Brain Tumor Detection using Deep Learning

Models

MSc Research Project

Ms Data Analytics

Nehal Deepak Sawant

Student ID: X19243464

School of Computing
National College of Ireland

Supervisor: Aaloka Anant

 National College of Ireland
Project Submission Sheet
School of Computing

Student Name: Nehal Deepak Sawant

Student ID: X19243464
Programme: Ms Data Analytics
Year: 2022
Module: MSc Research Project
Supervisor: Aaloka Anant
Submission Due Date: 31/01/2022
Project Title: Brain Tumor Detection using Deep Learning Models
Word Count: 5005
Page Count: 20

I hereby certify that the information contained in this (my submission) is information
pertaining to research I conducted for this project. All information other than my own
contribution will be fully referenced and listed in the relevant bibliography section at the
rear of the project.
ALL internet material must be referenced in the bibliography section. Students are
required to use the Referencing Standard specified in the report template. To use other
author’s written or electronic work is illegal (plagiarism) and may result in disciplinary
action.

Signature:

Date: 31st January 2022

PLEASE READ THE FOLLOWING INSTRUCTIONS AND CHECKLIST:

Attach a completed copy of this sheet to each project (including multiple copies). □
Attach a Moodle submission receipt of the online project submission, to □
each project (including multiple copies).
You must ensure that you retain a HARD COPY of the project, both for □
your own reference and in case a project is lost or mislaid. It is not sufficient to keep
a copy on computer.

Assignments that are submitted to the Programme Coordinator office must be placed
into the assignment box located outside the office.

Office Use Only

Signature:

Date:
Penalty Applied (if applicable):
 Brain Tumor Detection using Deep Learning Models
Nehal Deepak Sawant
X19243464

Abstract
Brain tumors are made up of abnormal brain cells. Brain cancer is classified into
benign and malignant tumors. Most tumors are diagnosed using magnetic resonance
imaging (MRI). The early detection of a brain tumor is crucial, since it is a fatal
condition. This study, therefore, employs Inception V3, VGG-16, and ResNet50
models, which are deep learning and transfer learning models, respectively. In this
study, data augmentation is proposed to minimize overfitting since limited MRI
data were used in the project. The study will use hyper-parameter tuning in order
to provide field workers with a more accurate model. Critically evaluated metrics
such as accuracy, precision and recall are used.In this study, VGG16, InceptionV3
and ResNet50 gives accuracy of 75%,67% and 94%. With the ResNet50 giving
better accuracy, it can efficetively detect brain tumor hence healthcare workers can
provide better treatment.

1 Introduction
It is critical in medical image processing to detect brain tumours for MR images. In
order to increase a patient’s treatment options and likelihood of survival, brain tumours
must be detected in their early stages. Detection by hand is highly subjective, differs
from doctor to doctor, and takes a long time to complete. On the other hand, accurate
detection enables doctors to make better decisions by allowing them to replicate findings
and store records electronically, enhancing assessment and treatment planning. In the
last few decades, neural networks (NN) and support vector machines (SVM) have become
increasingly popular for brain tumour diagnosis. The application of deep learning to brain
tumour detection automation, however, has recently gained attention because of its ability
to represent complicated structures, self-learn, and rapidly process large volumes of MRI
data.
This study analyzes brain tumours using a three-step deep network technique.

1. The early stages of model creation utilize multiple pre-processing approaches, in-
cluding the augmentation of data and normalization of MRI image data, to minimize
over-fitting.

2. Transfer learning models like VGG-16, InceptionV3, and ResNet50, are utilised to
build classification models. Hyperparameter tuning is performed during the model
building process and this helps in fine tuning the base models.

3. A better model is picked based on the evaluation and comparison of the performance
measures

1
With this three-stage procedure, clinical problems can be diagnosed using an MRI quickly
and accurately.

1.1 Background and Motivation

Different types of cells make up the human body, each with its own function. It is neces-
sary for bodies to remain healthy for cells to divide and expand in a controlled manner.
In the absence of proper cell division, cells multiply uncontrollably, resulting in the form-
ation of tumours. tumours in the brain are lumps of tissue containing abnormal cells.It
is important to diagnose tumours early, even if they are not malignant, so that treatment
can beginRaut et al. (2020).
To determine the location, size, behavior, and growth status of a suspected brain tumour,
a radiological investigation is necessary. It is simple to determine the correct treatment for
a patient based on this information, such as radiation, chemotherapy, optimum therapy,
and surgery. It is also important that the chances of survival of an infected patient may
increase with early detection of a brain tumour. Many types of new imaging technology
are transforming the medical industry. Most commonly used imaging techniques include
magnetic resonance imaging (MRI), computed tomography (CT), positron emission tomo-
graphy (PET), single photon emission computed tomography (SPECT), ultrasound, and
X-ray Brain tumor detection in MR image using superpixels, principal component ana-
lysis and template based K-means clustering algorithm (2021). MRI is the most common
non-invasive method for identifying changes in tissue composition. Due to its ability to
produce the highest contrast pictures of the brain and cancerous tissues, it is preferred
over other medical imaging techniques. MRI provides abundant information on clinically
suspected illness, as well as highly detailed images of the human body. Medical and
biological research also greatly benefit from MRI.Liu et al. (2014)
Cancer detection may be enabled by using models such as ResNet50, InceptionV3 and
VGG16, which are deep learning approaches. With these models, one can recognize faces,
analyze images, categorize, comprehend the weather conditions, and many more.
Deep learning, as a more specific and unusual technique, has sparked intense interest
across the board, including among the medical profession, for image analysis, with pre-
dictions that deep learning will control roughly $300 million of medical imaging market
by 2021, Sharma and Miglani (2020). A significant portion of the funding would go to
medical research, and the accuracy and outcomes would be improved for complicated
projects.
The process of transfer learning in deep learning is where patterns previously learned
from solving one problem are used to solve another instead of starting from scratch. By
using pre-trained models that were already trained on large benchmark datasets, transfer
learning can be achieved. The study applies the data augmentation technique to classify
brain tumours using MRI images as cancerous or non-cancerous through our study using
Brain MRI Image and also transfer learning models, such as Inception V3 and VGG-16,
are used. The data augmentation techniques are applied to the MRI slices for generaliz-
ation of results, increasing the dataset samples and reducing the chance of over-fitting.
Two stages define the purpose of this study:

1. Using deep learning and transfer learning, the study can identify brain tumours.

2. Doctors and healthcare experts will be able to diagnose brain tumours as soon as
possible using this approach.

2
1.2 Research Objectives
Research question of the study is defined and explained below
How accurately deep learning models using transfer learning can detect brain
tumour?
The given Table 1 states all the research objectives defined in the study and meant to solve
research question. The main objective of this study is to detect a brain tumour as early
as possible so that the health-workers can provide better treatment hence, improving
medical facilities. To achieve this objective various other objectives such as building
a strong deep learning model by identifying gap between deep learning and machine
learning is been defined.
Table 1 states all the objectives in detail with evaluation matrix.
Research Objectives Description Evaluation Metrics
First Objective The evaluation in study helps -
to recognize gaps of previous study
Second Objective Visualization and augmentation of data -
using pre-processing techniques
Third Objective Modeling and evaluating VGG16 Accuracy, precision
recall
Fourth Objective Modeling and evaluating ResNet50 Accuracy, precision
recall
Fifth Objective Modeling and evaluating InceptionV3 Accuracy, precision
recall
Sixth Objective Analyzing all the models based Accuracy, precision
on evaluation metrics recall
Table 1- Research Objective of this research

2 Related Work
2.1 Papers related to Brain tumour Detection using Machine
Learning Algorithms.
Machine learning algorithms are used to detect brain tumours in this study Sharma et al.
(2014) . Texture-based information (GLCM) is retrieved using Gray Level Cooccurrence
Matrix. In this suggested study, energy, contrast, correlation, and homogeneity are all
textural characteristics of the picture. 212 brain MR images were analyzed using the
Multi-Layer Perceptron and Nave Bayes machine learning algorithms, with the best sens-
itivity being 98.6% and 91.6 %. It is almost certainly possible to improve this accuracy
by analyzing a large set of data and extracting both intensity-based characteristics and
texture-based characteristics.
The authors, Mudgal et al. (2017) propose an efficient and precise model which requires
MRI images to be handled using a modified K-Means Clustering algorithm so that detec-
tion is reduced later on, and then using Mean Shift Detection to maximize the efficiency
of images entering the processing stage and to increase MRI contrast and intensity. Com-
paring XGBoost models with other comparable machines, as well as hybrid models that
employ these extreme models, on a large dataset presents a wealth of research opportun-
ities. Feature selection can be improved using PCA, LDA, or GA, as well as eXtreme

3
boosted machines. A number of alternatives have been presented for watershed manage-
ment, and the existing system can be improved.
In order to increase segmentation accuracy, a variety of strategies have been used in this
study Abbas et al. (2019). These included noise reduction and image enhancement. In
this data set, some of the samples had poor MR image quality, so these pictures have been
enhanced. This project now includes LIPC. Another way of reducing requirements is by
using LDA. Further, the work can be complicated and time-consuming due to the addi-
tion of underlying or textural attributes. Based on the division result, our LIPC-based
system provides more and more appropriate precision compared to other methodologies.
In this investigation, the dice score was 0.95. Future research will apply Deep Learning
models like CNN. In order to detect photos automatically, CNN is the best model.

2.2 Papers related to Deep Learning Models

In this study Bhanothu et al. (2020), a deep learning system was used to identify and
categorize brain tumours from MR images. Using the Faster R-CNN method, tumour
locations were identified and categorized into three categories: gliomas, meningiomas,
and pituitary tumours. Faster R-CNN was implemented using a deep convolutional net-
work architecture known as VGG-16. With the test dataset, a higher mAP was achieved
for identifying brain tumours. The proportion of the tumour’s area relative to the brain
may also be determined using this technique. In addition to skin lesion segmentation and
classification, this technology has applications in other medical fields.
Researchers, Grampurohit et al. (2020) are using deep-learning techniques such as the
Convolutional Neural Network (CNN) model and the VGG-16 architecture (created from
scratch) to determine where tumours are located in scanned brain images. Brain MRI
scans of 253 individuals were analyzed, 155 of which were tumourous and 98 of which
were not. The results of this research are compared with those of the CNN and VGG-16
models.
The characteristics of current pre-trained models are often derived from layers that are
different from natural or medical images. Taking this issue into account, this study
Noreen et al. (2020) suggests a concatenation and multi-level feature extraction method
for the early detection of brain tumours. As it uses Inception-v3 and DensNet201, this
model makes sense since these deep learning models are pre-trained. Two alternative
scenarios were examined using these two models for detecting and classifying brain tu-
mours. A pre-trained Inception-v3 model was used to combine the characteristics from
several Inception modules for brain tumour classification. These features were then used
to classify the brain tumour using the softmax classifier. A pre-trained DensNet201 was
used to extract features from multiple DensNet blocks.
Using CNN, the researchers Choudhury et al. (2020) are able to discriminate tumourous
MRI images from non-tumourous ones. It achieved 96.08 percent accuracy with a f-score
of 97.3. It consists of just three layers of a CNN and several preprocessing steps to pro-
duce results in 35 epochs. In this study, we aim to emphasize the relevance of diagnostic
and therapy prediction applications based on machine learning. Scientists plan to use
Convolutional Neural Networks to identify brain tumours using neutrosophical principles
in the future.
Brain tumours are caused by uncontrolled cell proliferation. MRI can be used to identify
this type of tumour, which helps to segment the tumour territory into different ways for
surgical and medical planning. The manual identification method is cumbersome and sub-

4
ject to errors. Several deep learning algorithms are used by researchers Sangeetha et al.
(2020) to identify the tumour site automatically. Millions of photos are trained and fine-
tuned within a short time. As an additional benefit, this work provides multi-iteration
classification algorithms based on CNNs such as GoogleNet, VggNet, and ResNet 50. In
terms of speed and accuracy, ResNet 50 is demonstrated to be superior to VggNet and
GoogleNet.
In this publication Raut et al. (2020), a CNN model is proposed for detecting brain tu-
mours. The first step is to enhance brain MRI scans so that there is enough data for deep
learning. Pre-processing of the photos follows, enabling them to be processed for the next
stage. It is based on MRI brain images that have already been processed to use training
characteristics to identify newly input images as tumours or not. A back propagation
technique reduces error during training and produces more accurate results. A picture is
constructed by removing extraneous information with autoencoders and segmenting the
tumour region using K-Means, which is an unsupervised learning algorithm.
The authors of this publication Nalepa et al. (2019), examined in detail several brain tu-
mour detection strategies. In addition to this, a deep-learning technique based on CNN
has been developed to detect and classify multiple forms of brain tumours. The limited
size of the medical imaging dataset is being improved using a data augmentation strategy.
As a data augmentation strategy for medical imaging, rotating at 90 degrees was found
to be more effective. The proposed model was also assessed against AlexNet and VGG-16
models from the point of view of sensitivity, precision, specificity, f1, and accuracy using
MATLAB scripts. Alternative networks did not perform as well as the suggested model,
which achieved an accuracy of 96.05 percent.
Convolutional neural networks are presented as a method of identifying tumours in brain
scans in this paper Seetha and Raja (2018). Input brain pictures are then used to train
the CNN model, and features are extracted. A fully connected layer followed by soft-
max activation is used to categorize the pictures based on the retrieved characteristics.
Harvard Medical School’s database of MR brain images is used to evaluate the strategy.
For both classification and training, CNN models like VGG16, ResNet, and Inception
are used. The experimented database offers 100 percent accuracy. Data augmentation
is used to improve segmentation accuracy. A learning rate that prevents overfitting the
model is chosen. MR brain images will be categorized in the future based on tumour
grade, allowing for better analytics and treatment planning.
In deep learning optical representation analysis Saxena et al. (2021) , convolutional neural
networks are often used. There are currently techniques available for detecting brain can-
cers, but they require large datasets and image processing techniques. Three of the main
components of the proposed system (CNN) are image enhancement, preprocessing im-
ages, and the use of convolutional neural networks. A deep learning algorithm and a
large dataset are used to create a system. In this study, the CNN outperformed all other
state-of-the-art training techniques with an accuracy of 87.42 percent and minimum dif-
ficulty.
Expertise and understanding of doctors are crucial in treating brain tumours. As a result,
an automated system for detecting brain tumours is critical for radiologists and clinicians.
This research focuses on the use of convolutional neural networks (CNNs) for predicting
brain malignancies and accurately interpreting magnetic resonance imaging (MRI) data.
An example of a method is the use of picture pre-processing algorithms that identify
extreme shapes and trim away black borders from images. The photos are then scaled
to the appropriate size. An Adaptive Moment Estimation (Adam) Optimizer was imple-

5
mented in our network in order to speed up the training process.In order to evaluate F1
score, the researchers used MRI dataset which was developed by their entire team. on
Kaggle to train and evaluate the best model.
This research Zhang et al. (2021) uses a novel cross-modality deep feature learning ap-
proach to detect brain cancers in multi-modality MRI data. Rather than rely on small
data sets, the focus is on mining patterns across multi-modal data sets. The cross-
modality feature transitions (CMFTs) and cross-modality feature fusions (CMFFs), both
of which essentially aim at learning rich feature representations by merging and transiting
knowledge obtained from different modes of data. A cross-modality deep feature learning
framework based on cross-modality deep learning is found to improve brain tumour seg-
mentation performance compared to baseline approaches and state-of-the-art methods as
demonstrated in extensive trials on the BraTS benchmarks.

3 Methodology
Specifically, the research was aimed at the development of a device to aid doctors in
diagnosing and improving life expectancy. An automated method could potentially re-
place the time-consuming and ineffective process of physically detecting brain tumours.
In a related study, deep learning proved to be a very effective method for detecting brain
tumours. It has been found that reviewing a variety of data sources for decision-making
can increase profitability and originality. The study follows CRISP-DM methodology to
detect if there’s a tumour present or not. Figure 1

Figure 1: CRISP-DM Methodology

3.1 Data Collection

It is easier to identify anomalies of tissue composition using MRI, which is a non-invasive
method. Because it produces a clear picture of the brain and cancer tissues better than
other imaging techniques, it is preferred over other methods. An MRI provides excellent

6
 Figure 3: Tumor Not present
Figure 2: Tumor Present

anatomical picture structures and abundant information for diagnosing clinically suspec-
ted illnesses. Additionally, it contributes to medical and biological research. One of the
most obvious benefits of MR imaging is the high spatial resolution with cross-section pic-
tures. The study uses Brain MR images to detect tumourous or non-tumourous images.
The given dataset has two folders namely yes which represent tumourous images as shown
in first image whereas the second folder represents non tumourous images as shown in the
second image. The dataset contains a total of 253 tumourous and non tumourous images.

3.2 Data Preparation

Once MR images are gathered, it is necessary to prepare these MR images before training
so that the models are efficient. Firstly dataset is divided into three folders namely ‘Test’,
‘Train’, ‘Val’ and sub-folders namely ‘yes’ and ‘no’ are created. In the next step image
normalization is performed, the process of normalizing alters the range of pixel intensity
values. The primary goal of this session is to identify extreme points in the image and
crop the edges. Noise is removed from the image by performing series of erosions and
dilations . Finding contours in the threshold image and then extreme points in MR image
is retrieved.
Figure 6 shows the cropping of image which is then applied to the dataset. In this session,

Figure 4: Image Cropping

data augmentation is also performed. Deep neural networks’ ability to generalize is

improved by data augmentation, also known as implicit regularization. Such information
is of critical importance when there is a shortage of high-quality data and gathering
additional samples is both costly and time-consuming. Medical picture analysis typically
presents this problem, particularly in the context of tumour delineationNalepa et al.
(2019). 5

7
 Figure 5: Data Augmentation Technique Applied

3.3 Modeling Techniques

3.3.1 VGG-16

Figure 6: VGG16 Architecture

The Visual Geometry group is abbreviated as VGG. The weights and convolutions
are divided evenly into 16 layers, and the output layer is completely linked to the other
layers. In the study, VGG-16 downloads the Imagenet weights with an input picture size
of (224,224,3); therefore, the first step is to obtain the weights from Imagenet.Figure 6
Wickrama Arachchilage and Izquierdo (2020) shows the architecture of VGG16. As
implied by the name, the VGG-16 model does not require any retraining.

8
 Figure 7: Model Summary of VGG16

3.3.2 InceptionV3

Figure 8: Inception Architecture

Figure 8 Wickrama Arachchilage and Izquierdo (2020) shows the architecture of In-
ceptionV3. The Inception-v3 convolutional neural network is based on the ImageNet
database and trained by the Google Brain Team. The system divides photos into 1000
categories depending on how many layers there are in the deep neural network. This
model is capable of learning rich representations of various photographs. Each input
image needs to be 224x224 pixels. With Inception-v3, the number of parameters is dra-
matically reduced by batch normalization, picture distortion, RMSprop, and several tiny
convolutions Sharma and Miglani (2020). The InceptionV3 model does not need to be
retrained as, as its name suggests, it has been trained on a huge data set and has been
used to classify a variety of classes. InceptionV3 will then be set to ’FALSE’ for the
output layer. Other than the output layer, all of the layers have been disabled. The final
step will be to introduce a layer of new classification to train the data set photographs.
As a result of the research, the layer will be flattened and the ’ReLu’ activation function
will be used to create a 1024-layer connected layer. A dropout layer of 0.1 is added to
eliminate overlapping. In the dense output layer is a function known as ’Softmax’ that
activates the output.

9
Figure 9: Inception Model

10
3.3.3 ResNet50

Figure 10: ResNet50 Architecture

Resnet-50 is a residual network consisting of 50 layers and Figure 10 Wickrama Arachchil-

age and Izquierdo (2020) shows an architecture of ResNet5 . Remaining features are called
residues. As a result of the layer’s learning, residue corresponds to its characteristics.
The training of these networks is much easier than that of traditional deep convolutional
neural networks. Moreover, they are also less prone to accuracy degradation.The Res-
Net50 model does not need to be retrained as, as its name suggests, it has been trained
on a large number of images and has been used to classify a variety of classes. Resnet50
will then be set to ’FALSE’ for the output layer. Other than the output layer, all of the
layers have been disabled. As a result of the research, the layer will be flattened and the
’ReLu’ activation function will be used to create a 1024-layer connected layer. A dropout
layer of 0.1 is added to eliminate overlapping. In the dense output layer is a function
known as ’Softmax’ that activates the output.

Figure 11: ResNet50 Model Summary

11
4 Design Specification
The design architecture specified in the study is used to complete the project. The
architecture includes a Data Layer, a Business Layer, and a Client Layer.

Figure 12: Proposed System Designed

1. Data Layer: In this layer, different steps such as collection of data, image pro-
cessing and data augmentation techniques are implemented. Initial dataset contains
253 MRI images of brain which states whether a patient is having a brain tumour
or not. The MRI image set includes overall of 155 pictures that depicts a tumour
in a brain and rest representing tumour not present. Once the dataset is gathered,
normalization technique is performed. In normalization technique, firstly a series of
erosions and dilations are performed to remove noise present in the images. In the
next step, finding contours in the threshold image and then extreme points in MR
image is retrieved. Finally, the images are cropped according to the extreme points.
Once image is pre-processed, Data augmentation is performed to scale, rotate and
to solve overfitting problem in a dataset.

2. Business Layer: Once data is pre-processed, it is ready to train the models.

In this layer various transfer learning models such as VGG-16, InceptionV3 and
ResNet50 are trained to achieve results. All these models have been built using
python language. In this section, various evaluation metrics are critically evaluated
to achieve the research objectives and also help health practioners to provide better
treatment by detect a brain tumour as early as possible.

3. Output Layer: After critically evaluating results, outputs are provided to health
practitioners to provide better and effective treatment. By studying classification
reports and graphs it will be easy to detect a tumour present in a brain or not. The
underlying architecture and accompanying requirements of the implementation are
presented in this section.

12
5 Implementation
5.1 Setup
The study uses Google Collaboratory with 13.6 GB RAM, this test is performed. Be-
cause big image data necessitates additional layers, all models run slower. GPUs are
used to accelerate the processing of transfer learning models in real time. TensorFlow
and Keras are used to implement these models. Google Collaboratory Notebook uses
Python 3. The information is stored in Google Drive. Numpy and Keras are used for
image normalization, argumentation, and cropping.

5.2 Data Handling

In the beginning, image data set is mounted from Google Drive to Google Colbartory.
The given data set includes two sub folders namely ‘yes’ and ‘no’. The dataset contains
a total of 253 images. In sub folder ‘yes’, it contains a total of 155 images whereas, sub
folder ‘no’ contains a total of 98 images. Data obtained from the API and data source is
excellent, and both of them are reliable. Python programming language is used to per-
form image pre-processing techniques such as cropping. Te given image dataset is divided
into three parts namely test, train and val. Normalization technique is performed. The
Image Augmentation, Keras, and Tenserflow libraries are utilized to establish a uniform
dimension for all inputs, making the model more general and specific. In the training
set, 70% of the data is placed, while the remaining 30% goes to the testing set. Study-
ing parameters that affect post-processing throughput, which all affect the extraction of
function patches, is crucial for this study. Finally, all parameters are recorded for the
uncommon method after the components are removed for further inspection. This can
be achieved through Transfer Learning.

5.3 Transfer Learning

A trained model is used to learn a new collection of data by transferring the information
from one problem to another. In this study, a transfer learning environment was em-
ployed. The pre-trained CNN designs include VGG16, ResNet-50,Inception-v3. Human
accuracy is being tested using the best available findings. A picture is input into the
networks, and they output the picture’s object name as well as the probability of its
category.Figure 13 explains workflow of ResNet50.

13
 Figure 13: WorkFlow of ResNet50

6 Evaluation
Under this section, the results various transfer learning models such as VGG16, Incep-
tionV3 and ResNet50 are evaluated. Various evaluation metrics such as confusion matrix,
accuracy, recall and model performance plots are critically evaluated. The accuracy and
loss of each process are calculated for each epoch while evaluating models. Accuracy and
loss charts are presented in this section. Based on the prediction and test results, the
test accuracy is calculated.

6.1 VGG16 Evaluation

Figure 14: Classification Report of VGG16 Model

From Figure 14, There were different epoch values tested and epoch=5 provided the
best accuracy. This model employs a batch size of 32. The number of accurately classified
predictions is known as precision,for labels ‘yes’ which is tumour positive and ’No’ which
is tumour negative the correct classification percentages are 88% and 69%. This recall
which is actually correct classification in this case is 58%, 92% for a brain tumour and
non-tumorous images. F1 is simply the sum of recall and precision

14
6.2 InceptionV3 Evaluation

Figure 15: Classification Report of Inception Model

With Inception-v3, the number of parameters is dramatically reduced by batch nor-

malization, picture distortion, RMSprop, and several tiny convolutions.
For brain tumor detection, InceptionV3 model provides accuracy values of 67 percent
with precision value of 73% for tumorous images and 63% for non-tumorous images. For
brain tumour and non-tumours images, the recall, i.e. actual values of true positives, is
53% and 80% respectively, from Figure 15 .
In order to achieve a better learning rate, the RMSprop optimizer is used, because the
output is multi-class classification categorical cross-entropy. Using the hyperparameters
adjusted, the epoch value for this model is 10. This model has a batch size of 32 and an
epoch value of 10.

Figure 16: InceptionV3 Model Loss Plot

The given Figure 16 illustrates Inception Model loss throughout the epoch process.
It is seen that the training losses fluctuated throughout the epoch process and at last
went down. Whereas, model loss of validation data was stable till epoch 3 and then was
fluctuating a bit till the end.

The given Figure 17 explains accuracy of training and validation tests through the
epochs. It is seen that for validation set the accuracy was stable at 50% and there was
no changes noticed while the accuracy of training model was 58% approx. initially then
there was a pattern of up and down and at last the final accuracy was more than 60%.

15
 Figure 17: InceptionV3 Model Accuracy Plot

6.3 ResNet50 Evaluation

From the Figure 18, the classification report of ResNet50 shows that the model has
achieved better accuracy of 94% which is highest than the other models. The implemen-
ted model can correctly identify 98% of images with tumour and 90% of images with
no tumour. The ResNet50 model gives better recall than other models as recall plays
a crucial role in medical field.The precision value is the number of correctly classified
predictions, so in this case for labels ‘yes’ which is tumour positive and ’No’ which is
tumour negative the correct classification percentages are 88% and 69%.
In order to achieve a better learning rate, the RMSprop optimizer is used, because the
output is multi-class classification categorical cross-entropy. Using the hyperparameters
adjusted, the epoch value for this model is the same as the previous one. This model has
a batch size of 32 and an epoch value of 10. The given figure illustrates ResNet50 Model
loss throughout the epoch process.

Figure 18: ResNet50 Classification Report

The model losses of ResNet50 through all the epochs have been getting low in Figure 19.
For validation set, there was a significant drop of model loss from 3.5 to 2 at step 1 epoch.
The model loss of validation set was decreasing gradually through out the epoch process.
The initial accuracy of training set was approx. 68% which was fluctuating in an up-

16
 Figure 19: ResNet50 Model Loss Plot

Figure 20: ResNet50 Model Accuracy Plot

17
ward way and at final epoch stage it was more than 90% for training data in Figure 20.
Whereas, accuracy of validation set was less than 80% but showed an upward trend .

6.4 Discussion

Figure 21: Comparing all three models

This study aims to improve the precision and accuracy of brain tumor identification.
It shortens the processing time and increases precision by using Google Colab.Three
transfer learning models such as VGG16, InceptionV3 and ResNet50 were built to detect
a tumor present in a brain or not. In evaluation method, all the models were critically
evaluated individually and explained model performance with the help of different plots.
In this section, Figure 21 gives overview of all the models evaluated in terms of accuracy,
recall and precision. VGG16 gives a total accuracy of 75% which is higher than Incep-
tionV3 Model and also VGG16 model’s precision and recall are better than InceptionV3.
ResNet50 gives the best accuracy of 94% which is better than other two models. Also
precision that is correctly classified tumour and non-tumour images are 91% and 98%
respectively. ResNet50 correctly identifies 98% of tumour presented images and 90% of
tumour not presented images, interpreted by recall parameter of ResNet50. Hence, Res-
Net50 model can detect a tumour present in a brain or not more efficiently and accurately.

Figure 22 illustrates the accuracy level of all three models. From the Figure 22 it is
clearly seen that ResNet50 gives better accuracy whereas VGG16 produces less accuracy
as compared.

7 Conclusion and Future Work

This study employed VGG16,InceptionV3 and ResNet50 to identify brain cancers using
magnetic resonance images (MRI). The suggested method enhances the accuracy by over
90% while also reducing runtime. The study uses multiple techniques such as normaliz-
ation to enhance the quality of image, data augmentation and transfer learning models
such InceptionV3, VGG16 and ResNet50 on brain tumour MR images data set. Com-
pared to earlier transfer learning algorithms, this model has exceptional accuracy and
consistency. This strategy enables the model for fast learning images which will give bet-
ter performance. Thus the study can efficiently and accurately detect tumor in a brain
which helps the health worker to provide better treatment as early as possible.
Patients with brain tumors may benefit from the proposed approach for identifying
tumors.Future research might use the proposed approach to address categorical classifica-

18
 Figure 22: Comparison of Accuracy

tion problems, like distinguishing between Gliomas, Meningiomas, and Pituitary tumors,
or identify other brain disorders. A proposed system may also provide early detection in
other clinical domains involving medical imaging, such as lung cancer and breast cancer,
both of which are highly fatal diseases worldwide.

7.1 Acknowledgement
In the research process, I appreciate the guidance, and invaluable response I received
from Prof. Aaloka Anant and would like to thank him for his guidance.

References
Abbas, K., Khan, P. W., Ahmed, K. T. and Song, W.-C. (2019). Automatic brain tumor
detection in medical imaging using machine learning, 2019 International Conference
on Information and Communication Technology Convergence (ICTC), pp. 531–536.

Bhanothu, Y., Kamalakannan, A. and Rajamanickam, G. (2020). Detection and clas-

sification of brain tumor in mri images using deep convolutional network, 2020 6th
International Conference on Advanced Computing and Communication Systems (IC-
ACCS), pp. 248–252.

Brain tumor detection in MR image using superpixels, principal component analysis and
template based K-means clustering algorithm (2021). Machine Learning with Applica-
tions 5: 100044.

Choudhury, C. L., Mahanty, C., Kumar, R. and Mishra, B. K. (2020). Brain tumor de-
tection and classification using convolutional neural network and deep neural network,

19
 2020 International Conference on Computer Science, Engineering and Applications
(ICCSEA), pp. 1–4.

Grampurohit, S., Shalavadi, V., Dhotargavi, V. R., Kudari, M. and Jolad, S. (2020). Brain
tumor detection using deep learning models, 2020 IEEE India Council International
Subsections Conference (INDISCON), pp. 129–134.

Liu, J., Li, M., Wang, J., Wu, F., Liu, T. and Pan, Y. (2014). A survey of mri-based
brain tumor segmentation methods, Tsinghua science and technology 19(6): 578–595.

Mudgal, T. K., Gupta, A., Jain, S. and Gusain, K. (2017). Automated system for
brain tumour detection and classification using extreme gradient boosted decision trees,
2017 International Conference on Soft Computing and its Engineering Applications
(icSoftComp), pp. 1–6.

Nalepa, J., Marcinkiewicz, M. and Kawulok, M. (2019). Data augmentation for brain-
tumor segmentation: A review, Frontiers in Computational Neuroscience 13: 83.

Noreen, N., Palaniappan, S., Qayyum, A., Ahmad, I., Imran, M. and Shoaib, M. (2020).
A deep learning model based on concatenation approach for the diagnosis of brain
tumor, IEEE Access 8: 55135–55144.

Raut, G., Raut, A., Bhagade, J., Bhagade, J. and Gavhane, S. (2020). Deep learning
approach for brain tumor detection and segmentation, 2020 International Conference
on Convergence to Digital World - Quo Vadis (ICCDW), pp. 1–5.

Sangeetha, R., Mohanarathinam, A., Aravindh, G., Jayachitra, S. and Bhuvaneswari, M.

(2020). Automatic detection of brain tumor using deep learning algorithms, 2020 4th
International Conference on Electronics, Communication and Aerospace Technology
(ICECA), pp. 1–4.

Saxena, P., Maheshwari, A. and Maheshwari, S. (2021). Predictive modeling of brain

tumor: A deep learning approach, in M. K. Sharma, V. S. Dhaka, T. Perumal, N. Dey
and J. M. R. S. Tavares (eds), Innovations in Computational Intelligence and Computer
Vision, Springer Singapore, Singapore, pp. 275–285.

Seetha, J. and Raja, S. S. (2018). Brain tumor classification using convolutional neural
networks, Biomedical & Pharmacology Journal 11(3): 1457.

Sharma, K., Kaur, A. and Gujral, S. (2014). Brain tumor detection based on machine
learning algorithms, International Journal of Computer Applications 103(1): 7–11.

Sharma, M. and Miglani, N. (2020). Automated Brain Tumor Segmentation in MRI

Images Using Deep Learning: Overview, Challenges and Future, Springer International
Publishing, Cham, pp. 347–383.

Wickrama Arachchilage, S. and Izquierdo, E. (2020). Deep-learned faces: a survey,

EURASIP Journal on Image and Video Processing 2020.

Zhang, D., Huang, G., Zhang, Q., Han, J., Han, J. and Yu, Y. (2021). Cross-modality
deep feature learning for brain tumor segmentation, Pattern Recognition 110: 107562.
URL: https://www.sciencedirect.com/science/article/pii/S0031320320303654

20