TUMOR DETECTION THROUGH

MRI BRAIN IMAGES

ROHIT ARYA
20MCS1009
INTRODUCTION
• Brain Tumor is caused by group of tissues which grows irregularly in the brain. Recently it has
become a major cause of death.
• These tumors can be detected through MRI brain scans.
• Images extracted through MRI Brain scans contains various information about various sections of
the brain, hence they can be used to determine the presence of tumor through various computer-
aided techniques to classify the image with presence of tumor or not.
• Manual detection of brain tumors from magnetic resonance images is a tedious and time-
consuming task. Also the accuracy depends upon the experience of expert. Hence, the computer
aided automatic detection has become important.
• So, the objective of this research is to detect tumors in MRI Brain images using various machine
learning and deep learning techniques to get the highest accuracy for classification.
BACKGROUND WORK

• Brain is a complex organ since it contains more than 10 billion working brain cells. Primary brain
tumors can be either malignant (contain cancer cells) or benign (do not contain cancer cells).
• A primary brain tumor is a tumor which begins in the brain tissue. If a cancerous tumor starts
elsewhere in the body, it can spread cancer cells, which grow in the brain. These type of tumors are
called secondary or metastatic brain tumors.
Types of Brain Tumor:-
1. Primary brain tumors: Tumors which originates in the brain cells are called as primary brain tumors.
In the case of primary brain tumors, sometimes they spread to other parts of the brain or to the
spine. But spreading to other organs occurs only rarely.
2. Metastatic brain tumors: Metastatic or secondary brain tumors are those which originate in other
parts of the body and then spread to the brain. These tumors are named according to the location
which they originate. Based on the malignancy of tumors originated, they are benign and malignant
brain tumors
BACKGROUND WORK

2.1 Benign brain tumors: From the name itself, it can be understood that the benign tumors are
the least aggressive ones. They originate from cells within the brain or from associated
parts of the brain and they will not contain cancer cells. They only grow slowly and also they
have clear borders i.e. their growth are self-limited and they will not spread into other tissues.

2.2 Malignant brain tumors: These tumors contain cancerous cells and their growth is not self-
limited. Often their borders are not clear. Also they grow rapidly and cover surrounding
brain tissue. Hence they will become life threatening if proper treatment is not taken at the
correct time.
PROBLEM DEFINITION

• Manual detection of brain tumors from magnetic resonance images is a tedious and
time-consuming task. Also the accuracy depends upon the experience of expert.
Hence, the computer aided automatic detection has become important.

• So, given any MRI image we want to classify them as tumorous or non-tumorous
using various machine learning and deep learning architectures
MOTIVATION

• Brain tumor is one of the most common major causes for the increase in Mortality among children
and adults in the world and manual detection of brain tumors from magnetic resonance images is a
tedious and time-consuming task. Also the accuracy depends upon the experience of expert. Hence,
the computer aided automatic detection has become important.
• So, the motivation behind this research is to decrease mortality rate through application of various
deep learning algorithms to detect brain tumor with highest accuracy.
 EXISTING WORK/ LITERATURE SURVEY

Heba Mohsen, Abdel-Badeeh M.Salem , “A machine learning technique for MRI brain images

• This study presents a proposed hybrid intelligent machine learning technique for Computer–Aided
detection system for automatic detection of brain tumor through magnetic resonance images.
• The technique is based on the following computational methods; the feedback pulse-coupled
neural network for image segmentation, the discrete wavelet transform for features extraction,
the principal component analysis for reducing the dimensionality of the wavelet coefficients, and
the feed forward backpropagation neural network to classify inputs into normal or abnormal.
• Accuracy on this architecture is:- 90% on test data
EXISTING WORK/ LITERATURE SURVEY

Masoumeh Siar, Mohammad Teshnehlab, “Brain Tumor Detection Using Deep Neural Network and Machine
Learning Algorithm”

• In this paper, a Convolutional Neural Network (CNN) has been used to detect a tumor through brain Magnetic
Resonance Imaging (MRI) images. This paper uses various machine learning models like decision tree and deep
learning model like CNN to give accuracy of the model. This paper uses CNN with softmax and CNN with RBF to
compare accuracy
• Feature extraction on the model is done using K-means clustering algorithm.
• Accuracy on test data:-
CNN + softmax = 91%
CNN + rbf = 90.2%
DT = 80%
EXISTING WORK/ LITERATURE SURVEY

Neha Rani, Sharda Vashisth , “Brain Tumor Detection Using Deep Neural Network and Machine
Learning Algorithm”

• Statistical analysis morphological and thresholding techniques are used to process the images
obtained by MRI. Feed-forward back-prop neural network is used to classify the performance
of tumors part of the image. This method results high accuracy and less iterations detection
which further reduces the consumption time. .

• Method discussed in the paper gives the accuracy of 88% on test data
 EXISTING WORK/ LITERATURE SURVEY
Rupika Nilakant, Hema P Menon, Vikram K” A Survey on Advanced Segmentation Techniques for Brain
MRI Image Segmentation”
• This paper talks about the various image segmentation techniques which can be applied on brain images like
Neural Network Model, Self-Organizing Maps, Radial Basis Function, Back Propagation, Fuzzy C-Means,
Deformable Models, Level Set Models, Genetic Algorithm, Differential Evolutionary Algorithm, Hybrid Clustering
and Artificial Intelligence.
• Conclusions:
1. From the study of these papers, it has been observed that the technique of Fuzzy C-Means is commonly used
for all research. The traditional type of Fuzzy C-Means works well for segmentation in the exemption of noise
2. It is observed that there is no singular algorithm that is used to segment all the regions from the MRI of a brain
image. Multiple algorithms should be used to satisfy the segmentation results.
3. The survey above shows that the hybrid segmentation technique can be used for partitioning the MRI of a brain
image into more than 3 regions
4. Radial Basis Function neural network has advantages of tolerance to the input noise, i.e. the Gaussian and Salt
Peppered noise from the MRI of a brain image will give good results.
EXISTING WORK/ LITERATURE SURVEY

Praveen Gurunath Bharathi, Anita Agrawal “Multi stage classification and segmentation of brain
tumor”
• This paper deals with various machine learning and deep learning algorithms and deep learning
algorithm like Multi-Layer Perceptron(MLP),Support Vector Machine(SVM),Decision Tree(DT), Random
Forest to classify images based on presence of tumor like meningioma and glioma.
• Conclusions:
1. MLP gives accuracy of 87.6%
2. SVM gives accuracy of 49.5%
3. Decision Tree gives accuracy of 83.8%
4. Random Forest gives accuracy of 87.6%
EXISTING WORK/ LITERATURE SURVEY
-SUMMARY
• We have seen various machine learning and deep learning techniques such as back-
propagated neural network, convolutional neural network with softmax and rbf, Multi
Layered Perceptron ,SVM, Decision Tree, Random Forest

• Various image segmentation techiques can be used but it is observed that there is no singular
algorithm that is used to segment all the regions from the MRI of a brain image. Multiple
algorithms should be used to satisfy the segmentation results

• In all the methods discussed back propagated neural network and CNN gives the highest
accuracy of 91%.
PROPOSED WORK

• In this proposed work I tried to implement a CNN architecture with augmented data and a
sigmoid activation at the output to classify the images for the presence of tumor or not.

• This work can be further extended by using various other architectures like VGG-16,
Inception network, ResNets to further improve on accuracy.
 IMAGE PREPROCESSING

Our image consists of black region as background, we want to remove this

background and grab only the top, bottom, left and right side of brain. For this,
we will convert the image to grayscale and use guassian blurring to blur the
image and apply thresholding to remove any noisy points and grab the largest
contor to find the extreme points(Left, right, top, bottom) and crop the brain
image
DATA AUGMENTATION

The dataset contains 2 classes: yes and no which contains 253 Brain MRI
Images. The class yes contains 155 Brain MRI Images that are tumorous and the
class no contains 98 Brain MRI Images that are non-tumorous. Since this is a
small dataset, we can use data augmentation in order to create more images.
We can also get rid of imbalanced data as 61% of previous data contains yes
hence data augmentation will also solve this problem
CNN-ARCHITECTURE-1 WITH CONV(7,7)
CNN-ARCHITECTURE -1
Each input x (image) has a shape of (240, 240, 3) and is fed into the neural network. And, it goes through the following
layers:

A Zero Padding layer with a pool size of (2, 2).

A convolutional layer with 32 filters, with a filter size of (7, 7) and a stride equal to 1.
A batch normalization layer to normalize pixel values to speed up computation.
A ReLU activation layer.
A Max Pooling layer with f=4 and s=4.
A Max Pooling layer with f=4 and s=4, same as before.
A flatten layer in order to flatten the 3-dimensional matrix into a one-dimensional vector.
A Dense (output unit) fully connected layer with one neuron with a sigmoid activation (since this is a binary classification
task).
SIMULATION RESULTS-1

Accuracy-conv(7,7) :
After running through multiple epochs ,
As we can see through multiple epochs training loss and
validation loss decreases.
Training accuracy of the model is quite high 92.3% while the
validation accuracy is 85%
Accuracy on test data is 81%.
CNN-ARCHITECTURE-2 WITH (CONV(5,5))
CNN ARCHITECTURE-2

Each input x (image) has a shape of (240, 240, 3) and is fed into the neural network. And, it goes through the following
layers:

A Zero Padding layer with a pool size of (2, 2).

A convolutional layer with 32 filters, with a filter size of (5, 5) and a stride equal to 1.
A batch normalization layer to normalize pixel values to speed up computation.
A ReLU activation layer.
A Max Pooling layer with f=4 and s=4.
A Max Pooling layer with f=4 and s=4, same as before.
A flatten layer in order to flatten the 3-dimensional matrix into a one-dimensional vector.
A Dense (output unit) fully connected layer with one neuron with a sigmoid activation (since this is a binary
classification task).
SIMULATION RESULTS-2

Accuracy- (Conv(5,5))
After running through 5 epochs ,
As we can see through multiple epochs training loss decrease
but validation loss increases.
Training accuracy of the model is quite high 88% while the
validation accuracy is very low 53%
Accuracy on test data is 61%.
CNN ARCHITECTURE-3 WITH (DROPOUT
LAYER)
SIMULATION RESULTS-3

Accuracy (With dropout)

After running through 5 epochs ,
As we can see through multiple epochs training loss and
validation loss decreases.
Training accuracy of the model is 80.2% while the validation
accuracy is low 79%
Accuracy on test data is 81%.
CONCLUSION

Architecture-1 gives the best validation and test accuracy of 81% which uses
(7,7) covulation layer without any Dropout. Hence architecture 1 is well suited
for classification of MRI Brain Images for presence of tumor or not.

Architecture Training Accuracy Vaidation Accuracy Test accuracy

Architecture-1 92.3% 85% 81%
Architecture-2 88% 53% 61%
Architecture-3 80.2% 79% 81%
FURTHER IMPROVEMENTS

To implement various other Neural network Architecture like AlexNet, ResNet

,VGG-16,19 and applying transfer learning and compare their accuracy to get the
best model for classification of MRI Brain Images.

To come up with the new architecture for classification of MRI Brain images