9th International Conference on Computer and Knowledge Engineering (ICCKE 2019), October 24-25 2019, Ferdowsi University of

Mashhad

Brain Tumor Detection Using Deep Neural Network

and Machine Learning Algorithm
Masoumeh Siar Mohammad Teshnehlab
Department of Computer Engineering Department of Electrical Engineering
Science and Research Branch, Islamic Azad University K.N. Toosi University of Technology
Tehran, Iran Tehran, Iran
[email protected] [email protected]

Abstract—Brain tumor can be classified into two types: be- to the size, shape, location and type of tumor in the brain.
nign and malignant. Timely and prompt disease detection and Diagnosis of brain tumors in the early stages of the tumor’s
treatment plan leads to improved quality of life and increased start is difficult because it cannot accurately measure the size
life expectancy in these patients. One of the most practical and
important methods is to use Deep Neural Network (DNN). In and resolution of the tumor [2].
this paper, a Convolutional Neural Network (CNN) has been However, if the tumor is diagnosed and treated early in the
used to detect a tumor through brain Magnetic Resonance tumor formation process, the chance of patient’s treatment is
Imaging (MRI) images. Images were first applied to the CNN. very high. Therefore, the treatment of tumor depends on the
The accuracy of Softmax Fully Connected layer used to classify timely diagnosis of the tumor [3]. The diagnosis is usually
images obtained 98.67%. Also, the accuracy of the CNN is
obtained with the Radial Basis Function (RBF) classifier 97.34% done by a medical examination, with computer tomography
and the Decision Tree (DT) classifier, is 94.24%. In addition to or magnetic imaging. MRI imaging is a method that provides
the accuracy criterion, we use the benchmarks of Sensitivity, accurate images of the brain and is one of the most common
Specificity and Precision evaluate network performance. Accord- and important methods for diagnosing and evaluating the
ing to the results obtained from the categorizers, the Softmax patient’s brain. In the field of Medical Detection Systems
classifier has the best accuracy in the CNN according to the
results obtained from network accuracy on the image testing. This (MDS), MRI images provide better results than other imaging
is a new method based on the combination of feature extraction techniques such as Computed Tomography (CT), due to their
techniques with the CNN for tumor detection from brain images. higher contrast in soft tissue in humans [4].
The method proposed accuracy 99.12% on the test data. Due The proposed technique has used CNN to identify and
to the importance of the diagnosis given by the physician, the categorize the tumor from brain images of the brain. The
accuracy of the doctors help in diagnosing the tumor and treating
the patient increased. main difference between the main channel of the neural
network with the normal neural network is that it is able to
Index Terms—Brain tumor, deep neural network, Convo- automatically and locally extract the feature from each image
lutional neural network, magnetic resonance imaging, feature
extraction. [5]. These types of networks consist of neurons with weights
and biases that can be learned [6].
I. I NTRODUCTION Due to the results of CNN on the dataset, in order to
Brain tumors can be classified into two types: benign (non- improve the proposed method. Machine learning algorithm
cancerous) and malignant (cancerous). The malignant tumors is used to feature extraction. The algorithm used was the
can quickly spread to other tissues in the brain and lead clustering algorithm applied on data set, and then the images
to worsening the patient’s condition [1]. When most of the are applied to the CNN. The results showed that the proposed
cells are old or damaged, they are destroyed and replaced method has been successful. The purpose of extracting the
by new cells. if damaged and old cells are not eliminated property before applying to the CNN is that in some images
with generating the new cells, it can cause problems. The fatty masses are considered as tumors, or in some images
production of additional cells often results in the formation the tumor is mistakenly considered to be fat and should have
of a mass of tissue, which refers to the growth or tumor. increased medical error. Extracting the attribute initially and
Brain tumor detection is very complicated and difficult due before applying the CNN leads to improved network accuracy
and increased accuracy.
978-1-7281-0673-1/19/$31.00 ©2019 IEEE

363
 II. R ELATED WORK processing, which are used to distinguish the desired segments
or the shape (features) of a digital image or video stream.
In [7], an automated method is used to identify and cat-
egorize MRI images. This method is based on the Super
Pixel Technique and the classification of each Super Pixel. IV. D EEP LEARNING
Extremely randomized trees (ERT) classifier is compared with Deep learning is one of the recent useful kinds of machine
SVM to classify each super pixel into tumor and normal. This learning. In other words, learning is called deep-seated ar-
method has two datasets, which are 19 MRI FLAIR images chitecture. These architectures are in fact the same old nerve
and BRATS 2012 dataset. The results demonstrate the good networks that have become DNN. These networks are data-
performance of this method using ERT classifier. driven and feature engineering is done automatically and we
In [8], an automatic classification method is used to identify do not interfere with it, and this is precisely what makes
a tumor using a CNN with 3 × 3 small kernels. The method the accuracy and excellent performance of these networks in
obtained simultaneously the first position for the complete, different areas. It is in fact a deep learning of a set of nerve-
core, and enhancing regions in dice similarity, coefficient based techniques that learns features automatically from our
metric (0.88, 0.83, 0.77), at the BRATS Challenge 2013. In own input data [16].
[9], Alexnet model CNN is used to simultaneously diagnose
MS and normal tumors. The CNN was able to accurately A. Convolution Neural Network
classify 98.67% images correctly into three classes. In [10], a
The CNN are a special type of DNN whose structure is
multi-stage Fuzzy C-Means (FCM) framework was proposed
inspired by biology of cat’s vision cortex [17]. The CNN has
to segment brain tumors from MRI images.
a hierarchical structure and consists of several layers. CNN
In [11], An efficient and effective method which uses also includes, input layer, output layer, convolutional layers,
CNNs used for classification and segmentation. The proposed pooling layers, normalization layers and Fully Connected
method, used Image-Net for extract features. The results layers. CNN is different in terms of the number of layers
obtained 97.5% accuracy for classification and 84% accuracy employed, the size and number of images, as well as the type
for segmentation. In [12], multiphase MRI images in tumor of activation functions employed. In the CNNs, the parameters
grading have been studied and a comparison has been made are chosen experimentally and experimentally based on trial
between the results of deep learning structures and base neural and error [18]. In other words, each CNN consists of several
networks. The results show that the network performance layers, the main layers of which are the Convolutional layer
based on the sensitivity and specificity of CNN improved by and the Sub-sampling layer which have been introduced in the
18% compared to the neural networks. following parts:
In [13], a deep learning-based supervised method is intro-
duced to detect synthetic aperture radar (SAR) image changes. 1) Convolution layer: Natural squirrels have fixed proper-
This method provided a dataset with an appropriate data ties in the sense that statistics are part of the image exactly
volume and diversity for training the DBN using input images the same as the other parts. This means that learned features
and the images obtained from applying the morphological of one section of the image can also be applied to other parts,
operators on them. The detection performance of this method and similar features are used in all image sections. After the
indicates the appropriability of deep learning based algorithms features are acquired, the features of Convolutional layer are
for solving the change detection problems. used to categorize images [19] [6].
In paper [14], a completely automated brain tumor clas-
2) Sub-Sampling layer: Operations in this layer are done
sification method is proposed based on DNN. The proposed
to reduce the size of the input image. By this layer, we receive
networks have been designed to be used in low-grade and high
a vector of points at the end of the CNN. The aggregation or
glioblastoma disease images. In this paper, a new architecture
Sub-Sampling operation is used as the mean pooling or max
of CNN is presented. The proposed a cascading architecture
pooling [20].
is proposed in which the output of a core CNN is used as an
additional source of information for the next CNN.
V. M ETHODOLOGY
III. F EATURE EXTRACTION A. Dataset
In machine learning and image processing, feature are The data set images used in this paper include brain MRI
created from the initial dataset. which facilitates the learning images of 153 patients, including normal and brain tumors
process. When the input data of an algorithm is too large, patients who referred to imaging centers because of headaches
it can be converted to a smaller set of features. The process After examination and diagnosis of the doctor, the collected
of extracting a subset from the primary features set is called images included brain images of 80 healthy patients Include
feature extraction [15]. The selected features include informa- 1321 images which has 56 images for testing data and 515
tion about the input data, so that the reduced representation of images for the train data. 73 patient tumors Include 571 images
the agent can be done instead of the full initial data. One of which has 170 images for test data and 1151 images for the
the important application of feature extraction is in the image train data. Of the total number of patients with brain tumor

364
disease, 86 were women and 68 were men, whose age range D. Convolutional neural method
from 8 to 66 years old. Of a total of 153 patients, 1892 images Initially, the images were applied to the CNN without any
were collected, 1666 images for train data and 226 images for feature extraction methods. The size of the input images is
test images. The collected images originally had an initial size initially 227 × 227. The Alexnet architect was used to identify
of 512 × 512. and classify the images, which consisted of 5 Convolutional
layers and 3 layers of Sub-sampling layers, Normalization
B. Simulation layers, Normalization layers, Fully Connected layers and lastly
layer the classification layer [21]. The fully connected layers
In few cases, some areas of fat in the pictures are mistakenly have 4096 neurons. We have two classes in this layer: brain
detected as tumor, or the tumors may not be seen by the tumor patient and normal patient. The utilized CNN is shown
physician; the most exact diagnosis is completely depended in Fig. 2.
on the physicians skill. In this paper, the CNN has been
used for tumor detection through brain images. There were VI. S IMULATION R ESULTS AND D ISCUSSION
additional margins of the images gathered from the imaging
The CNN managed to accurately categorize the images into
centers. These margins were cropped to prevent the noise of
tumor patient and normal patient tumors with precision of
the images. One of the main reasons for using the feature
98.67%. According to the results of the CNN on the initial
extraction technique and combining it with the CNN is to
images, in order to improve the network performance a combi-
retrieve the feature extraction of the images in order to increase
nation of Clustering algorithm for feature extraction and CNN
the accuracy of the network. According to the results of the
is used. Other classifiers such as the, Softmax Fully Connected
CNN on the initial images, in order to improve the network
layer classifier, RBF classifier and the DT classifier in the
accuract, in this study, a new method which is a combination
CNN architecture have been used to evaluate the efficiency of
of Clustering algorithm for feature extraction and CNN is
the proposed technique. Also, the criteria for the Accuracy,
proposed.
Sensitivity, Specificity, Precision have been used to verify the
function of the classifier. As shown in Table I, the accuracy
C. Feature extraction method of the CNN is obtained by Softmax classifier used to classify
images obtained 98.67%. Also, the accuracy of the CNN is
The central clustering is a clustering method. This algorithm
obtained with the RBF classifier 97.34% and the DT classifier
has a duplicate procedure that iterative, for a constant number
94.24%. By the method proposed(combination of Clustering
of clusters, attempts to obtain points as cluster centers, which
algorithm for feature extraction and CNN+Softmax), accuracy
are in fact the same mean points belonging to each cluster.
increased to 99.12% on the test data.
And assign each sample data to a cluster that gives the
data a minimum distance to the center of that cluster. In
the simple type of this method, first the cluster centers are TABLE I
T HE RESULTS OBTAINED FROM THE CN ON TEST DATA IMAGES WITH THE
selected randomly. The points are assign to the cluster centers CLASSIFIER .
according to the degree of similarity, and thus new clusters are
obtained. In this paper, the first-order clustering algorithm has
been used to feature extraction of the Fig. 1 shows the image Methods Accuracy Specificity Sensitivity Precision False
obtained from applying the clustering algorithm to the image.
CNN+
Softmax 98.67% 94.64% 100% 98.26% 3

CNN+
RBF 97.34% 89.28% 100% 96.59% 6

CNN+
DT 94.24% 85.71% 100% 95.37% 13

According to the results obtained from the categorizers,

the Softmax classifier has the best accuracy in the CNN.
After reviewing the results obtained from different categorizers
in the CNN, the SoftMax classification has been used in
the proposed method. Initially, the dataset were given to the
Fig. 1. Applying the clustering algorithm to the image. traditional CNN.

365
 Fig. 2. Proposed CNN to gender detection using MRI images

According to the obtained results, 3 images from the total Fig. 4 of the images categorized by the proposed network is
of 226 test data images were misdiagnosed and categorized as mistaken.
shown in the Fig. 3.

Fig. 4. Images that are wrongly categorized by the proposed method.

Fig. 3. Images by the CNN are mistakenly classified.

The results of CNN using the proposed method (i.e. com- The diagram of the network accuracy process is also shown
bining the feature extraction algorithm and CNN-SoftMax) on the test images in Fig. 5.
on dataset are shown in Table II. The accuracy of proposed
method increased to 99.12% on the test data, which is an
improvement compared to the traditional CNN.

TABLE II
T HE RESULTS OF THE CNN AND PROPOSED METHOD ON THE DATA TEST
IMAGES .

Methods Accuracy Specificity Sensitivity Precision False

CNN+
Softmax 98.67% 94.64% 100% 98.26% 3

Proposed
method 99.12% 96.42% 100% 98.83% 2

After using the proposed method, one of the misclassified

images from the traditional CNN is classified correctly. The Fig. 5. Network accuracy process.

366
 The diagram of the network loss process is also shown on ACKNOWLEDGMENT
the test images in Fig. 6.
The images used in the paper were collected by the author
from three Iranian imaging centers in over a year. The centers
of imaging that have gathered images have been the center of
the depiction of wisdom, truth, and infirmities. The images
were also screened by doctors at Dr. Rahimian’s and Dr.
Nourian’s imaging centers and confirmed by their correct
results.

R EFERENCES
[1] M. Karuna and A. Joshi, Automatic detection and severity analysis of
brain tumors using gui in matlab, International Journal of Research in
Engineering and Technology, 10, pp. 586-594, 2013.
[2] KS. Aboody, A. Brown, et al, Neural stem cells display extensive
tropism for pathology in adult brain Evidence from intracranial gliomas,
Proceedings of the National Academy of Sciences, 97 (23), pp. 12846-
12851, 2000.
[3] A. joshi, D. H. Shah, et al, Survey of brain tumor detection techniques
through MRI images, International Journal of Research in Engineering
and Technology, 10, pp. 586-594, 2013.
[4] JP. Poonam, Review of image processing techniques for automatic detec-
tion of tumor in human brain, International Journal of Computer Science
Fig. 6. Network loss process. and Mobile Computing, 2(11), pp. 117-122, 2013.
[5] H. Cecotti and A. Graeser, Convolutional neural network with embed-
ded Fourier transform for EEG classification, Pattern Recognition,19th
International Conference on. IEEE, ICPR, pp. 1-14, 2008.
C ONCLUSION [6] R. Bayot and T. Gonalves, A survey on object classification using
convolutional neural networks, 2015.
In this paper, a new method based on the combination of [7] M. Soltaninejad, et al, Automated brain tumour detection and segmenta-
feature extraction algorithm and the CNN for tumor detection tion using superpixel-based extremely randomized trees in FLAIR MRI,
from brain images is presented. The CNN is capable of International journal of computer assisted radiology and surgery, 12(2),
pp. 183-203, 2017.
detecting a tumor. The CNN is very useful for selecting [8] S. Pereira, et al, Brain tumor segmentation using convolutional neural
an auto-feature in medical images. Images collected at the networks in MRI images, IEEE transactions on medical imaging, 35(5),
centers were labeled by clinicians, then, tumor screenings were pp. 1240-1251, 2016.
categorized into two normal and patient classes. A total of [9] Halimeh Siar, Mohammad Teshnehlab, Diagnosing and Classification
Tumors and MS Simultaneous of Magnetic Resonance Images Using
1666 images were selected as train data and 226 images were Convolution Neural Network, 7th Iranian Joint Congress on Fuzzy and
taken as a test data. The proportion of image categorization Intelligent Systems (CFIS), 2019.
in two classes was proportional from the ratio of patients [10] L. Szilagyi, et al, Automatic brain tumor segmentation in multispectral
MRI volumes using a fuzzy c-means cascade algorithm, In 2015 12th
to healthy subjects. Images were applied to the CNN after international conference on fuzzy systems and knowledge discovery
preprocessing. In order to evaluate the performance of the (FSKD), IEEE, pp. 285-291, 2015.
CNN, has been used by other classifiers such as the RBF [11] Y. Xu, et al, Deep convolutional activation features for large scale brain
tumor histopathology image classification and segmentation, In 2015
classifier and the decision tree classifier in the CNN architec- IEEE international conference on acoustics, speech and signal processing
ture. The accuracy of the CNN is obtained Softmax classifier (ICASSP), pp. 947-951, 2015.
98.67% categorization. Also, the accuracy of the CNN is [12] Y. Pan, et al, Brain tumor grading based on neural networks and convo-
lutional neural networks, In 2015 37th Annual International Conference
obtained with the RBF classifier 97.34% and the DT classifier of the IEEE Engineering in Medicine and Biology Society (EMBC), pp.
94.24%. In addition to the Accuracy criterion, we use the 699-702, 2015.
benchmarks of Sensitivity, Specificity and Precision evaluate [13] F. Samadi, G. Akbarizadeh, et al, Change Detection in SAR Images
using Deep Belief Network: a New Training Approach based on Morpho-
network performance. According to the results obtained from logical Images, IET Image Processing, 2019.
the categorizers, the Softmax classifier has the best accuracy [14] M. Havaei, et al, Brain tumor segmentation with deep neural networks,
in the CNN. The CNN has been able to categorize accurately Medical image analysis, 35, 18-31, 2017.
[15] E. Alpaydin, Introduction to Machine Learning , London: The MIT
98.67% images in two normal and patient classes; and from Press. pp. 110, 2009.
a total of 226 images, three images have been constrained by [16] W. Zhang, K. Itoh, et al, Parallel Distributed Processing Model with
the CNN. Using the proposed method of feature extraction local Space-Invariant Interconnections and its Optical Architecture, Ap-
plied optics, 29(32), pp. 4790-4797, 1990.
and applying to the CNN. The accuracy of proposed method [17] LY. LeCun, K. Kavukcuoglu, et al, Convolutional networks and appli-
increased to 99.12% on the test data, which is an improvement cations in vision, In ISCAS, pp. 253-256, 2010.
compared to the traditional CNN. Due to the importance of the [18] HH. Aghdam, EJ. Heravi et al, Guide to Convolutional Neural Networks,
A Practical Application to Traffic-Sign Detection and Classification,
diagnosis given by the physician, the accuracy of the doctors Springer, 2017.
help in diagnosing the tumor and treating the patient increased [19] D. Ciresan, U. Meier, et al, Multi-column deep neural networks for
high medical accuracy of the proposed method. image classification. Int.Conf. Computer Vision and Pattern Recognition,
IEEE, pp. 1202.2745, 2012.

367
[20] B. Fasel, Robust face analysis using convolutional neural networks,
20040, 2002.
[21] A. Krizhevsky, I. Sutskever, et al, Imagenet classification with deep con-
volutional neural networks, In Advances in neural information processing
systems, pp. 1097-1105, 2012.

368