Brain Tumor Classification using Machine and Transfer Learning

Iliass Zine-dine

, Jamal Riffi

, Khalid El Fazazy

, Mohamed Adnane Mahraz

, Hamid Tairi

LISAC Laboratory, Faculty of Sciences Dhar El Mehraz (F.S.D.M),Sidi Mohamed Ben Abdellah University (U.S.M.B.A),

Fès , Morocco

Keywords: The Brain Tumor, Transfer Learning, VGG-16, Deep Features Extraction.

Abstract: Brain tumor classification is a controversial problem in computer-aided diagnosis (CAD). Conventionally,

cancer diagnosis depends mainly on its early prediction. Accordingly, the improvement of technology and the

rise of machines and deep learning facilitate the tasks of tumors’ detection and diagnosis while limiting human

intervention. Transfer learning has been widely adopted in several applications due to its performance. In the

present paper, we have combined VGG-16 and several classifiers for brain tumor classification. Indeed, after

the fine-tuning step of VGG-16, we have fed the extracted features to the classifiers. The proposed approach

has achieved efficient results and has outperformed several state-of-the-art studies in the topic of brain tumors

in terms of precision (98.7%), recall 98.7, F1-score 98.7%.

1 INTRODUCTION

We cannot deny the fact that cancer orders as a

principal cause of death, and an essential factor

determining life expectancy in every nation all over

the earth (Bray F), as in 2020, a brain tumor forms

1.6% of all tumors that can affect the human body,

and 2.5% of fatal tumors (Sung), its constantly

increasing incidence all across the world makes it a

priority to the World Health Organization (WHO) in

terms of screening, diagnosis, and treatment

(Idlahcen Ferdaous, 2020). To be more precise, a

brain tumor is an abnormal growth of tissue in the

brain. Unlike other tumors, brain tumors grow by

regional extension and infrequently metastasize

outside the brain (DeAngelis, 2001). Consequently,

the classification of brain tumors represents a

challenge because of their spread in different

positions. Our goal is to classify the cancer in MRI

images.

Several approaches have been proposed to

automatically predict the existence of a tumor in a

medical images such as those based on machine

learning. These approaches could be divided into two

https://orcid.org/0000-0001-7134-4888

https://orcid.org/0000-0003-0818-7706

https://orcid.org/0000-0000-0000-0000

https://orcid.org/0000-0002-0966-9654

https://orcid.org/0000-0002-5445-0037

main steps: the features extraction and the

classification model. Sharma et al. (Sharma, 2014)

found that human life takes a crucial position in

society and more precisely in the domain of medicine,

it is for that the intervention, the detection, and the

classification of the brain diseases appears very

necessary, the resolution of this problem

classification is frequently based on extracting and

classifying features using a machine learning

algorithm. In this regard, Javeria Amin et al. (Amin J.

S., 2018) proposed a new method to detect and

classify brain diseases (ischemic strokes, gliomas) at

an early step, using machine learning and features

extraction algorithms.

Despite the positive outcomes of these

approaches, they remain limited, especially when

dealing with vast databases. Another sort of approach

based on Deep Learning is proving to be very

effective while handling extensive data. Heba

Mohsen et al. (Mohsen, 2018) presented a new

approach using a DNN learning method to classify

brain tumor images by using fuzzy C-means to

segment the images, and also the discrete wavelet

transforms (DWT) to extract the features, and

principle component analysis (PCA) technique for

566

Zine-dine, I., Rifﬁ, J., El Fazazi, K., Mahraz, M. and Tairi, H.

Brain Tumor Classiﬁcation using Machine and Transfer Learning.

DOI: 10.5220/0010762800003101

In Proceedings of the 2nd International Conference on Big Data, Modelling and Machine Learning (BML 2021), pages 566-571

ISBN: 978-989-758-559-3

reducing the dimensions, and Deep Neural Network

(DNN) for classification. Milica M. Badža et al.

(Badža, 2020) presented a new CNN method to detect

and classify brain tumors, depending on their kind

(Meningioma, Glioma, Pituitary).

To improve the studies made on the topic of brain

tumors, we used a deep learning interface (API

FastAI), which allows us to resize the images and

speed up the training of the applied models.

Transfer learning is defined as an approach that

allows using the information of a model that has

already been trained to learn another target data set.

Arshia Rehman et al. (Rehman, 2020) employed and

explored CNN approaches (AlexNet, GoogleNet, and

VGG-16) to detect and classify brain tumors, using

MRI images.

In this paper, we propose an architecture based on

transfer Learning using the VGG-16 model to our

data set in order to get the feature extractions for the

machine learning models for the purpose of having a

higher prediction score.

The remainder of the paper is organized as

follows: in the second section, we present the related

work, followed by the material and the methods, then

the experimental results, ending with a conclusion.

2 RELATED WORK

The main aim of this part is to describe previous

studies related to the detection and classification of

brain tumors using machine learning and deep

learning models. Many researchers who used

Machine Learning classifier. George, D. N. et al.

(George, 2015) proposed architecture to segment

brain tumors and detect regions of MRI images. In

this regard, Ali and Hanbay (Ari Ali, 2018) proposed

a method that included three main steps: image

preprocessing, image classification with ELM-LRF,

along tumor extraction using image processing

techniques. Amin et al. (Amin, Sharif, Raza, &

Yasmin, 2018) proposed a methodology to segment

and classify the brain tumor using Deep Neural

Networks (DNN) and magnetic resonance images

(MRI).

Deep Learning is another powerful approach to

classifying problems. H. Sultan et al. (HOSSAM H.

SULTAN, 2019) suggested a deep learning approach

that uses a convolutional neural network (CNNs) for

the detection and classification (Meningioma,

Glioma, Pituitary) kinds of brain tumors. Amin Kabir

Anaraki et al. (Kabir Anaraki, 2019) presented a new

approach using CNNs, and a genetic algorithm (GA)

to classify brain tumors based on magnetic resonance

imaging. Correspondingly, Hüseyin and Engin

(Hüseyin Kultu, 2019) proposed a new approach to

categorising liver and brain tumors based on the CNN

efficiency in feature extraction, the capability of the

discrete wavelet transform in signal processing, and

the ability of memory long-term in signal

classification. Parnian Afshar et al. (Afshar,

Plataniotis, & Mohammadi, 2019) proposed a method

to classify brain tumors into three categories:

Meningioma, Pituitary, and Glioma, based on CNN

usage via CapsuleNet architecture and MRI images,

these are frequently used approaches for early

prediction of brain disease.

Although these approaches give perfect results,

another technique based on data pre-training is very

effective when processing extensive data. Several

methods are involved in this challenge; S. Deepak et

al. (Deepak, 2019) applied a pre-trained deep

network, based on GoogLeNet to classify problems

(MR images brain tumors) via transfer learning, in

this regard, Swati et al. (wati, et al., 2019) proposed a

new method that uses the deep neural network, and

trained CNN on the short data set, using a pre-trained

deep CNN model, and proposed a block-based fine-

tuning approach based on transfer learning.

3 MATERIAL AND METHODS

This research study has adopted three main steps

method. First, image preprocessing; second, image

representation; last but not least, image classification.

The structure of the proposed method is illustrated

in the figure below. The acquisition of images are the

first step of this method, the second step is the

preprocessing of images (cropping, resize, and

splitting increase) via the FastAI interface,

afterwards, the classification step by transfer learning

VGG-16, then, we transform the output images into

arrays, in the last step in this study, various Machine

Learning classification algorithms have been used to

compare their performance including Random Forest,

Support Vector Machine, Decision Tree, Gaussian

Naive Bayes, and K-Nearest Neighbor.

Brain Tumor Classiﬁcation using Machine and Transfer Learning

567

Figure 1: Diagram of Brain Tumor Classification Method.

We used a deep learning interface (API FastAI) to

get a reasonable learning rate, Tensorflow, Pytorch,

Keras, Pandas, Numpy, and Sklearn as libraries to

build our models. During training, in order to make

the model converge to the maximum state, the

number of epochs was 80, and the batch size was 32

for all models. We apply RMSprop as an optimizer,

binary_crossentropy as loss, and accuracy as metrics

for the VGG-16 parameters model. Our experiments

were run on the Kaggle notebook, which gave us 16G

RAM and a GPU kernel.

Figure 2: Sample image preprocessing (function cropped).

The last step of our method consists of classifying

and detecting the tumor in the images. The output of

the VGG-16 (features extraction) is fed to several

machine learning models such as Random Forest,

Support Vector Machine, Decision Tree Classifier,

Gaussian Naive Bayes, and K-Nearest Neighbor.

Figure 3: Sample MRI, no brain tumor.

Figure 4: Sample MRI, yes brain tumor.

4 EXPERIMENTAL RESULTS

4.1 Dataset

We obtained the data on which we carried out this

study (Brain MRI Images for Brain Tumor Detection)

from a Kaggle competition (Rakotomamonjy, 2008).

After augmenting and preprocessing the image, we

split the resulting data set into three categories: data

training which includes 1444 (70%) of images, data

validation that contains 310 (15%) of images, and

data testing that includes 309 (15%) of images, as

shown in the following table(1), and then we build

our VGG-16 model on our dataset. The outputs

(features extraction) that we get will be the inputs of

each machine learning model.

Table 1: Representation examples.

Step Number of examples

Train 1444 (70%) of ima

Validation 310 (15%) of images

Test 309 (15%) of images

As we have already stressed, Transfer learning (S.

Deepak, 2019) is an approach that allows using the

information of a model that has already been trained

to learn another target task. Thus, in order to improve

the studies made in this field of brain tumors, we have

applied FastAI (Francis, 2021), which is a library that

allows the processing of images, the construction of

deep learning models in a simpler and faster way.

The implementation of the VGG-16 model is

based primarily on the input layer of data, the

different types of the layers (Conv, Pooling), and the

dense layer on output. After the application on our

dataset, we haven’t only have gained a test accuracy

of 98.7% but also, outputs which transformed into

matrices to exploit them as inputs in the different

models of machine learning.

BML 2021 - INTERNATIONAL CONFERENCE ON BIG DATA, MODELLING AND MACHINE LEARNING (BML’21)

568

Figure 5: Loss and Accuracy of VGG-16 model.

The loss is the sum of the errors made for each

example in the training or validation sets. So we

assume that "the lower the loss, the better the model".

Precision is a measure of the performance of a

classification model (R. Prashanth, 2016).

Informally, precision indicates the percentage of

accurate predictions made by the model. Following

the same path, accuracy is defined as:

𝐜𝐨𝐫𝐫𝐞𝐜𝐭 𝐩𝐫𝐞𝐝𝐢𝐜𝐭𝐢𝐨𝐧 𝐧𝐮𝐦𝐛𝐞𝐫

𝐭𝐨𝐭𝐚𝐥 𝐩𝐫𝐞𝐝𝐢𝐜𝐭𝐢𝐨𝐧 𝐧𝐮𝐦𝐛𝐞𝐫

1

For binary classification, trueness can also be

calculated in terms of positives and negatives as

follows:

𝐕𝐏  𝐕𝐍

𝐕𝐏𝐕𝐍𝐅𝐏𝐅𝐍

2

4.2 Applied Machine Learning Models

In this paragraph, we will cite the Machine Learning

algorithms used to classify brain tumors in this work.

Support vector machines (SVM) (Joachims) are

supervised machine learning methods; they are linear

classifiers, interested in solving discrimination and

regression problems.

The Random Forest Algorithm (Breiman, 2001)

is a machine learning method that applies to several

decision trees formed on subsets of data.

The decision tree (Quinlan, 1993) is a method that

is built in the graph of a tree and gives a group of

choices, the ends or the leaves of the tree represent

the different possible decisions, several fields such as

medicine, commerce, security use this approach

A k-nearest-neighbor (K-NN) algorithm

(Peterson, 2009) is a data classification method that

predicts the probability that a data point is a part of

one set or the other.

The naive Bayesian classification (Liu Ximeng,

2016) is a probabilistic classifier using Bayes'

theorem.

4.3 Performances Metrics

ROC curve (Receiver Operating Characteristic curve)

(Hand, 2009) is a sensitivity/specificity function; it is

a performance measure that allows the characteristics

of a binary classifier to be evaluated.

ROC curve VGG-16

ROC curve SVM

ROC curve Random Forest ROC curve Decision Tree

ROC curve Gaussian Naive

Bayes

ROC curve K-Nearest

Neighbor

Figure 6: ROC curve Machine Learning models

A confusion matrix (S Visa, 2011) is a tool used

to assess the performance of a classification problem.

A confusion matrix is an array of a two-dimensional

Brain Tumor Classiﬁcation using Machine and Transfer Learning

569

Table 2: Comparison between machine learning models prediction.

Model Accuracy Precision Recall F1_score

VGG-16 0.987 0.987 0.987 0.987

SVM 0.825 0.827 0.805 0.830

Random Forest 0.887 0.887 0.887 0.887

Decision Tree 0.686 0.686 0.686 0.686

Naive Bayes 0.595 0.622 0.595 0.585

K-NN 0.796 0.824 0.796 0.6794

array: rows contain predicted values, and columns

have actual values.

VGG-16

SVM

Random Forest

Decision Tree

K-Nearest Neighbor

Gaussian Naive Bayes

Figure 7: Confusion Matrix

In this study, we built five different machine learning

models, starting with the implementation and

augmentation of the dataset to have more images for

the training, the evaluation, and the test of the model

built. We applied the VGG-16 model, followed by

getting the features extraction of this model and

consider them as inputs for all the machine learning

models built. In the end, we get the results of each

model.

After applying the VGG-16 model on our dataset,

along with considering its outputs (features

extraction) as inputs of the following models, the

prediction results based on our machine learning

models achieved 98.7% in terms of accuracy (test).

They ranked as follows (SVM: 82.5%, Random

Forest: 88.7%, Decision Tree: 68.6%, Naive Bayes:

59.5%, and 55.6% K-Nearest Neighbor).

5 CONCLUSION AND FUTURE

WORK

In the manuscript, we have built several machine-

learning models to classify brain tumors. Firstly, we

have implemented and augmented the data set in

order to have more images for the training. Secondly,

we have split the data set into training, validation, and

testing steps. After applying the VGG-16 model, we

extracted the outputs of this model (features

extraction) and considered them as inputs for all the

machine-learning models built. Finally, we get the

results of each model, and we compared them. Our

experimental findings are remarkable bearing in mind

the fact that they demonstrate the capability of deep

learning pre-trained models toward a promising

computer-aided diagnosis in digital pathology.

Eventually, this study is an initiation to other

researches, in light of the fact that some cancer

patterns can not be gleaned by human examination.

REFERENCES

Afshar, P., Plataniotis, K., & Mohammadi, A. (2019).

Capsule Networks for Brain Tumor Classification

based on MRI Images and Course Tumor Boundaries.

Amin, J. S. (2018). Detection of Brain Tumor based on

Features Fusion and Machine Learning.

BML 2021 - INTERNATIONAL CONFERENCE ON BIG DATA, MODELLING AND MACHINE LEARNING (BML’21)

570

Amin, J., Sharif, M., Raza, M., & Yasmin, M. (2018).

Detection of Brain Tumor based on Features Fusion

and Machine Learning.

Ari Ali, H. D. (2018). Deep learning-based brain tumor

classification and detection system.

Badža, M. M. (2020). Classification of brain tumors from

MRI images using a convolutional neural network.

Bray F, L. M. (s.d.). The ever‐increasing importance of

cancer as a leading cause of premature death

worldwide. Cancer. In press.

Breiman, L. (2001). Random Forests, Machine Learning.

DeAngelis, L. M. (2001). Brain Tumors.

Deepak, S. A. (2019). Brain tumor classification using deep

CNN features via transfer learning.

Francis, C. R. (2021). Alzheimer’s Disease Prediction

Using Fastai.

George, D. N. (2015). Brain tumor detection using shape

features and machine learning algorithms.

Hand, D. J. (2009). Measuring classifier performance: a

coherent alternative to the area under the ROC curve.

HOSSAM H. SULTAN, N. M.-A. (2019). Multi-

Classification of Brain Tumor Images Using Deep

Neural Network.

Hüseyin Kultu, E. A. (2019). A Novel Method for

Classifying Liver and Brain Tumors Using

Convolutional Neural Networks, Discrete Wavelet

Transform and Long Short-Term Memory Networks.

Idlahcen Ferdaous, M. A. (2020). Cnn-based approach for

cervical cancer classification in whole-slide

histopathology images.

Joachims, T. M. (s.d.). large-scale SVM learning practical.

Kabir Anaraki, A. A. (2019). Magnetic resonance imaging-

based brain tumor grades classification and grading

via convolutional neural networks and genetic

algorithms.

Liu Ximeng, L. R. (2016). Privacy-Preserving Patient-

Centric Clinical Decision Support System on Naïve

Bayesian Classification.

Mohsen, H. E.-D.-S.-H.-S.-B. (2018). Classification using

deep learning neural networks.

Peterson, L. (2009). K-nearest neighbor.

Quinlan, R. (1993). Programs for Machine Learning.

Morgan Kaufmann Publishers .

R. Prashanth, D. R. (2016). High-Accuracy Detection of

Early Parkinson's Disease through Multimodal

Features and Machine Learning.

Rakotomamonjy, A. G. (2008). BCI competition III:

Dataset II.

Rehman, A. N. (2020). A Deep Learning-Based Framework

for Automatic Brain Tumors Classification Using

Transfer Learning.

S Visa, B. R. (2011). Confusion Matrix-based Feature

Selection.

S. Deepak, P. A. (2019). Brain tumor classification using

deep CNN features via transfer learning.

Sharma, K. K. (2014). Brain Tumor Detection based on

Machine.

Sung, H. F. (s.d.). Global cancer statistics 2020:

GLOBOCAN estimates of incidence and mortality

worldwide for 36 cancers in 185 countries.

wati, Z., Zhao, Q., Kabir, M., Ali, F., Ali, Z., Ahmed, S., &

Lu, J. (2019). Brain tumor classification for MR images

using transfer learning and fine_tuning.

Brain Tumor Classiﬁcation using Machine and Transfer Learning

571