Brain Tumor Detection Using Deep Learning

Arya Anandan Nisha S M

Department of Computer Science and Assistant Professor, Department of
Engineering Computer Science and Engineering
Royal College of Engineering and Royal College of Engineering and
Technology Technology
Thrissur, India Thrissur, India
aryaanandan2001@gmail.com nisharanjith1810@gmail.com

ABSTRACT: Brain tumor is considered as one of the various diagnostic methods are computed tomography,
aggressive diseases, among children and adults. Brain tumors magnetic resonance imaging, tissue biopsy etc.
grow very fast and if not treated well, the survival chances of
the patient are very less. Early detection of brain tumors is Better treatments are now available for brain tumors.
very important. Proper treatment planning and accurate There is a chance of focal neurological deficits, such as
diagnostics is at the topmost priority to improve life motor deficit, aphasia or visual field defects in the
expectancy of the patients. The abnormal growths of cells in treatment. Side effects can be avoided by measuring
the brain are called tumors and cancer is a term used to tumor size and time to tumor progression (TTP)[4] .
represent malignant tumors. The best technique to detect Estimation of density of affected areas can give a better
brain tumor is Magnetic Resonance Imaging. The MRI images measurement in therapy.
are examined by the radiologist. MRI scan images are taken to
analyze the disease condition. Hence an automated brain Deep learning is a machine learning technique that
tumor detection system is required to detect tumors at its early instructs computers what to do as a human think and do in
stage. Density of the tumor can be estimated from the a scenario. In deep learning, a computer model is able to
segmented mask and it will help in therapy. Deep learning do classification tasks from images, sound or text.
technique is employed to detect abnormality from MRI Sometimes human level performance is being exceeded
images. Multi-level thresholding is applied to segment the by deep learning techniques. One of the most popular
tumor region. Number of malignant pixels gives the density of neural networks is an artificial neural network that has a
the affected region. collection of simulated neurons. Each neuron acts as a
node and by links each node is connected to other
Index Terms: Medical Image Processing, Brain tumor, MRI, nodes[5].
Artificial neural network, CNN.
The aim of this paper is to build a system that would
I. INTRODUCTION help in cancer detection from MRI images through the
The early detection and treatment of brain tumor helps convolution neural network. The proposed method was
in early diagnosis which aids in reducing mortality rate. tested and compared with the existing classification
Image processing has been widespread in recent years techniques to determine the accuracy of the proposed
and it has been an inevitable part in the medical field method.
also. Human body is made up of many organs and brain is
vital of all. The abnormal growth of cells in the brain
causes brain tumor. Brain tumor is also referred to as
intracranial neoplasm. The two types of tumors are
malignant and benign tumors. Standard MRI sequences
are generally used to differentiate between different types
of brain tumors based on visual qualities and contrast
texture analysis of the soft tissue. More than 120 classes
of brain tumors are known to be classified in four levels
according to the level malignancy by the World Health
Organization (WHO) [1].
All types of brain tumors evoke some symptoms based
on the affected region of the brain. The major symptoms
may include headaches, seizures, vision problems, Figure 1: MRI image
vomiting, mental changes, memory lapses, balance losing
etc [2]. Incidence of brain tumors are due to genetics, II. RELATED WORKS
ionizing radiation mobile phones, extremely low Image segmentation and classification is one of the
frequency magnetic fields, chemicals, head trauma and major tasks in machine learning and it is widespread in
injury, immune factors like viruses, allergies, infections, clinical diagnosis also. Mircea Gurbin, Mihaela Lascu,
etc[3]. The malignant tumors, also known as cancerous and Dan Lascu et al. [6] proposed a method consisting of
tumors, are of two types - primary tumors, which start Continuous Wavelet Transform (CWT), Discrete Wavelet
from the brain, and secondary tumors, which originate Transform (DWT) and Support Vector Machine (SVM).
somewhere and spread to the brain. The risk factors for
brain tumor are exposure to vinyl chloride,
neurofibromatosis, ionizing radiations and so on. The

.
It uses different levels of wavelets, and by training, the III. PROPOSED METHOD
cancerous and non-cancerous tumors can be identified. System architecture of the proposed system is shown
The computation time is longer for the proposed method. in figure 1. The components are image acquisition, pre-
Somasundaram S. and Gobinath R. et al. [7] processing, segmentation, Training and Splitting and
explains the present status of detection and segmentation classification.
of tumour through deep learning models. For deeper A. Image Acquisition
segmentation, 3D based CNN, ANN and SVM is used. Different bio-medical image records are available for
Damodharan S. and Raghavan D. et al.[8] address the study of brain tumour detection. Conventional
segmentation of pathological tissues (Tumor), normal methods are Computer Tomography (CT) and Magnetic
tissues (White Matter (WM) and Gray Matter (GM)) and Resonance Imaging (MRI). Positron Emission
fluid (Cerebrospinal Fluid (CSF)), extraction of the Tomography, Cerebral Arteriogram, Lumbar Puncture,
relevant features from each segmented tissues and Molecular testing are also used for brain tumour
classification of the tumor images with Neural Network detection. But these are expensive. MRI is working with
(NN). the principle that both the magnetic field and radio waves
can create an image of the interior of the human body by
detecting the water molecule present. Portable and
miniaturised MRI machines are developed now to avoid
the complexity of conventional scanning methods. MRI
has a better resolution and contains rich information. The
MRI dataset from the kaggle uploaded by Navoneel
Chakrabarty has been used here[12].It contains 98 normal
brain images and 155 abnormal images. In this dataset,
‘yes’ means tumour images and ‘no’ means healthy
images. The augmentation process is also applied here to
increase the number of samples. Augmentation step
contains a rotation range of 10 degrees, width shift range
of 0.1, height shift range of 0.1, brightness range of
(0.3,1.0), horizontal and vertical flip. A total of 2530
images were selected from the augmented data. The final
dataset contains 980 normal and 1550 abnormal images.

(a)
Image Image Pre- Classificat
Training
Acquition Processing ion
&Spliting

(b)

Figure 2 : System architecture of brain tumour detection

(a) Conventional Method (b) Deep learning
Figure 3 : Brain MRI dataset (a) Normal (b) Tumour
G. Hemanth, M. Janardhan and L. Sujihelen et al.[9]
states that with the appropriate use of data mining
classification technique early detection of tumour is made B. Pre processing
possible. It uses an automatic segmentation method based The aim of the pre-processing step is preparing the
on CNN. Reema Mathew A. and Dr. Babu Anto P. et al. brain images for further processing [13].This process
[10] stated that by the segmentation of MRI tumour mainly depends on the data acquisition device which has
region can be identified. With the help of radiological its own intrinsic parameters. Gray scale or 2D conversion
evaluations, the size and location of tumours can be is needed, if the raw data is in 3D. Median filtering is best
identified. Here the segmentation is done manually and it suited for biomedical images to avoid noise. The dataset
is time consuming. The pre processing is done using contains images in different resolutions. As part of the
anisotropic diffusion filters. The segmentation and augmentation process , each image is rotated and scaled to
classification is done using support vector machines. Wei a standard format. Histogram equalisation helps to
Chen, Xu Qiao, Boqiang Liu, Xianying Qi, Rui Wang and enhance the image quality. Contrast limited adaptive
Xiaoya Wang et al. [11] propose a novel method based histogram equalisation algorithm is applied to enhance the
on the features of separated local squares. Super pixel images.
segmentation, feature extraction and segmentation model C. Training & Splitting
construction are done on this proposed method for brain
The train-test split procedure is used to estimate the
tumour segmentation.
performance of machine learning algorithms. It can be used for
classification or regression problems and can be used for any
supervised learning algorithm.

.
The train-test procedure is appropriate when there is a
sufficiently large dataset available.

D. Classification
Many machine learning approaches are being
implemented in disease detection from brain images.
Artificial neural networks can be used here to classify, if
the features are extracted in an order[16].An ANN
classifier assumes one feature that is not related to any
other feature.
Deep learning techniques will be effective here
to classify tumour image without segmentation. A deep
neural network can be created with Convolutional nueral
network algorithm[17]. General architecture of
convolutional neural networks is shown in figure 6. In
deep learning , the feature is extracted from the entire
image automatically. Convolution in the CNN
architecture performs this operation. Number of feature
maps increases with the increase in CONV layer.
Reduction of dimension is required to initiate
training.Pooling layer down samples the feature
dimension. Fully connected layers manipulate the score of
each label. Softmax layers prepare the model with feature
and class score. The CNN architecture is slightly modified
in its dimension for training the brain tumor images.
The modified model architecture is listed in Table 1: Modified Model architecture
table 1.
The model is compiled in Keras with ‘Adam’
optimizer and ‘Binary cross entropy’ loss. Default
learning rate of 0.001 is used. The model is trained with
batch size of 32 for 24 epochs. For the test images our
trained model gives an accuracy of 95.6%.
For those images classified as having brain tumour,
the tumour location is detected using a combination of
multilevel thresholding, morphological operations and
contour extraction.

…… (1)

Figure 6 : General architecture of CNN

Where T is the mean of Maximum to Minimum intensities
of the image. Morphological Open function is used to
segment the regions.

.
 Morphological Open function is used to segment the images were collected from Kaggle dataset. The count of
regions. Contours of all regions are plotted and the region the data is insufficient for modelling a deep neural
with maximum area contains tumour region. network. So 2530 images have been created with
augmentation technique. The extracted cropped images
The Density of the tumour area can be estimated using are then resized to (240, 240) resolution. Keras (with
Gaussian kernel distribution.
Tensorflow backend) framework is chosen creating the
model. Two types of segmentation at different level are
implemented to analyse the performance of the system.
Segmentation was done before and after classification.
………. (2) From the performance analysis, segmentation after the
classification gives better result.
This algorithm is faster in execution for normal MRI
images. If it identifies the abnormal images, it goes to the
IV. RESULTS AND DISCUSSIONS next step ,ie : segmentation.
ROC curve shows the relation between sensitivity and
specificity .

Figure 7 : Normal Images Results

Figure 9 : ROC plot of Normal cases

Figure 10 : ROC plot of Abnormal cases

𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = (𝑇𝑃+𝑇𝑁)
= 0.98
(𝑇𝑃+𝑇𝑁+𝐹𝑃+𝐹𝑁)

𝑆e𝑛𝑠i𝑡𝑣i𝑡𝑦 = 𝑇𝑃
= 0.97
𝑇𝑃+𝐹𝑁

𝑆𝑝e𝑐ifi𝑐i𝑡𝑦 = 𝑇𝑁
= 0.99
𝑇𝑁+𝐹𝑃

Table 2: Performance Analysis

Algorithm Over all Accuracy

Nandpuru[18] 96.77%

El-Dahshan[19] 97%
(a) (b) (c) (d) Ibrahim[20] 96.33%
Figure 8 : Tumour detection results : (a) Input Image Rajini[21] 90%
(b) Abnormality Detection (c) Tumour region detection
(d) tumour mask for density estimation
Proposed Method 98%
Objective of the proposed system is to classify
malignant brain tumour from the MRI images. 253 MRI

.
 IV. CONCLUSION AND FUTURE SCOPE Approach”, 3rd International Conference on Trends in
Electronics and Informatics (ICOEI), Tirunelveli, India, 2019
This paper provides a new method for detecting [10] A. R. Mathew and P. B. Anto, “Tumor detection and
brain tumour by deep learning method. The early classification of MRI brain image using wavelet transform and
detection of cancer helps timely and effective treatment. SVM”, International Conference on Signal Processing and
Kaggle dataset contains good quality of MRI images for Communication (ICSPC), Coimbatore, 2017
research purposes. Different segmentation algorithms [11] W. Chen, X. Qiao, B. Liu, X. Qi, R. Wang and X. Wang,
were experimented. From this , multilevel thresholding
“Automatic brain tumor segmentation based on features of
and OTSU thresholding are the best methods for the
dataset. Convolutional Neural Network with modified separated local square”, Chinese Automation Congress (CAC),
approach helped to get a result with accuracy 98%. Jinan, 2017
Density estimation method is also proposed using [12] Navoneel Chakrabarty, “Brain MRI Images for Brain
Gaussian kernel distribution. Tumor Detection Dataset” , Kaggle , April 2019
[13] S. Poornachandra and C. Naveena, "Pre-processing of MR
This system can be improved to support with a web Images for Efficient Quantitative Image Analysis Using Deep
interface. Detection of different diseases can be also Learning Techniques," 2017 International Conference on Recent
identified from the MRI images. Apart from the density Advances in Electronics and Communication Technology
some other parameters can also estimated for therapeutic (ICRAECT), Bangalore, 2017, pp. 191-195, doi:
purposes. 10.1109/ICRAECT.2017.43.
[14] Mohammed Thanveersha N., et al. “Automatic Brain
REFERENCES Hemorrhage Detection Using Artificial Neural Network”,
[1] David N. Louis, Arie Perry, et al. , “The 2016 World Health International Journal of Innovations and Implementations in
Organization Classification of Tumors of the Central Nervous Engineering(ISSN 2454- 3489), 2019, vol 1
System: a summary” , Acta Neuropathol , Springer may 2016 [15] Soumya R S , et al. “Advanced Earlier Melanoma Detection
[2] Pär Salander, A Tommy Bergenheim, Katarina Hamberg, Algorithm Using Colour Correlogram”, 2016 International
Roger Henriksson, Pathways from symptoms to medical care: a Conference on Communication Systems and Networks
descriptive study of symptom development and obstacles to (ComNet) | 21-23 July 2016 | Trivandrum
early diagnosis in brain tumour patients, Family Practice, [16] J. A. Akhila, C, Markose , et al. "Feature extraction and
Volume 16, Issue 2, April 1999, Pages 143–148, classification of Dementia with neural network," 2017
[3] McKinney PA ,”Brain tumours: incidence, survival, and International Conference on Intelligent Computing,
aetiology”,Journal of Neurology, Neurosurgery & Psychiatry Instrumentation and Control Technologies (ICICICT), Kerala,
2004;75:ii12-ii17. India, 2017, pp. 1446-1450
[4] Heimans, J., Taphoorn, M. Impact of brain tumour treatment [17] Avigyan Sinha, Aneesh R. P., “Real Time Facial Emotion
on quality of life. J Neurol 249, 955–960 (2002) Recognition using Deep Learning”, International Journal of
[5] Malavika Suresh, et al. “Real-Time Hand Gesture Innovations and Implementations in Engineering(ISSN 2454-
Recognition Using Deep Learning”, International Journal of 3489), 2019, vol 1
Innovations and Implementations in Engineering(ISSN 2454- [18] H. B. Nandpuru, S. S. Salankar, and V. R. Bora, ‘‘MRI
3489), 2019, vol 1 brain cancer classification using support vector machine,’’ in
[6] M. Gurbină, M. Lascu and D. Lascu, “Tumor Detection and Proc. IEEE Students’ Conf. Electr., Electron. Comput. Sci., Mar.
Classification of MRI Brain Image using Different Wavelet 2014, pp. 1–6.
Transforms and Support Vector Machines”, 42nd International [19] E.-S.-A. El-Dahshan, T. Hosny, and A.-B.-M. Salem,
Conference on Telecommunications and Signal Processing ‘‘Hybrid intelligent techniques for MRI brain images
(TSP), Budapest, Hungary, 2019 classification,’’ Digit. Signal Process., vol. 20, no. 2, pp. 433–
[7] Somasundaram S and Gobinath R, “Early Brain Tumour 441, Mar. 2010.
Prediction using an Enhancement Feature Extraction Technique [20] W. H. Ibrahim, A. A. A. Osman, and Y. I. Mohamed,
and Deep Neural Networks”, International Journal of Innovative ‘‘MRI brain image classification using neural networks,’’ in
Technology and Exploring Engineering (IJITEE), ISSN: 2278- Proc. Int. Conf. Comput., Electr. Electron. Eng. (ICCEEE), Aug.
3075, Volume-8, Issue-10S, August 2019 2013, pp. 253–258.
[8] Damodharan S and Raghavan D, “Combining Tissue [21] N. H. Rajini and R. Bhavani, ‘‘Classification of MRI brain
Segmentation and Neural Network for Brain Tumor Detection”, images using knearest neighbor and artificial neural network,’’
The International Arab Journal of Information Technology, Vol. in Proc. Int. Conf. Recent Trends Inf. Technol. (ICRTIT), Jun.
12, No.1, January 2015 2011, pp. 563–568
[9] G. Hemanth, M. Janardhan and L. Sujihelen, “Design and
Implementing Brain Tumor Detection Using Machine Learning