weights such that the loss or the
error is decreasing with each iteration. A convolutional layer
is often coupled with a pooling layer and we can connect
multiple such convolutional layer-pooling layer pairs. In the
end, we can have few Dense layers and dropout layers for
the final learning process. Dropout layers are used for
tackling the overfitting problem. The last output layer does
the classification job and can have only 1 neuron in case of a
binary classification task or more than 1 for a multi-class
classification task. Special types of neural networks called
Capsule networks which encode the spatial information well
along with the probability of the object being there in the
image are gaining popularity and they are used in some
recent works [18][19].
Brain tumors in MRI scans (or any other scans) are
identified by abnormal blobs in the brain. These blobs or
regions have a different illumination than the rest of the
brain and are usually brighter than the background.
However, the process of segmenting the tumors in MRI
images is a very difficult task. The tumors have different
sizes, textures, and even the positions where they are found.
If we consider segmenting the tumor by properties such as
illumination, we may face issues such as overlapping pixel
intensities with the normal tissues. Identification and
segmentation of brain tumors in MRI images is important as
it indicates the presence of the abnormal tissues for
treatment or patient follow-up purposes. Most brain tumors
also cause edema [14] which is also a factor that distorts the
nearby structures and can change the pixel intensities around
the tumor. CNN along with some preprocessing techniques
can give accuracy comparable to or even higher than
humans.
In this paper, we have presented a novel brain tumor
detection method in MRI images. The first step was image
preprocessing involving morphological techniques as well
as histogram equalization to enhance our dataset. Then, the
tumor and non-tumor image classification is carried out
using the Convolutional Neural Network (CNN). This paper
strongly emphasizes the use of image processing as digital
images suffer from various problems such as the
illumination problem discussed above and so without proper
image preprocessing techniques, even CNNs can get
misguided to learn incorrect features and produce wrong
outputs. It is also necessary that we choose the right
techniques because a major problem with image processing
for datasets is that the chosen technique might be beneficial
to only a particular type of image instead of generalizing
well to the entire dataset. As such, while experimenting with
the techniques, we found some techniques to be harmful to
the dataset even before fitting a model to it while some
techniques looked promising and so we decided to go ahead
with the training process.
In this paper, we start by discussing the methodology we
used – ie starting from the dataset followed by discussing
the effects of different image processing techniques on the
dataset and then preparing the proper dataset format for
training, image augmentation, and training and through the
performance metrics to the result and conclusion.
II. METHODOLOGY
A. Block Diagram
B. Dataset
The input dataset was mostly made up of a subset of a
dataset[1] consisting of 3762 tumor images and the subset
contained 2297 images. The subset selection was done
based on removing the images which might have
misdirected the model training. Another small dataset of 253
images was added[2]. This dataset has 155 tumor images
and 93 non-tumor images. For more non-tumor images, all
105 non-tumor images from another dataset were used[3].
The non-tumor images folder was named “no_tumor” in the
original dataset on Kaggle. The images were preprocessed
and then a 70%-30% split was performed to get the training
and validation dataset. The preprocessing which was applied
consisted of histogram equalization followed by opening.
The resultant dataset was upsampled to get the final dataset
of 4222 images consisting of 1861 training tumor images,
563 training non-tumor images, 1463 validation tumor
images, and 315 validation non-tumor images. Upsampling
was done as the dataset should be large enough for the
model.
Test Dataset consists of 20 randomly picked images
from the internet out of which 10 had tumors and 10 didn’t.
1.
2.
Figure 1 and 2: Sample non-tumor images from dataset.
3.
4.
Figure 3 and 4: Sample tumor images from dataset.
Dataset
Image Preprocessing
Train- Validation Split
Image Augmentation
Training
Testing
C. Image Preprocessing
5.
6.
Figure 5 and 6: Image Processing on tumor images
7.
8.
Figure 7 and 8: Image Processing on non-tumor images
1. Global Thresholding:
This method proved to be quite ineffective. The
reason being, the different levels of illumination,i.e
the different pixel intensities in different images.
Thus, a global threshold cannot be chosen for all
the images, because of their different pixel
intensities. In the tumor images, the tumor is
brighter than the rest of the brain area. If we set a
threshold value, which separates the bright tumor
portion and the darker brain portion, it disrupts
with some non-tumor images in which the entire
brain portion is bright. Setting a threshold will
classify these non-tumor images as tumor images,
hence proving to be ineffective. See figure 7 where
global thresholding makes the brain MRI image
look like there is a tumor but the true label states
otherwise.
2. Adaptive Thresholding:
Adaptive thresholding does not use a fixed
threshold value for all pixels in the image. Instead,
the threshold is calculated based on the range of
intensities in the pixel’s local neighborhood.
Therefore, different regions of the image will have
different threshold values. This allows the
thresholding to be done dynamically for different
images. We can observe from the adaptive
thresholding output that the outlines are
highlighted. The issue is that even in the area of the
tumor, the outlines are about equally abundant as
there are in non-tumor regions. This makes it
difficult to distinguish the tumor from the rest of
the brain.
3. Sobel filter
Sobel filter is an edge detection filter and consists
of a pair of convolutional kernels. The first of the
pair detects edges running in the vertical direction
while the other detects edges running in the
horizontal direction.
Figure 9: Sobel Filter Kernels
This technique has a tragic effect on the images.
The Sobel filter leads to the introduction of
unwanted noise resulting in ample false detections.
Hence this technique proved to be completely
ineffective.
4. High Pass filter
A 3x3 high pass filter kernel consisting of 8 in the
center and -1 everywhere else is applied to the
images. Since it is a derivative filter, it highlights
the edges and turns the background black. Here it
does a similar job and in doing so, we make the
tumor indistinguishable from the background
which is a tragic effect. So we cannot use this
technique for our task.
5. Median Blur
The median filter is good at eliminating Salt and
Pepper noise. It does so by replacing the central
values of the window of the image under
consideration with the median of all the values in
that window. Since the pixel intensities of the salt
and pepper noise pixels are on the extreme ends ie.
around 0 for black and around 255 for white in an
8-bit image, while computing the median, they lie
on the extreme ends of the list of pixel intensities
(arranged in ascending order for median
computation) of a window. Therefore they rarely
become the median and the non-noise pixels
become the central value; eliminating the noise
pixels. Here we observe that the edges and
boundaries are preserved. Here very little
difference is observed between the original and
preprocessed image. Hence it is not very effective.
6. Histogram Equalization
This preprocessing method results in an overall
improvement. This is because histogram
equalization normalizes the pixel intensities,
thereby normalizing some of the illumination
problems. The technique does a great job in
normalizing the pixel intensities of the brain MRI
image in figure 7. The image belongs to the non-
tumor class but the whiter patch on the top left part
of the brain looks like a tumor although it is not.
Histogram equalization reduced the visual
difference between that area and the rest of the
brain. This was an encouraging result and we can
proceed to fit a Convolutional Neural Network
(CNN) model on histogram equalized images.
7. Dilation
Dilation is an image preprocessing technique
which adds pixels to the boundaries of objects.
White regions such as tumors grow in size after
dilation due to the addition of white pixels on the
boundaries and the gaps in the white regions get
filled as well (See the tumor region in Figure 6). If
we take a structuring element of size 3x3
containing all 1s and we convolve on our image,
any background pixel that is surrounded by at least
1 white pixel, will get changed to foreground pixel.
Thus there is an overall decrease in darker regions
and an increase in the whiter regions thus
highlighting the tumor more. This technique looks
promising and we will proceed with this image
processing technique as well.
8. Erosion
Erosion works opposite of dilation as it removes
the pixels from the boundaries of the objects.
White regions such as tumors shrink in size after
erosion and the gaps and holes in the white regions
increase in size. If we take a structuring element of
size 3x3 containing all 1s and we convolve on our
image, any foreground pixel that isn’t surrounded
everywhere by white pixels, will get changed to
background pixel. This can lead to the tumors (the
main focus of our work) being played down.
Although at first thought, erosion seems to be bad
for this job, erosion combined with dilation
(opening or closing) can prove to be effective.
Opening (erosion followed by dilation) removes
the small objects from the foreground which can be
used to put the focus on the tumor region. Closing
(dilation followed by erosion) removes the small
gaps and holes in the foreground which can be used
to eliminate the darker central regions of the tumor.
Conclusion: Histogram Equalization, erosion, and dilation
were the promising techniques for this task and we tried
different combinations of these for the preprocessing
pipeline.
Training on histogram equalized images gave an accuracy of
98% on the train set and 80% on the test set which suggests
overfitting.
Training on opened (erosion and dilation) images gave an
accuracy of 96% on the train set and 84% on the test set.
Training on histogram equalized and closed (dilation and
erosion) images gave an accuracy of 96% on the train set
and 84% on the test set.
The preprocessing pipeline opted for in the end was
Histogram Equalization followed by opening. This is the
novelty of our method as the 2 techniques chosen in the end
handle the illumination problem very well and also enhance
the morphological features of brain tumors which will help
the CNN.
D. Split the dataset into Train, Validation, and Test sets
The processed dataset has to be divided into Train and
Validation datasets. The training dataset is the one on which
CNN is trained while after each epoch or iteration, the
learned model till that iteration can be tested on a validation
set. The validation set becomes a type of test data as the
model was not trained on that; thereby becoming unseen
data.
Through the metrics we use on the validation dataset, we
can track the progress of the model.
A 70% -30% split on the original dataset gave us the train
and validation set.
Images were duplicated for a good dataset.
Test Dataset:
It has the same folder hierarchy as those of the train and
validation datasets. It contains 20 images downloaded from
the Internet out of which 10 contain tumors whereas 10 do
not.
E. Image Augmentation
Image Augmentation is the artificial increase in the
image dataset by applying specified transformations such as
rotation, horizontal flipping, etc. The need for image
augmentation arises when the existing dataset is not
sufficient for the neural networks or if there is a scope
improvement in the neural net’s performance if more data is
provided. Thus augmentation increases the size of the
dataset. It also prevents the neural network from
memorizing the data by adding some spatial variations to
the images, thus preventing overfitting. The
ImageDataGenerator of TensorFlow makes a generator for
images that can be fed into the neural network for training
and testing purposes.
The need for augmentation here is because neural
networks get better when the datasets are large and since our
dataset contains 4222 images, we should increase it. We can
create multiple variants of the image. Consider a cat image,
if we mirror it from left to right or top to bottom, it still
stays a cat. If we rotate the image, then also it stays a cat.
We can create such variants manually and store them in the
dataset but by doing this, we will require more memory and
time.
ImageDataGenerator creates such variants internally
thereby taking extremely less time and not taking a lot of
memory.
Augmentation applied :
1. All images are resized to 150x150 pixels.
2. Rotation of image between -30 degree and +30
degree.
3. Horizontal shift by 20% of the image width to the
left or right.
4. Vertical shift by 20% of the image height to the top
or bottom.
5. Shear - Distort the image along an axis.
6. Zoom in the image by 20%
7. Apply horizontal flip to the image
Figure 10-a: Original Image
Figure 10-b: Shear
Figure 10-c: Rotation
Figure 10-d: Horizontal Flip
F. Training
We used a convolutional neural network as our model as
CNNs are the neural networks that are best suited for
images.
Transfer learning [15] has been applied which means the
training our neural network will do will be based on a pre-
trained network. We have used a pre-trained model that has
already learned a lot of complex features. The pre-trained
model used is ResNet101v2 which will become our base
model on top of which we will fine-tune our task to classify
tumor and non-tumor images.
Reason for using ResNet :
In the case of normal neural networks, an increase in the
number of layers will decrease the error up to a particular
point but will then start increasing again. These additional
layers should not hamper the training so the solution was
provided in ResNet.
ResNet employs a skip connection where the output of
previous layers will not only affect the next layer but also
layers quite ahead of it. Using Regularization can stop the
adverse effects of additional layers.
On top of the ResNet, a flatten layer was used to flatten
out the matrix output of the previous layer into an array. A
dense layer for further training and to avoid any case of
overfitting, a Dropout layer was used. A Dropout layer
drops nodes with a probability p such that any
overdependence of the output on only particular nodes is
avoided. The value of p used was 0.2.
The activation function used was Rectified Linear Unit
(ReLU) and a sigmoid activation in the last layer as it is a
binary classification-yes or no problem.
ReLU activation function outputs y=x if the input x is
positive and 0 otherwise. It has gained popularity as an
activation function for the non-output layers because it is
easier to train and it also tackles the problem of vanishing
gradients. Sigmoid activation function outputs 1/ (1+e^-x)
for an input x. It outputs a probability between 0 and 1 for
the input to belong to the positive class (1). Sigmoid is used
in the output layer of the model and since it outputs a single
value ie. the probability, there is only 1 neuron in the last
layer.
The model was compiled using Adam as the optimizer
with a learning rate of 0.0002 and the loss used was binary
crossentropy and then the entire model was trained for 20
epochs.
G. Performance Metrics
Performance metrics measure the performance of a
model based on the predictions made v/s the true labels. The
3 metrics were accuracy, precision, and recall. F1 score is
another metric that makes use of precision and recall.
Accuracy is the percentage of correctly classified
data points. Accuracy is not a good metric as it fails to
suggest anything in the case of imbalanced classes. Consider
10 images out of which 9 are tumor images and 1 is a non-
tumor image. If the model learns badly and predicts every
image as tumor images, then also the accuracy would be
90% in this case which is good on paper but it fails to tell us
that the model was bad.
Precision is a metric that says out of all the images
which the metric classified as tumor images, how many of
those were tumors. Suppose the model identifies an image
to be a tumor image, the person can consult a doctor to
check if there’s a tumor. In this case, there is no health risk
ie in case of a false positive, the person will only have to
spend that extra money and time for consulting a doctor.
Recall says that of all tumor images, how many of
those did the model predict that there is a tumor. This is an
extremely important metric and the one we will focus on in
this task. Suppose if a person had a tumor, and the model
classifies it as non-tumor. The person would not consult a
doctor and could die due to the lack of attention given to
that case. Health risk increases if the model predicts a false
negative.
F1 score is a metric that conveys the balance
between precision and recall. It is the harmonic mean of
precision and recall and penalizes the model a lot even when
only one of them is low.
.
Figure 11: Metrics
In the above figure:
TP - True Positive
FP - False Positive
TN - True Negative
FN - False Negative
III. RESULT
We trained our CNN model for 20 epochs and we
recorded the performance metrics after the 20th epoch.
Accuracy
(%)
Precision
(%)
Recall
(%)
F1
Score
(%)
Training
Set
97.94
98.76
98.55
98.65
Validation
Set
99.33
99.45
99.73
99.59
Testing
Set
95
100
90
94.74
Table 1: Metrics values
The high values of the performance metrics are indicators of
a well-trained model for the given dataset and the absence of
underfitting which is good. The recall seems to be lower for
the test set than on the validation set which can be attributed
to the fact that the model was trained on a dataset of MRI
images of a particular distribution while random images
from the Internet might not necessarily belong to that
distribution so the test data can be completely foreign to our
database.
We also have the graphs below depicting the metrics values
associated with the training and validation sets after each
epoch. This can show the entire learning process of the
model and even the presence of overfitting or underfitting
(if any).
Figure 12-a: Accuracy
Figure 12-b: Precision
Figure 12-c: Recall
(The red line graph - Training set
The blue line graph - Validation set)
From the graphs, we can observe that the training process
was smooth and the less gap between the training and
validation lines, indicates high generalization to the
validation images and absence of overfitting.
The test images were plotted and we could see which
images were correctly labeled and which were not.
Figure 13: Wrongly Predicted Test Image
The 1.0 image title in red indicates that the image should
have been classified as a tumor image (1.0) but it got
classified as non-tumor (0.0). The reason for the
misclassification of the above image may be attributed to
the fact that the test images may not necessarily belong to
the same distribution as that of the training or validation set.
Some already existing methods:
While looking at the research papers, we realized that the
standard pipeline for image processing looked something
like greyscaling followed by skull stripping ie. all non-brain
tissues are removed using contouring or segmentation. This
is followed by thresholding using histogram analysis to find
out the appropriate valley region for the threshold value.
The grayscale black and white image is then eroded. In
some places, Berkeley wavelet transform is applied as an
effective segmentation technique. This pipeline was used by
Nilesh Bhaskarrao Bahadure et al. in their paper [4].
A lot of these papers used machine learning techniques that
did not require as many dataset images as deep learning
techniques such as CNN. Hence the experimentation in
these papers was carried out using very few images.
One of the most commonly used techniques is SVM
(Support Vector Machines) which uses what's known as a
kernel trick to find an optimal boundary between the
classes. The reason for its popularity is that it is a robust and
accurate algorithm. In Arakeri et al’s paper [5], the
preprocessing involved rotation correction, image denoising
followed by shape and texture feature extraction, and then
the result was fed into an ensemble learning method that
used SVM, ANN, and KNN. It was used to classify benign
and malignant tumors.
Other papers to use SVM are Mariam Saii et al’s paper [9],
and L. Guo et al’s paper [10]. In the former paper, at the
preprocessing level, an anisotropic diffusion filter is applied
to the image for denoising it. In the latter paper, the immune
algorithm was used to search the optimal weights and
parameters.
Principal Component Analysis (PCA) is also a popular
algorithm that has its use in representing multivariate data
tables into sets of smaller variables. This algorithm is
extensively used for feature selection in a lot of research
papers and this brings down the very large feature count to a
reasonable one. In Sachdeva J et al’s paper [6], 856 regions
of interest (SROI) are extracted by the content-based active
contour model. 218 intensity and texture features are
extracted from these SROIs. Then PCA was used for
dimensionality reduction and the result was fed into an
ANN.
Deep learning techniques such as ANN (Artificial Neural
Network) are also used but it is a very basic and generalized
type of neural network and gets combined with other
techniques such as PCA and KNN (K Nearest Neighbors) to
provide good results.
The metric used was Accuracy in most of the cases and the
best result was observed in D.Sridhar et al’s paper who
performed dimensionality reduction using DCT and then
used a Probabilistic neural network (PNN) for classification.
The accuracy was 100% [7].
Apart from MRI images, tumor classification has also been
performed on CT scan images as seen in Padma
Nanthagopal et al’s paper [8]. A wavelet-based statistical
texture feature set was derived from 2 level DWT. Genetic
algorithm and principal component analysis were used for
optimal texture feature selection. Then SVM was used for
classification purposes.
Most of the research works done on this topic were on very
small datasets whereas our work was done on a
comparatively larger dataset giving it a better ability to
generalize. Our novel method emphasizes the power of
image processing – easy, simple to implement but very
useful at the same time.
IV. CONCLUSION
Brain tumors especially the malignant ones are
considered almost incurable and fatal. The need for early
detection arises from the fact that brain tumors can have
symptoms that do not seem to be alarming at first. The most
common symptom of brain ailments is a headache which
worsens over time in the case of brain tumors. Hence there
are lots of cases where the fatality from brain tumor
increased due to the diagnosis not being done early.
Brain tumor diagnosis begins with an MRI scan which is
followed by studying a tissue sample for determining the
type of tumor. MRI scan can also reveal additional details
such as the size of the brain tumor.
This paper presents a novel method involving image
processing techniques for image manipulation which would
aid our CNN model to classify tumor and non-tumor images
better. Image Processing techniques helped us solve the
illumination issues and brought the tumor into focus. Data
augmentation was used to reduce the chances of overfitting,
as it artificially expands the size of a training dataset, thus
bringing out an improvement in the performance and the
ability of the model to generalize. Transfer learning is also
used as a pre-trained model, ResNet101v2 was used as the
base model, upon which further training was applied to tune
our task. The system recorded an adequate accuracy of
97.94% with an excellent training recall of 98.55 % and
validation recall of 99.73%.
There are limitations to our work as there are small chances
that the image preprocessing applied can damage the
information which makes a tumor image appear non-tumor
in the eye of the CNN model. The input image should be of
a good enough size because if it’s not, after resizing the
image to the size we have set in the image augmentation
step ie. 150x150, the image can become unsuitable for use
(Figure 14-a and 14-b). For future improvements, we can
use ensemble techniques and combine the performance of
different models for better performance.
Figure 14-a: Before Resizing
Figure 14-b: After Resizing
In conclusion, image processing proved to be effective
in solving the illumination problems of the different images
and reducing the noisy details thereby bringing the tumor in
focus. Different variants of the images were created using
image augmentation techniques which augmented the
images and internally created more images for the model.
CNN combined with transfer learning proved to be an
effective training model which can be seen in the extremely
good values of the three performance metrics.
REFERENCES
[1] Jakesh Bohaju, (2020, July). Brain Tumor, Version 3. Retrieved
September 22, 2020, from
https://www.kaggle.com/jakeshbohaju/brain-tumor.
[2] Navoneel Chakrabarty, (2019, April). Brain MRI Images for Brain
Tumor Detection, Version 1. Retrieved September 18, 2020, from
https://www.kaggle.com/navoneel/brain-mri-images-for-brain-tumor-
detection.
[3] Sartaj, (2020, May). Brain Tumor Classification (MRI), Version 2.
Retrieved September 23, 2020 from
https://www.kaggle.com/sartajbhuvaji/brain-tumor-classification-mri
[4] Nilesh Bhaskarrao Bahadure, Arun Kumar Ray, Har Pal Thethi,
"Image Analysis for MRI Based Brain Tumor Detection and Feature
Extraction Using Biologically Inspired BWT and SVM", International
Journal of Biomedical Imaging, vol. 2017, Article ID 9749108, 12
pages, 2017. https://doi.org/10.1155/2017/9749108.
[5] Arakeri, M.P., Ram Mohana Reddy, G. An intelligent content-based
image retrieval system for clinical decision support in brain tumor
diagnosis. Int J Multimed Info Retr 2, 175–188 (2013).
https://doi.org/10.1007/s13735-013-0037-5
[6] Sachdeva J, Kumar V, Gupta I, Khandelwal N, Ahuja CK.
Segmentation, feature extraction, and multiclass brain tumor
classification. Journal of Digital Imaging. 2013 Dec;26(6):1141-1150.
DOI: 10.1007/s10278-013-9600-0.
[7] D. Sridhar and I. Murali Krishna, "Brain Tumor Classification using
Discrete Cosine Transform and Probabilistic Neural Network," 2013
International Conference on Signal Processing, Image Processing &
Pattern Recognition, Coimbatore, 2013, pp. 92-96, doi:
10.1109/ICSIPR.2013.6497966.
[8] Padma Nanthagopal, A., Sukanesh Rajamony, R. Automatic
classification of brain computed tomography images using wavelet-
based statistical texture features. J Vis 15, 363–372 (2012).
https://doi.org/10.1007/s12650-012-0140-3.
[9] Mariam Saii, Zaid Kraitem, Automatic Brain Tumor Detection in
MRI Using Image Processing Techniques, Biomedical Statistics and
Informatics. Vol. 2, No. 2, 2017, pp. 73-76. doi:
10.11648/j.bsi.20170202.16.
[10] L. Guo, L. Zhao, Y. Wu, Y. Li, G. Xu and Q. Yan, "Tumor Detection
in MR Images Using One-Class Immune Feature Weighted SVMs,"
in IEEE Transactions on Magnetics, vol. 47, no. 10, pp. 3849-3852,
Oct. 2011, doi: 10.1109/TMAG.2011.2158520.
[11] Nair M, Varghese C, Swaminathan R. Cancer: Current Scenario,
Intervention Strategies and Projections for 2015. NCMH Background
Papers; 2015. [Google Scholar] [Ref list]
[12] Yeole BB. Trends in the brain cancer incidence in India. Asian Pac J
Cancer Prev. 2008;9:267–70.[PubMed][Google Scholar]
[13] Dasgupta, Archya et al. “Indian data on central nervous tumors: A
summary of published work.” South Asian journal of cancer vol. 5,3
(2016): 147-53. doi:10.4103/2278-330X.187589.
[14] Papadopoulos MC, Saadoun S, Binder DK, Manley GT, Krishna S,
Verkman AS. Molecular mechanisms of brain tumor edema.
Neuroscience. 2004;129(4):1011-20. doi:
10.1016/j.neuroscience.2004.05.044. PMID: 15561416.
[15] Niklas Donges (2019, June 16). What is transfer learning? Exploring
the popular deep learning approach. Builtin. https://builtin.com/data-
science/transfer-learning.
[16] American Association of Neurological Surgeons. (n.d.).
“Classification of Brain Tumors”.
https://www.aans.org/en/Media/Classifications-of-Brain-Tumors
[17] Charles Patrick Davis. (2020, August 24). “How do you get Brain
Cancer?” https://www.medicinenet.com/brain_cancer/article.htm
[18] Vijayakumar, T. "Classification of Brain Cancer Type Using Machine
Learning." Journal of Artificial Intelligence 1, no. 02 (2019): 105-
113.
[19] Pandian, A. Pasumpon. "Identification and classification of cancer
cells using capsule network with pathological images." Journal of
Artificial Intelligence 1, no. 01 (2019): 37-44.