DOI:10.31557/APJCP.2019.20.7.

2095
Brain Tumour Segmentation Using Convolutional Neural Network with Tensor Flow

RESEARCH ARTICLE Editorial Process: Submission:01/16/2019 Acceptance:07/09/2019

Brain Tumour Segmentation Using Convolutional Neural

Network with Tensor Flow
M Malathi*, P Sinthia

Abstract
Introduction: The determination of tumour extent is a major challenging task in brain tumour planning and
quantitative evaluation. Magnetic Resonance Imaging (MRI) is one of the non-invasive technique has emanated as
a front- line diagnostic tool for brain tumour without ionizing radiation. Objective: Among brain tumours, gliomas
are the most common aggressive, leading to a very short life expectancy in their highest grade. In the clinical practice
manual segmentation is a time consuming task and their performance is highly depended on the operator’s experience.
Methods: This paper proposes fully automatic segmentation of brain tumour using convolutional neural network. Further,
it uses high grade gilomas brain image from BRATS 2015 database. The suggested work accomplishes brain tumour
segmentation using tensor flow, in which the anaconda frameworks are used to implement high level mathematical
functions. The survival rates of patients are improved by early diagnosis of brain tumour. Results: Hence, the research
work segments brain tumour into four classes like edema, non-enhancing tumour, enhancing tumour and necrotic tumour.
Brain tumour segmentation needs to separate healthy tissues from tumour regions such as advancing tumour, necrotic
core and surrounding edema. This is an essential step in diagnosis and treatment planning, both of which need to take
place quickly in case of a malignancy in order to maximize the likelihood of successful treatment.

Keywords: Brain tumour- magnetic resonance imaging- convolutional neural network- segmentation

Asian Pac J Cancer Prev, 20 (7), 2095-2101

Introduction avoid injury to the sites of language, motor sensory

function during therapy. The manual segmentation of brain
Dimililera et al., (2016) discussed Brain tumour is tumour is labour sensitive and the segmentation results
the abnormal growth of cells within the brain. There are based on the operators experience and their subjective
two main types of tumours like benign and malignant decision making. Hence there is a need for fully automatic,
tumours. Generally, tumour is classified into primary objective and reproducible segmentation methods. There
and secondary tumour. Primary tumour starts within the are many challenges and fully automatic algorithms for
brain and secondary tumour will spread to the other parts brain tumour segmentation, because of its high variability
of the body. There are many medical imaging methods of brain tumour size, shape, regularity, location and their
available like X-ray, CT (Computed Tomography) and heterogeneous appearance. Rémi et al., (2018) discussed
MRI (Magnetic Resonance Imaging). The research article the proposed work classifies the high grade gliomas.
uses MRI brain images, because of its high resolution Generally, astrocytomas are the most common type of
and good quality of an image. After capturing MRI brain glioma found in both adults and children. Astrocyotomas
image, it is necessary to separate the tumour region from are classified into low grade gliomas and high grade
the MRI brain image. Accurate segmentation of medical gliomas. The work uses fully automatic convolutional
images helps the radiologist for radiotherapy planning. neural network. The convolutional neural network is
The most common type of tumours is such as astrocytoma. implemented in python programming. The anaconda is
oligodondroglima and glioblastama. The presented work one of the frameworks for machine learning concept,
mainly segments glioma types of brain tumour. The in which neural network tool for training of BRATS
astrocytoma is the most common type of glioma which is database is implemented using tensor flow. By using this,
formed by star shaped cells. Segmentation of tumour from the accuracy of segmentation is improved and it has the
brain images is critical, because of its complex structure features to process the larger dataset. The performance of
of the brain tissues so, it is compulsory to provide further the segmentation of MRI brain tumour is compared with
diagnosis. Further, it is difficult to perform accurate the ground truth images of BRATS database 2015. The
delineation in radiotherapy. Since, it is compulsory to dice co-efficient is the parameter to define the accuracy

Saveetha Engineering College,Chennai, India. *For Correspondence: [email protected]

Asian Pacific Journal of Cancer Prevention, Vol 20 2095

M Malathi and P Sinthia

of automatic segmentation. uses deep convolutional neural network algorithm to

locate the tumour which has an irregular shape. Hence,
Relevant work and our contributions the survival rate of the patient is increased with accurate
The various brain tumour segmentation techniques segmentation of brain tumour. The problem of overfitting
have been discussed as follows: is removed by introducing max-out and drop out layers in
Dong et al., (2017) proposed noninvasive magnetic the patch processing. The proposed algorithm also uses a
resonance techniques as a diagnostic tool for brain tumour preprocessing method to remove the unwanted noise and
to identify brain tumour without ionizing radiation. post processing help to remove small false positive using
Manual segmentation of the 3D MRI volumes needs morphological operators. The research work uses BRATS
larger time, and the performance is mainly based on the 2013 dataset for segmentation.
operators experience. Hence, the author recommended a Wang et al., (2017) recommended that convolutional
u-net based deep convolution network. This segmentation neural network provides the state of the art performance
is implemented on BRATS 2015 datasets, which contain for automatic medical image segmentation. But, it does
220 high grade glioma brain tumour and 54 low grade not provide robust results for clinical use. It has the
tumour cases. The performance of our proposed method limitation to the lack of generalizability of previously
was compared to the manual delineated ground truth U-net unseen object classes. The problem is rectified by novel
based deep neural network provides the superior results deep learning based interactive segmentation framework
for the core tumour regions. by relating CNN into bounding box and scribble based
Haveri et al., (2017) illustrated a brain tumour segmentation pipeline. The proposed method makes the
segmentation using deep neural networks to glioblastomas CNN model adaptive to a specific test image, which may
(both low and high grades) MRI image. This kind of be unsupervised or supervised.
brain tumour appears anywhere in the brain and also Cui et al., (2018) developed a novel automatic
it has any shape, size and contrast. The article utilizes segmentation based on cascaded deep learning
the convolutional neural network as a machine learning convolutional neural network. It has two sub networks;
algorithm. It exploits both local and global features for tumour localization network (TLN) and a intra tumour
tumour segmentation. The author uses BRATS dataset classification network (ITCN). The tumour region from
for research work. the MRI brain slice is separated using tumour localization
Isin et al., (2016) discussed that segmentation of brain network and ITCN helps to label the defined tumour region
tumour is one of the challenging tasks in medical field. into multiple sub-regions. The work was performed on
The lifetime of the patient is improved by early diagnosis multimodal brain tumour segmentation (BRATS, 2015)
of brain tumour. Manual segmentation of brain tumour for dataset, which had 220 high grade glioma (HGG) and 54
large amount of data is a time consuming process. Hence low grade glioma (LGG) cases. The evaluation can be
there is a need for automatic segmentation. Now a days the performed by dice coefficient, positive predictive value
automatic segmentation uses deep learning techniques for (PPV) and sensitivity.
segmentation. It provides efficient segmentation for a large Khawaldeh et al., (2018) offered a widespread machine
amount of MRI based image data. The article reviewed the learning technique for medical image classification and
state of the art methods of deep learning. The conventional segmentation. The approach uses conv net for classifying
automatic segmentation methods need prior knowledge brain medical images into healthy and unhealthy brain
into probabilistic maps or selecting highly representative images. The implemented method classifies the brain
features for classifiers, which is challenging task. But, tumour into low grades and high grades. It uses alex
the convolutional neural network method automatically krizhousky network deep learning architecture to classify
learns the corresponding complex features for both healthy the MRI brain tumour. The tumour classification is
brain tissues and tumour tissues from multimodal MRI br performed on the whole image rather than pixels.
ain images. Chinmayi et al., (2017) conferred a method for MRI
Pereira et al., (2016) described gliomas is one of brain tumour segmentation and classification using
the aggressive types of brain tumour which leads to Bhaltacharya co-efficient. The unwanted skull portions
the short lifetime with their highest grade. Since, the were removed using anisotropic diffusion filter. Further, it
MRI techniques produce large amount of data, manual uses a fast bounding box algorithm to extract the tumour
segmentation needs reasonable time. The automatic area. It uses deep learning CNN to train the MRI brain
brain tumour is one of the challenging tasks due to the tumour image. Finally, the results of the proposed method
large, spatial and structural variability among brain. The compared in terms of accuracy, similarity index, PSNR
author suggested a new segmentation method based on and MSE. The results will help the radiologist to identify
a convolutional neural network with small 3×3 kernals. the size and position of a tumour.
The small kernel helps the deep architecture to avoid Konstantinos et al., (2017) developed a segmentation
over fitting by assigning the fewer number of weights in of brain lesion which is a challenging task, was performed
the network. And also using intensity normalization as a using the 3-D convolutional neural network. The dual
preprocessing along with convolutional neural network pathway architecture was used to extract the local and
provides effective segmentation. Further, the proposed larger contextual information, which operates an input
work is implemented BRATS 2013 data base. image at multiple scales. The false positives which were
Hussian et al., (2017) suggested a segmentation removed by using 3D fully connected conditional random
algorithm to detect the gliomas based brain tumour. It field. The segmentation process was used to separate
2096 Asian Pacific Journal of Cancer Prevention, Vol 20
 DOI:10.31557/APJCP.2019.20.7.2095
Brain Tumour Segmentation Using Convolutional Neural Network with Tensor Flow
lesion on multichannel MRI with traumatic brain injuries,
brain tumours and ischaemic stroke. The 3-D CNN is an
effective method, which provides good segmentation
without increasing the computational cost and the number
of training parameters.
In modern years, several methods have been
implemented to automatically segment MRI brain
tumours. These methods can be basically divided into
two categories respectively hand- crafted features and
classifier method based on traditional learning such as
support vector machine (SVM) and random forest. The
second method is based on fully automatic methods
based on deep learning using the Convolutional Neural
Network (CNN). The first category method use manually
separated features and it is given as input to classifiers.
After determination of hand crafted features. The
classifiers do not modify the features during training.
But in second category the features, specifications can
be adapted to do specific task of training data. Currently
deep neural network CNNs mostly used in the computer Figure 1. Predictive Model of Proposed of Work
vision community.

Maerials and Methods The article uses the HGG image of BRATS 2015. Pulse
sequence images of brain tumour as shown in Figure 2.
The research article uses convolutional neural network
for MRI brain tumour segmentation using tensor flow. Preprocessing
Normally, the segmentation is performed using various The main challenging task is removing artifacts
tools like MATLAB, LABVIEW etc. The research article produced by inhomogeneity in a magnetic field or small
uses tensor flow based MRI brain tumour segmentation movements created by the patient during scanning. Many
in order to improve segmentation accuracy, speed and time bias is present in the scanning results, which affect
sensitivity. Segmentation can be performed on BRATS the segmentation results, particularly in the computer n
MRI brain images and results are compared in terms of based models. Chmelika et al., (2018) demonstrated work
dice co-efficient. uses n41TK bias correction for the T1 and T1C images
in the data set. The n4T1K bias correction removes the
Python Based Convolutional Neural Network intensity gradient on each scanning images. Additionally,
Many researchers use MAT LAB to implement the noise reduction is also performed by median filter in
segmentation process. Our research work utilizes the order to standardize the pixel intensities. Hence, noise
python programming to implement the segmentation reduction and bias correction helps to improve the data
of MRI brain tumour. The features are listed below in processing and provides the better segmentation, multiple
order to choose python programming to implement the radio frequency pulse sequences can be used to provide
research work the different types of tissue. In BRATS data base there
1. Python code is more compact and readable than are four different sequences available for each image like
MATLAB fluid attenuated recovery (FLAIR), T1, T1 contrasted and
2. The python data structure is superior to MATLAB T2. The chemical and physiological characteristics can
3. It is an open source and also provides more graphic be obtained from these pulse sequences, which result in
packages and data sets contrast between the individual classes. The architecture
Hence, the proposed work utilizes python programming for the proposed work as shown in Figure 3.
instead of MAT LAB. There are some additional python
packages used during the implementation process of our
research work through python. The flow diagram of the
proposed work as shown in Figure 1.

MRI Brain data acquisitions

The recommended method was tested and implemented
on the BRATS 2015 dataset, which had 220 high grade
glioma (HGG) and 54 (LGG) low grade glioma patient
scanned images. Each patient in the BRATS 2015 dataset
multimodal MRI was available and also four scanning
sequences were implemented for every patient using T1
weighted (T1), T1 weighted imaging with gadolinium Figure 2. Pulse Sequence Images of Brain Tumour
enhancing contrast (T1C), T2 weighted and FLAIR.
Asian Pacific Journal of Cancer Prevention, Vol 20 2097
M Malathi and P Sinthia

Figure 3. The Architecture of Convolutional Neural Network

Neural network covers larger region and provides global information. In

For image recognition, neural network is one of order to reduce the computation burden produced by the
the powerful tools to perform segmentation. MRI is larger global patch we used down sampling of all global
one of the most commonly used imaging techniques to patches. The down sampling process is also called as
capture MRI brain images. Automatic segmentation is a pooling.
challenging task because of its large spatial and structural
variability. Hence, the proposed system implements the Convolution section
automatic segmentation method based on CNN exploring There are multiple layers in the convolution section,
small 3×3 kernals. The small size kernals help to design which help to sequentially identify the features using
deeper architecture by using fewer number of weights in convolution operations. The captured features are low
the network. level features like edges and corner relationship between
neighbouring voxels. The feature maps are the output of
CNN algorithm the convolutional layer. The complete convolution region
Raphael Prevost et al., (2018) discussed CNN consists of three different sub paths. The path 1 and 2
algorithm performs the voxel-wise classification problem. utilize the same local patch with various filter sizes to
The tumour or lesion portion is separated from the internment various neighbourhood patterns, at the same
background by calculating the probability of each image time subpath 3 uses downsampled global patch to provide
voxel belonging to the target is known. The CNN network global features. The convolutional layer calculates the
has four sections input and convolution sections. The output of the neurons that are connected to either local or
input layer processes the input image in order to produce global regions in the input. The convolution is the process
the designed image patches. The convolution section of performing dot product between their inputs and their
process the designed image patches, in which multilayer receptive field to which they are connected to in the input
convolutional filters operates and output feature maps. volume. For our research work
Further, the fully connected layer that groups all feature L- be the depth of the convolution filter stack
maps. The classification section estimates a prediction a1 (1Ɛ [1,L]) represents the number of feature maps
score to classify the every image voxel and provides a in the lst layer and
segmentation map. Fil (iƐ [1,αl] ) represents the ith feature map of the lst
layer Fil is calculated by
Input section
The input section generates the image patches for Fil = g (
the remaining of the network. Chena et al., (2017)
implemented The method performs the classification Where g(.) is the PReLU function. The PRELU is a
based on voxel-wise, where each voxel is classified neuron activation function defined by
based on the linear and nonlinear relationship between g(x)= x, x >0
the focal voxel’s intensity and its neighbours. The input ax x≤0
3D image size is large; hence, calculation of linear and a is the network parameter
nonlinear relationship between all voxels in the complete Wj1l is the filter connecting the jth feature map in the
image is complex. Hence, the entire image is divided into l-1th layer and bil is the bias of the artificial neuron model
smaller patches in order to find the relationship within a ReLU is the rectified linear units.
particular region instead of the entire image. It reduces the It is one of the most popularly used activation
computational time and also memory space. It separates functions. It is represented as
both local and global patches as input for the convolution R(x)= max( 0, X) ie
section. For every extraction process, the central voxel is If X < 0 R(x)=0
chosen randomly and extracted concentric local and global If X>=0, R(x)=X
image patches. The neighbouring voxel around the central Most of the machine learning techniques refers to
voxel provides the local information and global patch this function because of its simplicity and also it avoids

2098 Asian Pacific Journal of Cancer Prevention, Vol 20

 DOI:10.31557/APJCP.2019.20.7.2095
Brain Tumour Segmentation Using Convolutional Neural Network with Tensor Flow
vanishing gradient problem. But, it should be used only Results
within the hidden layer of neural network model. For
example, output layers, softmax function will be used The BRATS data base provides different target
to calculate the probabilities for the classes. During the structures edema (label 1), advancing tumour.
training phase, some gradients can be fragile and die. (Label 2), non-advancing tumour (label 3), necrotic
Hence, it needs a weight update, further, it never activates tumour core (label 4) and background.
on any data point again. It can be avoided by another The proposed work uses either colour, gray scale or
modification called leaky ReLU. It provides small slope intensity images with a default size of 220×220. The
to keep updates alive. automatic brain tumour detection of a patient consists of
two important stages, namely, image segmentation and
Fully connected region edge detection. The next important step is adding path
The special information is preserved in the fully from the image data set file to the system directory. It can
connection section by fusing of all of the feature maps be implemented by an anaconda frame work to allocate
produced by the convolution section. The operation of data sets from computer memory. It can be performed
fully connected section is similar to the convolution either manually or by programming using the python OS
section in which each co-efficient node performs like a inbuilt operation.
convolution filter with the size (1,1,1) The main objective of segmentation is to cluster pixels
into image region and it helps in identifying a region of
Classification section interest ie., to locate tumour and other abnormalities. The
The classification section helps to provide probability pixels segmented portion is compared with normal brain
for each voxel. Most of the Convolutional neural network image in the jupytor note book. This kind of comparison
uses softmax function in order to map the feature maps helps to locate abnormal parts of brain tumour patient. The
into categorical probabilities. next step is edge detection. The work uses cannon edge
detection in order to classify the brain tumour portion of
Jp(

Where xv and Csv are the vth target voxel’s position and
ground truth,

Where (f1L, f2L, f3L) is the Lth layer feature map size.
S ϵ [1, B] where the B is the size of a batch, PCsv is the
output of the softmax function. The parameter θ is denoted Figure 4. Segmented Image of Brain Tumour
as filter co-efficients and bias. It can be found by training
through the stochastic gradient descent method. SGD
is one of the stochastic approximations of the gradient
descent optimization method in order to minimize an
objective function iteratively. For every iteration, the
method estimates the gradient from a subset of labeled
dataset and also it helps to increase the speed of training
in a huge training set.

Training of network
The training model is created using Keras. The 4 layer
sequential model was trained on selected 33× 33 patches
of MRI images in order to classify the center pixel. Every
input has 4 channels, among four, one channel helps for
each imaging sequence and the net can learn the relative Figure 5. Result of Brain Tumour Segmentation
pixel intensities of each given class.
Validation of data set
The proposed method uses BRATS database for
evaluating the brain tumour segmentation methods.
BRATS provides each patient’s T1 weighted MRI with
gadolinium contrast (T1c) and T2 weighted fluid attenuated
inversion recovery (FLAIR) images. This kind of data set
compares our automatic delineation method with other
algorithms.

Figure 6. Brain Tumour Segmentation with Different

Labels
Asian Pacific Journal of Cancer Prevention, Vol 20 2099
M Malathi and P Sinthia

the segmented image. The white portion shown is tumour.

Sensitivity =
In the resultant, red represents the necrotic core of the
tumour, yellow represents the advancing tumour, blue The table provides the Dice co-efficient value for
represents non-advancing tumour and green represents the different ground truth brain images.
swelling part of the tumour called edema. The results of In conclusion, the segmentation of brain tumour plays
the segmentation process are shown in Figure 4, 5 and 6. an important role in diagnostic procedures. The accurate
segmentation helps in clinical diagnostic, but also helps
Discussion to increase the lifetime of the patient. In this paper a
segmentation of brain tumour implemented using CNN
Dice similarity coefficient architecture. The algorithm relates both local and global
In mathematical notation, dice score is the number of features, because it helps to perform the segmentation
false positives divided by the number of positives added accurately. The training and testing speed is increased by
with the number of false positives. using max pooling, max out and drop out complement
the learning process. The speed is increased by reducing
Experiments and performance evaluation the features in the fully connected layer. Reducing of
Five-fold cross validation method is used for parameters also causes the reduction of overfitting.
evaluation of HGG and LGG data respectively. For The result shows that the implemented method helps
every patient, four subnormal regions are validated. It is in detection of enhancing tumour as well as specifying
represented by tumour to the actual tumour region only.
a) The complete tumour region (It includes all the four
intra tumoural classes label 1,2,3 and 4) Conflicts of interest
b) The core tumour region (It excludes edema region, None .
(label1), includes label (2, 3, 4)
c) The advancing tumour (only label 2) Acknowledgements
For every tumoural region, the performance of
segmentation has been evaluated using the dice similarity The authors are grateful to the mysterious reviewers for
co-efficient and the sensitivity. The DSC measures the the insights and the invaluable recommendations provided
overlap between the manual delineated brain tumour in order to progress the initial version of this manuscript.
regions and the segmentation results of our fully automatic Also, the authors are grateful to the anonymous assistant
method that is who provided help in order to perform the statistical
analysis.
DSC=
TP- True positive References
FP –False positive
FN-false negative Chena L, Bentley P, Rueckert D (2017). Fully automatic acute
ischemic lesion segmentation in DWI using convolutional
Additionally, sensitivity is also used to measure the neural networks. Neuro Image Clin, 15, 633-53.
number of TP and FN that is Chinmayi P, Agilandeeswari L, Prabu Kumar M, Muralibabu
K (2017). An efficient deep learning neural network based
brain tumor detection system. Int Pure Appl Mathemat,
117, 151-60.
Table 1. Parameters for Measurement of Various Ground Chmelika J, Jakubiceka R, Waleka P, et al (2018). Deep
Truth Brain Images convolutional neural network-based segmentation and
S.No Input image Output image Parameters classification of difficult to define metastatic spinal lesions
1 Dice co- efficient:0.73 in 3D CT data. Med Image Analysis, 49, 76-88.
Advancing tumour:0.76 Cui S, Mao L, Jiang J, Liu C, Xiong S (2018). Automatic
Sensitivity:0.82 semantic segmentation of brain gliomas from MRI images
Non advancing:0.83 using a deep cascaded neural network. J Healthcare
2 Dice co- efficient:0.73 Engineering, 2018, 1-14.
Advancing tumour:0.7 Dimililera N, Ilhan A (2016). In Proceedings of 12th International
Sensitivity:0.80 Conference on Application of Fuzzy Systems and Soft
Non advancing:0.86 Computing, August 29-30, Effect of image enhancement on
3 Dice co- efficient:0.77 MRI brain images with neural networks, Vienna, Austria.
Advancing tumour:0.76 Dong H, Yang G, Liu F, Mo Y, Guo Y (2017). Automatic brain
Sensitivity:0.81 tumor detection and segmentation using U-net based fully
Non advancing 0.83 convolutional networks. MIUA, 3, 1-12.
4 Dice co- efficient:0.69 Havaei M, Davy A, Warde Farley D, et al (2017). Brain tumor
Advancing tumour:0.71 segmentation with Deep Neural Networks. Med Image
Sensitivity:0.82 Analysis, 35, 18-31.
Non advancing 0.82 Hussain S, Anwar SM, Majidmil M (2017). In Proceedings of
5 Dice co- efficient:0.65 39th International Conference of the IEEE Engineering
Advancing tumour:0.72 in Medicine and Biology Society, July, Brain Tumor
Sensitivity:0.80 Segmentation using Cascaded Deep Convolutional Neural
Non advancing:0.80
Network, Jeju, South Korea.
2100 Asian Pacific Journal of Cancer Prevention, Vol 20
 DOI:10.31557/APJCP.2019.20.7.2095
Brain Tumour Segmentation Using Convolutional Neural Network with Tensor Flow
Ism A, Direkoglu C, Sah M (2016). In proceedings of 12th
International Conference on Application of Fuzzy Systems
and Soft Computing, August 29-30, Review of MRI based
brain tumor image segmentation using deep learning
methods, Vienna, Austria.
Khawaldeh S, Pervaiz U, Rafiq A, Rami SA (2017). Noninvasive
grading of glioma tumor using magnetic resonance imaging
with convolutional neural networks. Appl Sci, 8, 1-17.
Konstantinos KK, Ledig C, Virginia, FJ, et al (2017). Efficient
multi scale 3D CNN with fully connected CRF for accurate
brain lesion segmentation. Med Image Analysis, 36, 61-78.
Pereira S, Pinto A, Alves V, Carlos AS (2016). Brain tumor
segmentation using convolutional neural networks in MRI
images. IEEE Trans Med Imaging, 35, 1240-51.
Prevost R, Saleh M, Jagodaa S, et al (2018). 3D freehand
ultrasound without external tracking using deep learning.
Med Image Analysis, 48, 187-202.
Rémi Y, Cédric Balsat YE, Verset L, et al (2018). Segmentation of
glandular epithelium in colorectal tumours to automatically
compartmentalise IHC biomarker quantification: A deep
learning approach. Med Image Analysis, 49, 35-45.
Wang G, Li W, Azuluaga M, et al (2018). Interactive medical
image segmentation using deep learning with image specific
fine tuning. IEEE Trans Med Imaging, 37, 1562- 73

This work is licensed under a Creative Commons Attribution-

Non Commercial 4.0 International License.

Asian Pacific Journal of Cancer Prevention, Vol 20 2101