Brain Tumor

Detection and
Segmenting
•Group members :
1) Mishi MAKADE
2) KGOSI Katlego

•Capstone Project Coordinator’s name(s):

Mr Kudakwashe MADZIMA
Project Layout
 Introduction
 Problem description
 Aims of the project
 Objectives of the project
 Relevance
 Significance of Study
 Limitation of Study
 Methodology
 Method of data collection
 Ethical Consideration
 System specifications
 Network Selection
 FinalModel
 Results
 Challenges
 References
Introduction
• A brain tumor is the growth of abnormal cells in the brain. There are
many types of brain tumors, but this study will focus on
meningioma, pituitary, and glioma tumor.
• Meningiomas start from the membranes that surrounds the brain and
surrounds the central nervous system (Badža et al.2020:1).
• Pituitary tumors are lumps that are inside the skull(Badža et
al.2020:1).
• Gliomas are a general term for tumors that arise from brain tissues
other than nerve cells and blood vessels (Badža et al.2020:1).
• Deep learning algorithms and Machine learning techniques have
been and are still playing a big role in detection of brain tumors.
• The deep learning models previously used to classify brain tumors
are slow and consist of many parameters.
Introduction
• As a result, training these models is
time consuming and requires a lot of
memory.
• This poses a great challenge when it
comes to deploying these models on
resource-restricted mobile devices.
• Mobile intelligence needs faster
mobile processors, more storage space,
smaller but more accurate models, and
even the assistance of other network
nodes.
Problem description
• Accurate and fast diagnosis of brain tumor is crucial as it involves
the selection of the most appropriate treatment method for saving a
patient’s life (Irmak, 2021:2).
• Previously implemented deep learning models on brain tumor
classification like vgg16,Resnet etc are complex and slow (Badža et
al.2020:9).
• Performing real-time performance with these models requires
dedicated hardware (Badža et al.2020:9).
• This may result in delayed diagnosis, treatment and therefore may
increase the chances of a patient’s death.
• The models are also difficult and impossible to deploy on mobile
devices due to their complexity(Anon 2021).
• The proposed Capstone project will address the problem of
complexity by using pre-trained EfficientNetB7 model.
• EfficientNetB7 is 8.4 times smaller than the best existing ConvNet
(Tan & Le 2020:1)
Aims
The aims of this project are to:
1. Develop and compare pre-trained models on
brain tumors detection.
2. Segment brain tumors using image processing
techniques.
Objectives
Gather Gather and perform pre-processing on the dataset.
and
perform

Develop Develop an EfficientNetB7 and compare it to

vgg16 for classification of brain tumors.
Compare the models based on accuracy and
complexity.

Segment Segment the brain tumor using image processing

techniques.
 Relevance

This project implements deep learning

models to classify images.
The use of deep learning models
relates this project to the Data Science
2A and Data Science 2B modules.
The project also implements image
processing and that relates it to the
Image Processing module taken by
third years.
Business Case  Healthcare industry is a very
important industry where best
accuracies need to be achieved.
 As brain tumor is one of top
cancers in men and women,
early and low-cost detection is
what is targeted.
 • The app developed in this project can be
implemented by less experienced medical
practitioners or low levelled practitioners,
example nurses and parents or individuals
who have access to MRI scanners.

Business • This app will provide a second opinion when

diagnosis of brain tumor is carried out in
Case hospitals.
• Due to the reduced complexity of
EfficientNetB7 the app can be deployed on
mobile devices therefore providing easy
access to the app’s services.
 • The development of smaller networks is important
as it is linked to the possibility of deploying the
models on mobile platforms and fast .
• This study’s findings will help medical
practitioners generate fast and reliable diagnosis of
Significance brain tumor through the implementation of
EfficientNetB7.
of study
 Literature Review
• Numerous techniques have been applied to brain tumor classification. One of
those techniques is machine learning (Irmak, 2021:2). Deep learning and
artificial intelligence have made a big impact in the field of medical image
analysis, more specifically in the field of disease diagnosis((Mehmood et al.
2020, 2021; Yaqub et al. 2020 cited in Irmak, 2021:2).
• Kang, Ullah and Gwak (2021:6) proposed a system that used pretrained deep
learning models to extract features and then fed those features to machine
learning models. They used many pretrained deep learning models including
vgg-16, Resnet and others and they also used many machine learning models.
They then studied the performance of each of those machine learning models.
• The results showed that DenseNet-169 deep feature alone is a good choice when
the size of the MRI dataset is very small and the number of classes is 2, the
ensemble of DenseNet-169, Inception V3, and ResNeXt-50 deep features is a
good choice when the size of MR images dataset is large and the number of
classes is 2, and the ensemble of DenseNet-169, ShuffleNet V2, and MnasNet
deep features is a good choice when the size of MRI dataset is large and there
are four classes (normal, glioma tumor, meningioma tumor, and pituitary tumor).
They also discovered that in most cases, SVM with RBF kernel outperforms
other ML classifiers for the MRI-based brain tumor classification task .
Literature Review cont.
• Then Badža and Barjaktarovi´c (2020:4) proposed a new CNN architecture and the classification
was performed using a T1-weighted contrast-enhanced MRI image database which contains three
tumor types. The CNN model created is simpler and it is possible to run it on conventional modern
personal computers. The performance of their network was evaluated using the four approaches:
combinations of two 10-fold cross-validation methods and two databases. The generalization
capability of their network was tested with one of the 10-fold methods, subject-wise cross-
validation, and the improvement was tested by using an augmented image database. The best
results for the 10-fold cross-validation method was obtained for the record-wise cross-validation
for the augmented data set, and the accuracy was 96.56%.
• From all the techniques implemeneted we discovered that the EfficientNetB7 hasn’t been
implemented to classify brain tumors. So, this study will develop the model in brain tumor
classification along with vgg-16 then compare the results based on accuracy and complexity.
Limitations of
study
• Time
• Data
Methodology: Method
of data collection
The dataset was downloaded from
Kaggle.
Dataset Sample
Methodology: System specifications
Programming language :
• Python

Libraries:
• Keras & tensorflow
• Sklearn
• Matplotlib
• Pandas

Models:
• vgg16
• EfficientNetB7
 • The MRIs came in different sizes and were
therefore resized as per requirement of the different
models presented in this study.
Methodology: • The “preprocess_input” function was used for pre-
processing the images along with
Image Pre- “ImageDataGenerator”.

processing • The “ImageDataGenerator” takes the original

image and replaces it with a new pre-processed
image.
Image Pre-processing
Results
 • All models were fine-tuned on the same
Methodology: parameters.
model • They were fine tuned on the following parameters
namely, optimizers, dropout, number of epochs and
development batch size.
• The goal of fine tuning was to get the optimum
and fine-tuning hyper-parameters for the final models.
 • Segmentation is the division of an image into
significant structures.
• Segmentation divides an image into its
Methodology: essential regions or objects, and it is an
important step in image analysis, object
MRI representation, visualization and numerous
Segmentation other image processing tasks.
• This study uses image processing to segment
a tumor.
Segmented
Image
Example
 Ethical Considerations
• The dataset for this study was obtained from Kaggle
and is therefore publicly available.
• Navoneel Chakrabarty and Swati Kanchan along with
the team members Sartaj Bhuvaji, Ankita Kadam,
Prajakta Bhumkar and Sameer Dedge(2020) collected
the dataset.
• The dataset is called Brain Tumor Classification
(MRI). The dataset consists of 2227 training MRIs and
429 test MRIs.
• There are 826 glioma tumor, 247 meningioma tumor,
827 pituitary tumor and 327 no tumor MRIs in the train
dataset. The test dataset consists of 98 pituitary, 100
glioma, 127 meningioma, and 104 no tumor MRIs.
Network Selection
• For our project a good model was one with few training
parameters which indicates the complexity of the model,
and one with high accuracy.
• We used two models namely, vgg-16 and Resnet pre-
trained networks.
• Both models were fine tuned through hyper-parameters.
• Then the final models were trained with the optimum
hyper-parameters.
Final
Model EffcientNetB7 Architecture

• We found that EfficientNetB7 gave best results for the chosen parameters
of interest, namely accuracy in terms of number of trainable parameters
and accuracy.
• EfficientNetB7 has a total of 66 million trainable parameters while VGG-
16 has 138 million trainable parameters, and EfficientNetB7 gave the
highest accuracy.
Results

Trainable Training Validation

Model Parameters Accuracy Accuracy
Vgg-16 138M 88% 92%

EfficientNetB7 66M 96% 98%

Challenges and Areas of Future Research
• Improve our segmentation techniques as
for other images it shows there exist a
tumor where there is none(in the image
on the side it shows as if there is a small
tumor it can be misinterpreted as an
early stage of a tumor)
• Improve and add more performance
measurement techniques of our models
• Badza,M & Barjaktarovi´,M.2020. Classification of Brain Tumors from
MRI Images Using a Convolutional Neural Network.applied sciences
13:1-13. References
• Kang,J & Ullah,Z &Gwak,J.2021. MRI-Based Brain Tumor
Classification Using Ensemble of Deep Features and Machine Learning
Classifiers.applied sciences 13:1-21.

• Irmak,E.2021. Multi-Classification of Brain Tumor MRI Images Using

Deep Convolutional Neural Network with Fully Optimized
Framework. Iranian Journal of Science and Technology, Transactions
of Electrical Engineering 22:1-22.

• Magadza,T.2021. Deep Learning for Brain Tumor Segmentation : A

Survey of State-of-the-Art. Journal of Imaging 22:1-22

• Anon.2021. Deployment of AI model in mobile devices(consulted 02

September 2021):
https://developpaper.com/deployment-of-ai-model-in-mobile-devices/

• Tan,M & Le,Q. 2019. EfficientNet: Rethinking Model Scaling for

Convolutional Neural Networks. Proceedings of the 36 th International
Conference on Machine Learning 7:1-11.