Brain Tumor Classification Using Convolutional Neural Network

Abstract

Misdiagnosis of brain tumor types will prevent effective response to medical intervention and decrease the chance of survival among patients. One conventional method to differentiate brain tumors is by inspecting the MRI images of the patient’s brain. For large amount of data and different specific types of brain tumors, this method is time consuming and prone to human errors. In this study, we attempted to train a Convolutional Neural Network (CNN) to recognize the three most common types of brain tumors, i.e. the Glioma, Meningioma, and Pituitary. We implemented the simplest possible architecture of CNN; i.e. one each of convolution, max-pooling, and flattening layers, followed by a full connection from one hidden layer. The CNN was trained on a brain tumor dataset consisting of 3064 T-1 weighted CE-MRI images publicly available via figshare Cheng (Brain Tumor Dataset, 2017 [1]). Using our simple architecture and without any prior region-based segmentation, we could achieve a training accuracy of 98.51% and validation accuracy of 84.19% at best. These figures are comparable to the performance of more complicated region-based segmentation algorithms, which accuracies ranged between 71.39 and 94.68% on identical dataset Cheng (Brain Tumor Dataset, 2017 [1], Cheng et al. (PLoS One 11, 2017 [2]).

Full text

Brain Tumor Classification Using
Convolutional Neural Network
Nyoman Abiwinanda, Muhammad Hanif, S. Tafwida Hesaputra,
Astri Handayani, and Tati Rajab Mengko
Abstract
Misdiagnosis of brain tumor types will prevent effective
response to medical intervention and decrease the chance
of survival among patients. One conventional method to
differentiate brain tumors is by inspecting the MRI images
of the patient’s brain. For large amount of data and
different specific types of brain tumors, this method is
time consuming and prone to human errors. In this study,
we attempted to train a Convolutional Neural Network
(CNN) to recognize the three most common types of brain
tumors, i.e. the Glioma, Meningioma, and Pituitary. We
implemented the simplest possible architecture of CNN;
i.e. one each of convolution, max-pooling, and flattening
layers, followed by a full connection from one hidden
layer. The CNN was trained on a brain tumor dataset
consisting of 3064 T-1 weighted CE-MRI images pub-
licly available via figshare Cheng (Brain Tumor Dataset,
2017 [1]). Using our simple architecture and without any
prior region-based segmentation, we could achieve a
training accuracy of 98.51% and validation accuracy of
84.19% at best. These figures are comparable to the
performance of more complicated region-based segmen-
tation algorithms, which accuracies ranged between 71.39
and 94.68% on identical dataset Cheng (Brain Tumor
Dataset, 2017 [1], Cheng et al. (PLoS One 11, 2017 [2]).
Keywords
Training loss Training accuracy Validation loss
Validation accuracy Overfitting
1 Introduction
On 2016, brain tumor was the leading cause of
cancer-related death in children (ages 0–14) in the United
States and ranked above Leukemia [3]. Brain and CNS
tumors are also the third most common cancer occurring
among teenager and adolescents (ages 15–39) [4]. Different
types of brain tumors require different medical interventions.
In conventional computer-aided diagnosis systems, the
tumor mass itself has to be identified and segmented before
it can be classified into different types. Upon tumor mass
segmentation, the segmented region is then subjected to
feature extraction and classification.
Recent studies of identification and segmentation of brain
tumor [5,6] found no universal system for accurate tumor
detection system regardless of its location, shape, and
intensity [6]. There are numerous proposed algorithms in
recent studies for feature extraction and classification of
brain tumors. Grey-level co-occurrence matrix (GLCM) [7–
9] is commonly used for extraction of low-level features.
Several other feature extraction algorithms which attempt to
handle the complex texture of brain tumor are Neural Net-
work [9,10], Bag-of-Words (BoW) [2,8], and Fisher Vector
[2]. One recent study showed that by using a combination of
adaptive spatial pooling and fisher vector algorithm, brain
tumor classification into Glioma, Meningioma, and Pituitary
can be achieved with 71.39–94.68% accuracy [2].
Conventional brain tumor classification methods com-
monly involve region-based tumor segmentation prior to
feature extraction and classification. In this paper, we propose
an automatic brain tumor segmentation/classification method
based on Convolutional Neural Networks. CNN consists of a
convolutional network to perform automatic segmentation
N. Abiwinanda (&)M. Hanif S. T. Hesaputra A. Handayani
T. R. Mengko
Bandung Institute of Technology, Bandung, West Java 40134,
Indonesia
e-mail: abiwinanda@outlook.com
M. Hanif
e-mail: mhaniffarhat@gmail.com
S. T. Hesaputra
e-mail: tafwidahesaputra@gmail.com
A. Handayani
e-mail: a.handayani@stei.itb.ac.id
T. R. Mengko
e-mail: tmengko@stei.itb.ac.id
©Springer Nature Singapore Pte Ltd. 2019
L. Lhotska et al. (eds.), World Congress on Medical Physics and Biomedical Engineering 2018,
IFMBE Proceedings 68/1, https://doi.org/10.1007/978-981-10-9035-6_33
183

and feature extraction, followed by a conventional neural
network to perform classification task. The well-known basic
architecture of CNN involves a Rectified Linear Unit (ReLu),
a convolution, and a pooling layer [11]. In contrast to con-
ventional methods which require prior segmentation of tumor
mass, our CNN approach does not involve region-based
pre-processing step. We validate our algorithm on the same
dataset which was used in previous publications [1,2].
2 Previous Work
Various methodologies have been developed in the past years
to recognize brain tumor in MRI images. These methods are
ranging from classical image processing to neural network
based machine learning approach. Jun Cheng et al. [2]
developed a tumor classification method that consists of two
phases: offline database building and online retrieval. In the
offline database phase, the brain tumor images are processed in
sequential steps. The steps are consisted of tumor segmenta-
tion, feature extraction, and distance metric learning. In the
online learning, the input brain image will be processed sim-
ilarly and compare the extracted feature with the learned dis-
tanced metrics which are stored in the online database. This
method does not use neural network approach but could
achieve a classification accuracy of 94.68%. On the other
hand, Gawande and Mendre [12] used Deep Neural Network
using autoencoders in order to classify the brain tumor. Image
segmentation and feature extraction had been implemented on
the image before it was processed by DNN layers. The texture
and intensity based features of the image were extracted with
help of Gray Level Co-occurence Matrix (GLCM) and Dis-
crete Wavelet Transform (DWT). In the final step, DNN layers
which consist of two autoencoders and one softmax layer were
performed for classification. Furthermore, Pereira et al. [13]
also exploring the used of Convolutional Neural Networks
(CNN) with small 3 3 kernels in order to get to the deeper
architecture and avoid the overfitting. They also investigated
the use of intensity normalization as the pre-processing step
before getting into the CNN layers. In this study, inspired by
those works, we investigate and explore the implementation of
deep CNN on classifying several brain tumor type diagnosis
problems in order to get better accuracy result.
3 Method
3.1 Convolutional Neural Network
CNN convolution layer is a network where an image will be
convolved with filters to produce feature maps. This feature
maps will be forwarded to the next convolution layer to
receive or extract another higher level features from the input
image. Between convolution layers, non-linearity functions
and down sampling operation is used to add non-linearity and
reduced the dimensionality of the image respectively. Max-
pooling usually used as the down sampling operation as it
reduced the dimension while preserving the dominant feature
in the feature maps. Just after the last convolution layer or
before the first layer of the neural network, flattening layer
exist to vectorize the feature maps. In the neural network or
classification phase, the flatten input vector will be forwarded
into the network to produce a number at each output neurons.
This number tells how much an input vector is classified as a
certain class. Usually a softmax activation function is used at
the output layer to normalize the output sum such that all
numbers at the output neuron will add up to one.
In the training phase, CNN use a learning or optimizer
algorithm to update the filters at the convolution layers and
weights at the neural network or fully connected layer. The
learning algorithm takes a classification error or loss as an
input and back propagates the error into the network to
update the filters and the weights.
3.2 CNN Architecture
In this paper, we use five different architectures to test the
accuracy of brain tumor classification. There are already well
defined CNN architecture such as AlexNet [14], VGG16
[15], and ResNet [16] but the architecture that are imple-
mented in this paper are much simpler that the one men-
tioned. The CNN architecture that are implemented in this
paper are summarize in the Fig. 1a–e.
3.3 Hyper-parameters Optimization
Hyper-parameter is a parameter in deep learning process
whose value can be set and tuned before the learning pro-
cess. This parameter will determine the algorithm to be used
in the learning process. Different model training algorithms
need different hyper-parameters that also affecting the result
of the learning process too. Hyper-parameter optimizer has
to be chosen and tuned so that the classifier will have the
most optimal way to solve the problem.
In this study, we will use ‘adam’optimizer in the learning
process which is a method for stochastic optimization by
utilizing the stochastic gradient descent principle. ‘Adam’
optimizer which stands for adaptive moment estimation is
chosen because of it’s advantage that can handle sparse
gradients on noisy problems.
184 N. Abiwinanda et al.

4 Data
In this paper, our CNN is trained using 3064 T-1 weighted
CE-MRI of brain tumor images. The dataset is provided by
Jun Cheng and was previously used in his paper [1,2]. This
dataset consists of 708 images with glioma, 1426 images with
meningioma, and 930 images with pituitary tumors. In our
training phase, we equalize the amount of images that are
used to train the CNN for each class or type of tumors. Out of
all available images, we only used 700 images from each
class where 500 of those images were used for training phase
and the other 200 images were used for validation phase. The
dataset was originally provided in matlab.mat format where
each file stores a struct containing a label which specify the
type of tumor for a particular brain image, patient ID, image
data in 512 512 uint16 format, vector storing the
coordinates of discrete points on tumor border, and a binary
mask image with 1 s indicating tumor region. In our paper we
only make use the label and image data in the.mat files
therefore our brain tumor classifier is a simple CNN network
which only takes image as an input. Figure 2a–c represent
example of the dataset from each of the classes.
5 Result
In our experiment, the hyperparameter at each layer such as
number and size of filters in the convolution layers, size of
maxpooling kernel, number of neurons in the fully con-
nected layers are held fixed. Only the depth of the archi-
tecture are varied between different architectures. The
architecture and hyperparameter that are used are served in
Table 1.
Fig. 1 a–eThe proposed
Convolutional Neural Network
Architecture 1–5
(a) (b) (c)
Fig. 2 aGlioma, bMeningioma,
cPituitary (each labeled in green)
Brain Tumor Classification Using Convolutional Neural Network 185

Sizes of the input images that are forwarded into the
network are 64 64. The original images are in the size of
512 512. This reduction is performed because of com-
putational cost reason. All architectures are compiled with-
out a GPU therefore to speed up the training phase smaller
image size is used.
All convolution layers in the architectures use 32 filters of
size 3 3. We use ReLu as our activation function as it
already the standard activation function used in image
classification task. The size of the maxpool kernel is 2 2
and all the fully connected layer (called ‘dense’in keras) use
64 neurons.
Finally, there are 3 neurons in the output layer since we
try to classify an image with three types of brain tumors
(glioma, meningioma, and pituitary). The activation func-
tions that are used at the output layer are softmax so that all
three output neurons are summed up to one.
Based on Table 1, each of architecture produces different
numbers of params and features depending on the depth of
the convolution layer and the fully connected network. The
Table 1 Architecture 1–5
Architecture 1 Architecture 2
Architecture 3 Architecture 4
Architecture 5
186 N. Abiwinanda et al.

architecture with deeper layers of convolution will have
fewer numbers of the trainable params.
Implementation of the above architectures will produce
four parameter values that will describe the success of the
classifier model in classifying the input image. The four
parameter values are loss and accuracy from the training set
and validation set. Accuracy is defined as the percentage of
the correct guesses by the classifier either for the training set
input or the validation set input. Loss is defined as feasible
error that represents the price paid for inaccuracy of pre-
dictions in classification problem. We use cross-entropy
method for loss calculation. The cross-entropy loss calcu-
lation can be represented in the mathematical form [17]
below:
Hðy;b
yÞ¼X
i
yilog 1
b
yi
¼X
i
yilog b
yið1Þ
yirepresents the result of the classifier output from class i.
While b
yirepresents the expected output from class i. The
classification accuracy of each architecture are presented in
Figs. 3,4,5,6and 7.
Based on Figs. 3,4,5,6and 7, the values of loss and
accuracy vary according to the implementation of the
architecture. The classifier model is said to be ‘good fit’if
the accuracy of training set and validation set tend to
increase for every epoch of training. However, if the accu-
racy of validation set tends to decrease while the accuracy of
training set increases, then the classifier model is estimated
to have overfitting. Overfitting happens when the model
learns the detail and noise in the training data thus reducing
its ability to generalize other datasets well [18].
By looking at the result of each architectures, we could
see that all architecture’s validation loss shows an increasing
trend with respect to the number of epoch except for
architecture 2. This indicates that architecture 2 is the best
architecture out of the five architectures at generalizing the
brain tumor images. The decreasing pattern in the validation
loss indicates that using the available training images,
architecture 2 could classify the unknown images in the
Fig. 3 Accuracy and loss of architecture 1
Fig. 4 Accuracy and loss of architecture 2
Fig. 5 Accuracy and loss of architecture 3
Fig. 6 Accuracy and loss of architecture 4
Brain Tumor Classification Using Convolutional Neural Network 187

validation set with medium performance. We conclude a
medium performance since the validation loss does not show
a perfect decreasing pattern as the number of epoch increase.
In the last epoch, architecture 2 could achieve a validation
accuracy of 84.19%.
Based on the best architectural choice on the previous
part, we try to vary the number of the filter in the convo-
lution layer into 64 filters and 128 filters we called it as
architecture 6 and architecture 7 respectively. In this
experiment, we will identify the effect of the filter numbers
in the convolution layer on the accuracy of the validation set.
The architecture and hyper parameters that are used are
served in the Table 2.
After implementation of those architectures, the classifi-
cation accuracy of those architectures are presented in
Figs. 8and 9.
From the result, the architecture that has the highest
validation accuracy is still architecture 2 with 32 filters in the
convolution layers. Although an increase in the number of
filters in the convolution layer does not necessarily con-
tribute to better CNN performance, in our case the number of
filter in a convolution layer does influence the accuracy of
the classifier. We stick or recommend the use of architecture
2 to classify a brain tumor since it has the highest validation
accuracy.
Fig. 7 Accuracy and loss of architecture 5
Table 2 Architecture 6–7
Architecture 6 Architecture 7
Fig. 8 Accuracy and loss of architecture 6 Fig. 9 Accuracy and loss of architecture 7
188 N. Abiwinanda et al.

6 Conclusion
In this paper, we introduced CNN to automatically classify
the three most common types of brain tumor, i.e. the Glioma,
Meningioma, and Pituitary; without requiring region-based
pre-processing steps. We identified an optimal CNN archi-
tecture (architecture 2) consisting of 2 layers of convolution,
activation (ReLu), and maxpool, followed by one hidden
layer of 64 neurons. Architecture 2 is the only architecture
that show a consistently decreasing pattern in the validation
loss as the number of epoch increases, leading to the highest
validation accuracy out of all five architectures. The training
and validation accuracies of architecture 2 at best is 98.51%
and 84.19%, respectively. These figures, although somewhat
lower, are still comparable to the accuracies of conventional
algorithms with region-based pre-processing, which per-
formed at 71.39–94.68% [1,2]. For future work, we con-
sider to include color balancing step into our CNN, to
improve classification accuracy of textured brain MRI pixels
[19]. Our algorithm may be implemented in as a simple
supportive tool for medical doctor in classifying brain tumor.
Disclosure The authors declare that they have no conflict of interest.
References
1. Jun Cheng. (2017). Brain Tumor Dataset (Version 5). Retrieved
from https://doi.org/10.6084/m9.figshare.1512427.v5
2. Cheng, J., Huang, W., Cao, S., et al. (2015). Classification via
Tumor Region Augmentation and Partition. PLoS One,10(10).
3. Varade, A. A., Ingle, K. S. (2017). Brain MRI Classification
Using PNN and Segmentation Using K Means Clustering,
International Journal of Advanced Research in Electrical, Elec-
tronics and Instrumentation Engineering,6(5), 6181–6188.
4. Cheng, J., Yang, W., Huang, M., et al. (2016). Retrieval of Brain
Tumors by Adaptive Spatial Pooling and Fisher Vector Represen-
tation. PLoS One,11(6).
5. Lavanyadevi, R., Machakowsalya, M., Nivehitha, J., et al. (2017),
Brain tumor classification and segmentation in MRI images using
PNN, IEEE Xplore.
6. Siegel, R. L., Miller, K. D., Jemal, A. (2016). Cancer statistics,
2016. CA: A Cancer Journal for Clinicans,66(1), 07–30.
7. Quinn, T., Ostrom, M.A., Gittleman, H., et al. (2016). American
Brain Tumor Association Adolescent and Young Adult Primary
Brain and Central Nervous System Tumors Diagnosed in the
United States in 2008–2012. Neuro-Oncology,18(1), i1–i50.
8. Gordillo, N., Montseney, E., Sobrevilla, P. (2013). State of the art
survey on MRI brain tumor segmentation. Magnetic Resonance
Imaging,31(8), 1426–1438.
9. Kaur, H., Sharma, R. (2016). A Survey on Techniques for Brain
Tumor Segmentation from MRI. IOSR Journal of Electronics and
Communication Engineering,11(5), 01–05.
10. Hasan, M. H., Meziane, F., et al. (2016). Segmentation of Brain
Tumors in MRI Images Using Three-Dimensional Active Contour
without Edge. Symmetry,8(11), 132.
11. Wu, J. (2017). Introduction to Convolutional Neural Networks.
National Key Lab for Novel Software Technology Nanjing
University, China.
12. He, K., Zhang, X., Ren, S., et al. (2015). Deep residual learning for
Image Recognition. CVPR, 770–778.
13. Browniee, J. (2016, March 21). Overfitting and Underfitting With
Machine Learning Algorithms. Retrieved from https://
machinelearningmastery.com/overfitting-and-underfitting-with-
machine-learning-algorithms/.
14. Zacharaki, E. I., Wang, Sumei., Chawla, S., et al. (2009).
Classification of Brain Tumor Type and Grade Using MRI Texture
and Shape in a Machine Learning Scheme.Magnetic Resonance in
Medicine, 62(6), 1609–1618.
15. Kingma, D. P., Ba, J. (2014). Adam: A Method for Stochastic
Optimization, arXiv preprint arXiv:1412.6980.
16. Gawande, S. S., Mendre, V. (2017). Brain Tumor Diagnosis Using
Deep Neural Network (DNN). International Journal of Advanced
Research in Electrical, Electronics and Instrumentation Engineer-
ing, 5(5).
17. Simonyan, K., Zisserman, A. (2015). Very Deep Convolution
Networks for Large-Scale Image Recognition, ICLR 2015.
18. Pereira, S., Pinto, A., Alves, V., et al. (2016). Brain Tumor
Segmentation Using Convolutional Neural Networks in MRI
Images. IEEE Transactions on Medical Imaging, 35(5), 1240–1251.
19. DiPietro, Rob. (2016, May 2). A Friendly Introduction to
Cross-Entropy Loss. Retrieved from https://rdipietro.github.io/
friendly-intro-to-cross-entropy-loss/.
Brain Tumor Classification Using Convolutional Neural Network 189