See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/358397341

Detecting Brain Tumor by Using Machine Learning and Image Processing

Techniques

Article · February 2022

CITATIONS READS

0 3,576

2 authors:

Ahmet Safa Karakoç Ahmet Yasir Cilvez

Abdullah Gul University Abdullah Gul University
8 PUBLICATIONS 7 CITATIONS 2 PUBLICATIONS 0 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Ahmet Safa Karakoç on 06 February 2022.

The user has requested enhancement of the downloaded file.

Detecting Brain Tumor by Using Machine Learning
and Image Processing Techniques

Ahmet Safa KARAKOÇ Ahmet Yasir CILVEZ

Electrical and Electronical Electrical and Electronical
Engineerings Department Engineerings Department
Abdullah Gul University Abdullah Gul University
Kayseri/Turkey Kayseri/Turkey
ahmetsafa.karakoc@agu.edu.tr ahmetyasir.cilvez@agu.edu.tr

Abstract— In this article, classification of two types of brain Gray level co-occurrence matrix (GLCM) was used to the
tumors explained. In the method of the research, image filtering, extract the second order statistical texture features for motion
segmentation, enhancement is used in image processing part and estimation of images which are introduced in [3]. In that
support vector machine is used in machine learning part. As a project, Angular second moment, Correlation, Inverse
result of this research, test images are classified into two classes
with minimum 62% accuracy rate.
difference moment and Entropy features are extracted by
using Vivado FPGA. These features have high classification
Keywords— Brain tumor, GLCM, SVM, image processing, accuracy and require less extraction time that means they
machine learning have some advantages. To conclude; GLCM is one of the
most efficient feature extraction methods.
I. INTRODUCTION
Brain tumors are one of the most dangerous illnesses that IBkLG classifier has been used to classification of normal,
can be seen all people in every age. The reason brain tumors benign and malignant tumors in brain MRI images by
are highly dangerous is that they are located in the nervous developing an automated method [4]. The IBkLG method is
system. Each year; approximately 12000 people are one of the successful methods that reach approximately 100%
diagnosed by this disease. According to Davies [1], between classification accuracy because of handling real valued
the 20-44 age, while the rate of survival is 50%, after 65 years features. In this proposed method; the system has ability to
old the survival rate decreases to 5%. Because of the get rid of noisy images by using median filtering. For the
importance of brain tumors, diagnosing earlier will help to preprocessing, the filtered image is converted gray to RGB
intervene rapidly. In addition, by using machine learning and and to l*a*b color space. After these steps; the image has
image processing methods, the missed diagnoses cases can be transferred to color-based segmentation process that make the
detected truly. Brain tumor detection is possible by images ready to feature extraction. For the feature extraction,
transferring MRI images to digital software. Transferred GLCM, Histogram of oriented gradients and canny edge
images can be classified according to tumor size and its detection has been used. The extracted features’ mean,
location. standard deviation, skewness, Root mean square value, and
median absolute values are calculated. From this values a
In this project; We will detect the brain images in our dataset table has been created to use in IBkLG
selected dataset using image processing and classify the classifier.IBkLG classifier has reached 86.6% classification
diagnosed tumors with machine learning techniques by using accuracy.
MATLAB. In literature review part; there are studies about
brain tumor detection and classification methods which were DWT+PCA+SVM is a developed model to detect and
useful for this project classify the normal and abnormal MR images of brain. [5].
DWT stands for discrete wavelet transform that extract the
information of the brain images without any loss. This
II. LITERATURE REVIEW method has advantages that can able to keep both frequency
K-means clustering is one of the pixel-based segmentation and location information. PCA stands for principal
methods that is introduced in [2]. This method is applied on component analysis. PCA is used the reduce the dimensions
two different data sets which taken from the gray feature of the images which will cause the worsened classification
vectors of the original MR brain images and the converted accuracy and heavy computation burden. For the SVM
RGB features to detect the tumor objects in magnetic section, kernel SVMs types has been tried for classification.
resonance (MR) images. Color translation, K-means Among them, GRB SVM has reached the most accurate
clustering and histogram clustering are the applications that method for the classification. As a conclusion,
combined in this method. Because of the not causing extra DWT+PCA+SVM method’s accuracy level is 99.38%.
overhead of the color space translation, this method is easy to
applying and is efficient. Skull stripping is one of the preprocessing methods for the
MRI brain images [6]. This method is the process of the brain
tissue segmentation from surrounding region that are the skull

XXX-X-XXXX-XXXX-X/XX/$XX.00 ©20XX IEEE

and nonbrain area. Morphology based methods has been used accuracy. So, we took the MRI images that is taken by same
to separate the brain area from skull. This method has 5 steps. shooting angle.
These are, image thresholding, morphological operators that
fill the gaps in extracted object, edge detection, creating a IV. METHODOLOGY
binary mask and combining the binary mask. As a a. Image Processing (Preprocessing)
conclusion, this algorithm has provided more efficient MRI Before starting classification, some methods have to be
images that are not affected from surrounding region of the applied in Image Processing to make the image workable.
brain. Firstly, skull stripping method (Figure4) has been
implemented to remove the surrounding area of brain. This
III. DATASET DEFINITION method helps to focus on the brain tumor part. Secondly,
We took the dataset from Kaggle [7]. In this dataset; there median filtering (Figure5) has been applied to remove the
are 3 different tumor types that are glioma, meningioma and noises from the brain MR images via built-in function in
pituitary and normal brain images. The description of the MATLAB. After the filtering is completed, normalization
tumor types will be introduced in order. has been applied. Normalization has been used to change the
range of pixel intensity values. Next methods after these steps
Gliomas are a general name used to describe a group of are image enhancement and image segmentation. Image
tumors that occur in the glial cells, the supporting tissue of enhancement is used to adjust the digital images to make
the brain. Symptoms include headache, seizures, nausea, and them more suitable for displaying. For this implementation
vomiting. (Figure1). Meningiomas are slow-growing and “imadjust” function has been used in MATLAB. Image
mostly benign tumors that arise from the membranes segmentation’s goal is to change the representation of image
surrounding the brain. Although they are not technically a to make easier to analyze and more meaningful. As a final
brain tumor since they do not originate in the brain, they step morphology operator has been applied to brain images.
cause a number of symptoms by creating pressure on the In a morphological operation, the value of each pixel in the
brain with their growth. (Figure2). Pituitary tumors are output image is based on a comparison of the corresponding
abnormal growths that develop in your pituitary gland. Some pixel in the input image with its neighbors. (Figure 6) [11].
pituitary tumors result in too much of the hormones that After all steps completed, our MRI images are ready for
regulate important functions of your body. Some pituitary GLCM that will extract the features of the images.
tumors can cause your pituitary gland to produce lower levels
of hormones. (Figure3).

Figure 4 Skull Stripping

Figure 1 Figure 2

Figure 5 Median Filtering

Figure 3
Dataset has been separated by 2 files that are training
images and test images. In training image file, there are 841
different images with different types. In testing image file,
there are 156 different images. For our project we did not use
all images. Because the shooting angle of the images are
different and to use these images together will decrease the
 Autocorrelation, Contrast, Correlation, Correlation,
Cluster Prominence, Cluster Shade, Dissimilarity, Energy,
Entropy, Homogeneity, Homogeneity, Maximum
probability, Sum of squares, Sum average, Sum variance,
Sum entropy, Difference variance, Difference entropy,
Information measure of correlation1, Information measure of
correlation2, Inverse difference (INV), Inverse difference
normalized (INN), Inverse difference moment normalized.

Table 1 GLCM Features Values

Figure 6 Figure 7

c. Classification (SVM)
There are many options for classification in Machine
learning. To select the best option for this project, we used
classification learner app in MATLAB. We trained our
dataset via this application. The highest accuracy has reached
Figure 8 with Kernel SVM. To train and cross validate fitcsvm
function has been used. The common syntax is;
b. Feature Extraction (GLCM Texture)
When the image processing part is completed, to extract
the features from the images GLCM Texture has been used. SVMModel=fitcsvm(X,Y,'KernelFunction','rbf',
Gray level cooccurrence matrix is a statistical method that ...'Standardize',true,'ClassNames',{'negClas
examines the texture by considering the spatial relationship s','posClass'});
of pixels.[6] The GLCMs are stored in a three-dimensional
matrix that (i x j x n). n represents the number of GLCM’s In our SVM model Gaussian RBF (Radial Basis
calculated where n is the number of GLCMs calculated. Function) is used. RBF is stands for Radial Based Function.
RBF kernel is a function whose value depends on the distance
from the origin or from some point.[10]

Figure 9 GLCM Algorithm

By implementing GLCM, some characteristic features are Figure 10 RBF Kernel

extracted. Contrast, Correlation, Energy, Homogeneity are V. RESULTS AND DISCUSSION
the most important features among them. Contrast calculates
the local variations of the GLCM matrix. Correlation The methods in Methodology section were applied to the
calculates the joint probability of occurrence of the pairs. two different data which contain Glioma and Meningioma
Energy shows the sum of squared elements in GLCM. tumor MRI images. According to results of the algorithm, the
Homogeneity calculates the closeness of the distribution of accuracy rate shows that the algorithm classifies the tumor
elements in GLCM [8],[9]. type. However, the rate is less than the expected value. One
of the possible reasons of this issue is number of low-quality
In this project, we extracted 22 different features of each images in the dataset. Moreover, because of the MRI nature,
brain images in our dataset. We created an Excel file to keep all of the images have different perspective even though some
these values. These values will be used for classification. are similar to each other. Another possible reason of the low
These features are listed below; accuracy is image processing part. Skull stripping algorithm
remove some important part of minority of the total images VII. REFERENCES
and if a K-means algorithm for segmentation due to the gray [1]https://www.medicinenet.com/brain_tumor_symptoms/vi
scale of the image is used successfully, the results can be ews.htm
more trustable. Even though we used first K-means algorithm
to detect tumor and working on just tumor instead of rest of [2] Ming –ni wu1 chia- chin lin2 chin-cheng-chan3, “Brain
the brain, there are some disadvantages have occurred and it Tumor Detection Using Color-Based K-Means Clustering
didn’t work every MRI images. That’s why, it was eliminated Segmentation” December 2007
from the algorithm. DOI:10.1109/IIHMSP.2007.4457697, PP 1-3
[3] P. Mohanaiah1, P. Sathyanarayana2, L. GuruKumar3,
“Image Texture Feature Extraction using GLCM Approach”
International Journal of Scientific and Research Publications,
Volume 3, Issue 5, May 2013 1 ISSN 2250-3153, PP 1-2
[4] Chauhan, S., More, A., Uikey, R., Malviya, P., & Moghe,
A.A. (2017). Brain tumor detection and classification in MRI
images using image and data mining. 2017 International
Conference on Recent Innovations in Signal processing and
Embedded Systems (RISE), 223-231.
[5] Yu-Dong Zhang and Lenan Wu, "An Mr Brain Images
Figure 11 Spreading of Classifier via Principal Component Analysis and Kernel
Autocorrelation vs Contrast Support Vector Machine," Progress in Electromagnetics
Research, Vol. 130, 369-388, 2012.
[6] Swiebocka-Wiek J. (2016) Skull Stripping for MRI
Images Using Morphological Operators. In: Saeed K.,
Homenda W. (eds) Computer Information Systems and
Industrial Management. CISIM 2016. Lecture Notes in
Computer Science, vol 9842. Springer, Cham.
https://doi.org/10.1007/978-3-319-45378-1_16
[7]https://www.kaggle.com/sartajbhuvaji/brain-tumor-
classification-mri?select=Training

Figure 12 Spreading of Entropy vs [8] Avinash Uppuluri (2021). GLCM texture features
(https://www.mathworks.com/matlabcentral/fileexchange/22
Correlation
187-glcm-texture-features), MATLAB Central File
In order to increase the accuracy rate, MATLAB’s own Exchange. Retrieved June 9, 2021.
algorithm that generated from ‘Classification Learner’ can be
[9] https://www.mathworks.com/help/images/texture-
used. Even though the accuracy rate of Support Vector
analysis-using-the-gray-level-co-occurrence-matrix-
Machine algorithm is more and more high, the results of our glcm.html
SVM algorithm gave less accuracy rate. Further, the number
of GLCM texture features can be rearranged to obtain more [10] https://towardsdatascience.com/support-vector-
accurate results. machines-svm-c9ef22815589
[11]
https://www.mathworks.com/help/images/morphological-
VI. CONCLUSION
dilation-and-erosion.html
As a conclusion; different types of brain tumor images
have been detected and classified successfully by using image
processing and machine learning techniques. Firstly, image
processing techniques has been applied to make the image
suitable for the classification. Secondly, GLCM Texture
Features method has been used to extract the characteristic
features of the brain images. Finally, RBF Kernel SVM
provides us to classify the brain images. To testing the
machine learning algorithm, 200 training examples of each
two group were used and algorithm was trained. The accuracy
rate obtained as minimum 62%.

View publication stats