JOURNAL OF CRITICAL REVIEWS

ISSN- 2394-5125 VOL 7, ISSUE 04, 2020

PREDICTIVE MODELLING OF BRAIN TUMOR

DETECTION USING DEEP LEARNING
Suraj Patil1 , Dr D. K. Kirange2, Varsha Nemade3

1 Computer Engineering Dept., MPSTME,Shirpur,NMIMS University, India

2 Computer Engineering Dept, J.T. Mahajan College of Engineering, Faizpur, India
3 Computer Engineering Dept., MPSTME,Shirpur,NMIMS University, India

Abstract. : In recent years, prediction and analysis of human brain tumor have
become one of the most challenging issues in healthcare science. Various machine
learning algorithms are designed to automate the process of detection of brain tumor.
Because of the popularity of computer vision in AI, the segmentation of tumor in
unstructured data set such as brain MRI and its analysis as become an important part
of the diagnosis of cancer at an early stage. The correct diagnosis is a very crucial and
critical step and depends on the expertise of doctors and radiologists. The deep
learning models are getting a lot of popularity in the detection of tumors because its
accuracy. In this paper, we designed deep learning architectures for detection of
tumors in Magnetic Resonance Imaging (MRI) image. In the proposed architecture,
firstly, the convolution neural network (CNN) architecture was designed from scratch
using Keras library; secondly, the architecture of CNN was tuned by adjusting hyper
parameter and increasing number of layers, and finally the transfer learning
mechanism was implemented by using weights of VGG16 architecture. The
performance of all models was evaluated using confusion matrix on validation and the
test data set. The result shows that adjusting hyper parameter and transfer learning the
accuracy of detection of tumor can be improved. In addition, this deep learning model
detects human brain tumors within seconds as compared to other machine learn- ing
algorithm.
Keywords: Magnetic Resonance Imaging(MRI), hyper parameter, CNN, VGG16,
Ke- ras,transfer learning.

1. Introduction
Artificial Intelligence in the healthcare domain is used to estimate the power of
human cognition to simplify the analysis of complicated medical data by using com-
plex algorithms and decision support systems. As the computing power of medical
data is increasing in terms of velocity, volume, and variability, finding meaningful
insight from medical data has become a challenging task. Health care data come in
structured and unstructured format [1]. The structured data is in the form of textual
information containing different features of specific diseases were as unstructureddata
are in the form of signals and medical images. Because of the popularity of com-
puter vision in AI, the segmentation of tumor in unstructured dataset such as brain
MRI and its analysis as become an important part of the diagnosis of cancer at an
early stage. The correct diagnosis is a very crucial and critical step and depends on the
expertise of doctors and radiologists. In such cases, the computer-aided diagnosis
system is used as the second option for diagnosis [2]. The problem with the traditional
computer-aided system is of false positive and false negative predictions done, con-
cerning the classification of tumor which can be life-threatening. Also,due tothe hete-
rogeneous and diffusive shapes of human organs such a liver, brain, tumor, etc., the
segmentation of these organs as become challenging task because a lot of overlapping
and low clearance ratio is seen between these organs. The physicians are finding it
difficult and thus need a second option to come the final conclusion of treatment ther-
apy for patient before any surgical operation decision. So there is a need to design an
algorithm which can process 2D medical images of CT scan devices or MRI efficient-
ly and classify weather given image contains tumor or not. The motivation is to de-
sign a robust model using deep learning techniques to improve performance of the
proposed model in terms computation processing, overfitting,learning mechanism and
accuracy. The most popular framework of deep learning is a convolution neural net-
1805
 JOURNAL OF CRITICAL REVIEWS
ISSN- 2394-5125 VOL 7, ISSUE 04, 2020

work architecture for medical image analysis. The transfer learning mechanism can be
used to reduce the cost of high computation to train a classifier for medical images
[3].
The organization of the paper is as follows: The second section describes various
approaches used for tumor detection and differences between them. The methodology
of designing deep learning model is given in the third section. The fourth section ex-
plains about experimental setup and dataset used for model building. The fifth section
summarizes results analysis and tuning of parameters of the deep learning model. The
last section highlights the research scope which can be extended as future work.

2. Literature Review
Various machine learning techniques for automatic detection and segmentation of
brain tumor are described in literature and their performance are evaluated to check-

accuracy [4]. Deep learning shows the great performance in the healthcare domain of
medical image analysis such as MRI, CT scan etc [5] and more in image-based cancer
detection and diagnosis [6]. The following table describes the various techniques
applied on images of brain tumor with datasets, techniques used and with their per-
formance measure.

Table1. Literature review Overview

Author Title Dataset Techniques used Performance Journal or Year

used Measure publisher
Pereira S Brain tumor seg- BRATS Deeper CNN with BRATS dice IEEE Trans- 2016
et al. mentation using CNN 2013 LReLu as activa- scores of 88%, action Medi-
in MRI images tion function 83% and 77% cal Imaging
for whole tumor,
core tumor and
active tumor
regions respec-
tively
Lina Chato Machine and deep BRATS SVM,KNN Linear Accracy=68.8% In Proceed- 2017
et.al learning techniques to 2017 Discriminant, De- ings of the
predict overall survival cision Tree, En- 2017 IEEE
of brain tumor patients semble and Logis- 17th Interna-
using MRI images tic Regres- tional Con-
sion,CNN ference on
Bioinformat-
ics and Bio-
engineering
(BIBE),
Washington
Amin J. et. Detection of Brain BRATS Gabor Wavelet DiceScores J. Ambient. 2018
al. Tumor based on Fea- 2012 Features Complete = 0.91 Intell.
tures Fusion and Ma- ,Histogramsof Non-Enhancing
chine Learning Oriented Gradient , = 0.89 Enhanc-
Local Binary Pat- ing = 0.90
tern and segmenta-
tion based Fractal
Texture Analysis
(SFTA) features
-Random forest
calssifier
Abiwinanda Brain tumor classi- figshare CNN,AlexNet,VG Validation InWorld 2018
et. al. fication using convolu- Cheng G16,ResNet Accuracy Conress on
tional neural network (Brain 84.19% In medical
Tumor physics and

1806
 JOURNAL OF CRITICAL REVIEWS
ISSN- 2394-5125 VOL 7, ISSUE 04, 2020

Dataset, Bio medical

2017) Engineer-
ing;Springer.
Deepak Brain tumor classifica- figshare Transferred Accuracy 98% Com- 2019
et.al. tion with deep CNN learned (Google puters in
features through trans- Net),deep CNN- biology and
fer learning. SVM medicine

Talo et.al Application of deep Dataset pre-trained Accuracy on 613 Cognitive 2019
transfer learning for available CNN ResNet34 images is 100% Systems
automating brain ab- Harvard Research
normality classifica- Medical
tion in MRI images. School
website

S Kumar et.al has described the hybrid architecture for detection of tumor, in which
features are extracted using DWT and a genetic algorithm is used for reducing the
numbers of features and finally SVM is applied for classification [13].Dong at.al
apply fully automatic segmentation on brain tumor by using U Net Deep learning
segmentation on BRATS 2015 datsaet [15]. Nasor et. al. has described the approach
by using machine learning techniques for detection of brain tumor at an early stage
using a combination of different techniques such as k-means, patch based processing
,object counting and finally tumor evaluation and they got accuracy 0.99 and dice
score 0.95[16].

3. Methodology:
The supervised classification modeling is done by designing deep learning CNN
architecture and transfer learning mechanism.
Design CNN model for tumor detection:
A convolution neural network is an artificial neural network, which uses convolu-
tion tricks to add convolution layers. We use these convolution tricks to preserve the
spatial structure in images which helps in classifying tumor. Initially the network was
designed with one convolution layer with 32 filters and kernel of size 2 as shown in
figure1. The max pooling was done with stride equal to 2 to preserve the spatial

property of brain tumor and 128 neurons were used in fully connected layers for final
prediction.

Fig1: CNN model

Since we have been using GPU based system with augmented data set, we mod-
ified the above network by adding extra dense layers and adjusted hyper parameters
as shown in figure2. The network consist of two layers with convolution filter size of
32 and 64 followed by two fully connected layers with adjusted hyper parameters.

1807
 JOURNAL OF CRITICAL REVIEWS
ISSN- 2394-5125 VOL 7, ISSUE 04, 2020

Fig2: CNN model with adjusted hyper parameter

The activation output volume size after applying the convolution operation is given
by equation 1.

𝑁 [𝑙] = 𝑁[𝑙 ] + 2𝑃[𝑙] − 𝑓 𝑙−1 +1 ....................(Eq 1)

𝑤ℎ 𝑤ℎ
𝑆[𝑙]
Where
𝑁[𝑙]= Height of input image
ℎ
𝑁[𝑙]=Width of input Image
𝑤
𝑁𝑐[𝑙]=Number of filters

𝑓𝑙 =Filter Size of Kernel

𝑃𝑙 =Padding Size
𝑆𝑙=Stride Size
The equation1 gives the linear output by applying the convolution operation, the
non-linearity is introduced by applying Relu function to the output volume by adding
some bias 𝑏[𝑙]. Therfore, net activation size for next layer is given by equation 2
𝑖
𝑎[𝑙]=Relu(𝑁[𝑙] + 𝑏 [𝑙 ]) ................. (2)
𝑤ℎ 𝑖
Transfer Learning

To improve performance of the deep learning model, the transfer learning tech-
nique can be used. In transfer learning pre-trained models are used to build model
instead of designing them from scratch. Here we have used VGG16 architecture
weights for training the model. In this architecture we have frozen higher layers and
trained lower layer using weights of VGG16 architecture. The VGG16 architecture
contains 16 layers of which pre-trained weights are used for training. The VGG16
architecture is shown in figure3.

Fig3: CNN model with VGG16 transfer learning mechanism

So one solution to minimize high computation cost is to use deep learning transfer
learning mechanism. The transfer learning mechanism uses pertained weights of CNN
such as VGG16 and can save time and computational power as well as speed up the
learning process. The VGG16 is used as a base model to train early layers of architec-
1808
 JOURNAL OF CRITICAL REVIEWS
ISSN- 2394-5125 VOL 7, ISSUE 04, 2020

ture and only last 4 layers are added during the training process as shown in figure3.
The activation units of this model are flattened, which acts as input to rest of layers
ofthe model. The dense layer of 128 units is used to take dot product with units of
flat- tened layer and generates final activation units by applying ReLu function to
gene- rates final activation on which sigmoid function is applied to get binary output
of tumor classification The table 2 and table 3 shows the summary of activations units
of two deep learning architectures of figure2 and figure3.

Table2.CNN model with adjusted hyper parameter

The table3 below shows number of parameters of VGG16 model and layers used
for training the sample size using transfer learning mechanism.

Table3.CNN model with transfer learning mechanism

1809
 JOURNAL OF CRITICAL REVIEWS
ISSN- 2394-5125 VOL 7, ISSUE 04, 2020

4. Experimental Setup:

The proposed model is designed and trained on NVIDIA K80 GPU. The model tested
and evaluated on brain MRI image dataset [14]. The brain MRI dataset contains 253
MRI images out of which 153 images are tumors and 98 are non tumors. The images
are resized to dimension 224 * 224 to fit to the VGG16 model for training. Since we
are using transfer learning mechanism to optimize the CNN model, the data augmen-
tation technique is used to increase size of the data set. After data augmentation tech-
nique, the tumorous images are 1095 and non-tumorous images are 980. The annota-
tion of the tumorous and non-tumorous images is done and is stored in the form of
numpy array. The CNN model with adjusted hyper parameter and CNN model with
transfer learning are trained for 25 epochs with batch size of 32 and accuracy and loss
is recorded. The batches are normalized to reduce overfitting and covariance shift.
The ground truth of validation dataset was done with MRI images from the radiolo-
gist.

5. Result Analysis and Discussion

The two different CNN models are trained separately and the performance was re-
corded. The CNN model architecture shown in figure 2 with adjusted hyper parameter
trained for 25 epochs. The layer1 of this model uses the kernel of size 7*7 with 32
filters and layer2 uses kernel of size 5*5 and 64 filters. The max pooling of size 4*4
was applied in two layers to preserve spatial properties of tumors. Finally, all neurons
coming from layer2 are flattened into single vector, which acts as an input to artificial
neural network. The trainable parameters are 58,369 and non-trainable parameters are
64 out of total parameters 58,433 in the network. The proposed architecture improves
accuracy of CNN architecture shown in figure1 from 72% to 80%. The accuracy and
loss plots are shown in figure. From plot is seen that model with hyper parameter
adjustment improves the accuracy both in validation and testing dataset. There is
slightly peak is seen in loss during training because of small sample of test dataset
which can be adjusted by increasing size of dataset, but as epoch increases the loss
decreases and it is in line with validation loss.

Fig4: CNN model with hyper paramter (a) ModelAccuracy (b) Model Loss
CNN model with Transfer Learning(c) Model Accuracy (d) Model Loss

1810
 JOURNAL OF CRITICAL REVIEWS
ISSN- 2394-5125 VOL 7, ISSUE 04, 2020

The model is further improved by transfer learning mechanism. We have used

weights of VGG16 architecture as base model by keeping last few layers intact with
kernel size 7*7 and 512 filters. The total trainable parameters are 3,211,521 parame-
ters out of total parameters 17,926,209.It is observed during experiment that by add-
ing more dense layers, the performance of the architecture shown in figure3 is im-
proved. The performance of the CNN architecture with adjusted parameter
andVGG16 CNN transfer learning model is evaluated using confusion matrix. The
performance of the model is evaluated separately on validation and testing dataset
accuracy is plotted as shown in figure5.
The accuracy of the CNN model with adjusted parameters for validation is
80% and for testing it comes to be 89% Fig (a)(b). By applying transfer learning
mechanism there is considerable improvement in the accuracy of the CNN model
with 88% for validation and 94% for testing. Also the F1 Score of CNN model is .80
and that of CNN using transfer learning model is .85. So by transfer learning mecha-
nism the performance of model has improved.

Fig5: Confusion matrix for CNN model with hyper paramter

(a)Validation Accuracy (b) Test Accuracy
Confusion matrix for CNN model with Transfer Learning
(c)Validation Accuracy (d) Test Accuracy

1805
1811
 JOURNAL OF CRITICAL REVIEWS
ISSN- 2394-5125 VOL 7, ISSUE 04, 2020

6. Conclusion and Future Work

.
In this paper, we have proposed different styles of CNN architectures and compared
their performance for brain tumor detection. First we started with very simple archi-
tecture and recorded its accuracy and then the model is tuned by adjusting hyper pa-
rameter and increasing number of filters and layers. The results show that adjustment
of hyper parameters increases the accuracy of CNN model. Further model accuracy
was increased by using VGG16 has base model and keeping other layers of the model
has intact. The use of transfer learning mechanism shows significant improvement in
accuracy and F1 score of tumor detection model. The future work will be to design a
CNN model for 3D brain MRI images to get geometrical spatial properties of tumor.
Also further research work can be extended for selecting optimal weights from the set
of ensemble architectures, so that learning and prediction time can be reduced. The
different approaches of hybrid architectures such as cascading of different CNN archi-
tectures, ensemble of CNN with other machine learning models can be explored to
improve the accuracy of deep learning models.

References

1. D. L. Pham, C. Xu, and J. L. Princem, ―Current methods in medical image

segmenta- tion,‖ Annual Review of Biomedical Engineering, vol. 2, no. 1, pp.
315–337, 2000.
2. Hong, Liang, et al. "Big data in health care: Applications and challenges." Data
and In- formation Management 1.ahead-of-print (2018).
3. Lakshmi, A., and T. Arivoli. "COMPUTER AIDED DIAGNOSIS SYSTEM
FOR BRAIN TUMOR DETECTION AND SEGMENTATION." Journal of
Theoretical & Applied In- formation Technology 64.2 (2014).
4. Lundervold, Alexander Selvikvåg, and Arvid Lundervold. "An overview of deep
learning in medical imaging focusing on MRI." Zeitschrift für Medizinische
Physik 29.2 (2019): 102-127.
5. Shen, Dinggang, Guorong Wu, and Heung-Il Suk. "Deep learning in medical
image analy- sis." Annual review of biomedical engineering 19 (2017): 221-248.
6. Hu, Z.; Tang, J.; Wang, Z.; Zhang, K.; Zhang, L.; Sun, Q. Deep Learning for
Image-based Cancer Detection and Diagnosis—A Survey. Pattern Recognit.
2018
7. Pereira S, Pinto A, Alves V, Silva C A. Brain tumor segmentation using
convolutional neural networks in MRI images. IEEE Trans Med Imaging
2016;35(5):1240–1251.
8. Chato, L.; Latifi, S. ,‖Machine learning and deep learning techniques to predict
overall survival of brain tumor patients using MRI images‖, In Proceedings of the
2017 IEEE 17th
International Conference on Bioinformatics and Bioengineering (BIBE),
Washington, DC, USA, 23–25 October 2017; pp. 9–14
9. Amin, J.; Sharif, M.; Raza, M.; Yasmin, M. ,―Detection of Brain Tumor based on
Features Fusion and Machine Learning‖, J. Ambient. Intell. Humaniz. Comput.
2018
10. Abiwinanda, N.; Hanif, M.; Hesaputra, S.T.; Handayani, A.; Mengko, T.R. Brain
tumor classification using convolutional neural network. In World Congress on
Medical Physics and Biomedical Engineering; Springer: Berlin, Germany, 2018;
pp. 183–189.
11. Deepak, S., & Ameer, P. M. (2019). Brain tumor classification using deep CNN
features via transfer learning. Computers in biology and medicine, 111, 103345.
12. Talo, M., Baloglu, U. B., Yıldırım, Ö.,& Acharya, U. R. (2019). Application of
deep trans- fer learning for automated brain abnormality classification using MR
images. Cognitive Systems Research, 54, 176-188.
13. S Kumar, Sanjeev, Chetna Dabas, and Sunila Godara. "Classification of brain
MRI tumor images: a hybrid approach." Procedia computer science 122 (2017):
1806
1812
 JOURNAL OF CRITICAL REVIEWS
ISSN- 2394-5125 VOL 7, ISSUE 04, 2020

510-517.
14. https://www.kaggle.com/navoneel/brain-mri-images-for-brain-tumor-detection,22
May 2020.
15. Dong, Hao, et al. "Automatic brain tumor detection and segmentation using u-net
based fully convolutional networks." annual conference on medical image
understanding and analysis. Springer, Cham, 2017.
16. Nasor, Mohamed, and Walid Obaid. "Detection and Localization of Early-Stage
Multiple Brain Tumors Using a Hybrid Technique of Patch-Based Processing, k-
means Clustering and Object Counting." International Journal of Biomedical
Imaging 2020 (2020).

1807
1813