3rd International Conference on Emerging Technologies in Computer Engineering: Machine Learning and Internet of Things (ICETCE-

2020), 07-08 February 2020, (IEEE Conference Record # 48199)

Non-invasive Brain Tumor Detection using

Magnetic Resonance Imaging based Fractal
Texture Features and Shape Measures
Manu Gupta K.Sasidhar
Department of Electronics and Computer Engineering,Sreenidhi Department of Electronics and Computer Engineering,
Institute of Science and Technology, Sreenidhi Institute of Science and Technology,
Hyderabad, India Hyderabad, India
[email protected] [email protected]

Abstract—This study presents a novel non-invasive versatile contrast between different brain tissues and does
quantitative feature set using magnetic resonance imaging not use ionizing radiations [6].
(MRI) for diagnosis of brain tumor and their grade
classification. Texture features using segmentation-based In routine clinical diagnosis, qualitative evaluation of
fractal texture analysis (SFTA), and selected shape measures MR images is carried by experienced radiologists to identify
were extracted from the segmented tumor volume to tumor location and size to perform grading of tumor as LG
differentiate low-grade (LG) tumor and high-grade (LG) or HG. But it is very difficult to perform brain tumor
brain tumor. Classification of the tumor grade is performed diagnosis by this approach based on qualitative analysis as
with support vector machine (SVM) classifier and testing and tumors vary largely in their appearance and shape. Also, this
training dataset are obtained using k-fold cross-validation approach is qualitative, prone to operator bias and is very
method. The brain detection method proposed in this study time-consuming.
gave an overall specificity and sensitivity of 86% and 88%
respectively. Also, an accuracy of 87% was achieved while In the recent past, quantitative techniques have been
classifying LG and HG brain tumor. developed for diagnosing brain tumor and its classification.
Researchers have extracted various features from brain
Index Terms— Brain Tumor, Feature classification, SFTA, images in order to classify them. Texture features extracted
Shape Measures, Texture feature, Magnetic Resonance Imaging using statistical approaches are used most frequently for
medical image analysis [7]. Gray-level co-occurrence matrix
I. INTRODUCTION (GLCM) parameters were used by Jafarpour et al. [8] for
Brain tumor is caused due to uncontrolled division of extracting features.Further, using an artificial neural
cells in the brain or around it. The presence of tumor at a network(ANN) classifier, MR images are classified as
given location in brain affects the activities controlled by normal and abnormal. Tessamma et al. [5] and Ain et al. [9]
that part of the nervous system. Tumor growth can displace presented the system based on GLCM features and first-
the surrounding brain tissue leading to a shift in brain order histogram parameters for the diagnosis of abnormality
midline structures [1]. Brain tumors can also cause an in MR images. Singh et al. [10] extracted texture properties
increase in intracranial pressure and expansion of ventricles. from brain MRI slices with gray level run length matrix
The other effects of brain tumor include behavioral changes, parameters. The SVM classifier is used in this study to
poor coordination of limbs, memory loss etc. [2]. Brain classify brain MR images. These systems have a major
tumor is one of the main causes of increase in death rates drawback that these are computationally expensive.
among adults and children around the world and constitutes The authors in [6] and [11] used a spectral approach for
about 1–2% of all malignancies [3]. The statistics from the measuring texture feature set from brain MR images by
Central Brain Tumor Registry of the United States, applying discrete wavelet transform (DWT) method. Neural
CBTRUS, stated that brain tumors have exceeded leukemia network based classifier was used in study by Dahshan et al.
and are the major causes of cancer deaths in children of 0- [6] todifferentiate normal and abnormal MR
19yrs in the United States [4]. Hence, it is imperative to images.However, SVM is utilized in [11] for classifying
develop methods for the early diagnosis and treatment of tumors. Statistical and non-statistical approaches were
brain tumors to obtain a hopeful prognosis. combined by Kumar et al. [12] for texture features analysis
Brain tumors are generally classified as high grade (HG) of brain tissues. In this model DWT followed by the GLCM
and low grade (LG) based on their growth patterns and method is applied to compute texture features from brain
histological characteristics. LG tumors grow slowly, are images. Lastly, the detected brain tumors were classified as
non-cancerous (benign) and generally have well-defined benign or malignant using SVM. The limitation of DWT
borders. HG tumors grow rapidly, have ill-defined borders based systems is that were not able to produce high
and are cancerous (malignant) in nature [5]. Various classification accuracy because of shift-variance and poor
imaging methods are used to identify brain tumor, such as directionality property of DWT [13].
magnetic resonance imaging (MRI), positron emission In this study, a novel feature set for brain tumor
tomography (PET),computed tomography (CT), etc. MRI is diagnosis including texture features extracted using fractal
the most commonly used method for diagnosing brain analysis (SFTA based) and selected shape measures is
tumors among various imaging modalities. It offers a proposed. The proposed features are computed from the

978-1-7281-1683-9/20/$31.00 ©2020 IEEE

Authorized licensed use limited to: ANNA UNIVERSITY. Downloaded on May 27,2020 at 11:20:18 UTC from IEEE Xplore. Restrictions apply.
segmented tumor volume obtained using an algorithm based inhomogeneous variations in intensity reduce the quality of
on the symmetry characteristics of the brain. The tumorous image and blur it, which results in decreasing the
brain MR images were classified as benign (LG) or information corresponding to high-frequency regions.
malignant (HG) using k-fold cross-validation based linear Therefore to enhance the results of tumor segmentation and
SVM classifier, with the extracted feature set as input. It is feature extraction process, firstly, the process of bias
demonstrated that the new feature set works efficiently with correction was carried out on brain MRI volumes. FMRIB
higher accuracy for distinguishing different grades (LG or Software Library (FSL) [16] was used to perform bias
HG) of brain tumor. correction of MRI image volumes.
II. METHODOLOGY C. Brain Tumor Segmentation
The methodology implemented in proposed study for Segmented tumor volume from brain MR Images was
the brain tumor diagnosis and classification is shown in Fig. obtained using an approach presented in our previous work
1. [17]. T2-weighted MR images were used for separating
tumor region. The method is described briefly in this
section.
First, the tumor regions in the brain were located using
the symmetry characteristic of the brain. In this process, the
brain was split into two symmetric halves using axial planes.
These half brain regions were then compared voxel-wise to
identify tumor tissues by applying subtraction operation.
The non-symmetric regions appeared with high intensity in
output images. Next, contrast stretching process using piece-
wise operation was carried on the subtracted images to
improve the differentiation among tumorous regions and
other areas using the following equation:

Ice(p,q)
( , )

where I(p,q) represents the input image, E signifies the slope

function and L refers to mid-line of moving from dark to
bright gray levels.
Lastly, a region having a maximum area was identified
amongst all the regions present in contrast enhanced image.
This region was then labeled as a tumorous region while
remaining regions were masked with zero gray value and
merged with the background. This process was implemented
on all the brain MRI slices so as to obtain complete
segmented tumor volume for each brain tumor
patient.Fig.2b shows the segmentation results for one of the
brain tumor patients (in one of the brain volume MRI slices).
Fig. 1. Methodology followed for brain tumor diagnosis
The output from the segmentation stage was a binary
A. Image Database
mask with a foreground having intensity ‘1’ as separated
The volumes of brain tumor MRIused in this study have tumor region and background have intensity ‘0’, for all the
been taken from the database of MICCAI 2012 Challenge brain MRI images considered. This segmented tumor mask
[14]. These images have been obtained from University of is denoted as ‘MB(p,q)’.
Bern, ETH Zurich, University of Debrecen and University
of Utah. In database, T1-weighted, T1 contrast enhanced
(T1-CE), FLAIR and T2-weighted MR images were
available for all the patients, which were co-registered to the
T1-CE image. The database also contains the ground truth
for the tumor. T2-weighted MR images provide better
contrast for tumor tissues [6,15] and hence are used
frequently by researchers for brain tumor diagnosis
[6,8,11,15].Thus, for the present study, T2-weighted brain
MR images are used for tumor diagnosis and its grade
classification. Fig. 2. (a) Preprocessed T2-weighted image; (b) segmentation results; (c)
segmented image multiplied with the T2-weighted image, for one of the
B. Preprocessing MRI slices.
Radiofrequency (RF) field non-uniformity, known as RF
in homogeneity or bias field, produces signal intensity
deviations that lead to distorted MR images. These

94

Authorized licensed use limited to: ANNA UNIVERSITY. Downloaded on May 27,2020 at 11:20:18 UTC from IEEE Xplore. Restrictions apply.
D. Feature Extraction • Eccentricity was computed as the ratio of major
The texture of tumor region and its shape were important axis length to the minor axis length for the region
parameters that have been used to identify tumor grade and considered. It has a value between zero and one.
its degree of malignancy. The following sub-sections • Perimeter was calculated by computing the
explain the feature set extracted in the proposed study. distance between each neighboring pair of pixels
along the border of the separated tumor region.
1) Fractal based Texture Features: Texture features
were computed from the separated tumor volume by The proposed feature set, consisting of SFTA algorithm-
applying the method of segmentation-based fractal texture based texture features and shape measurements, was then
analysis (SFTA) [18]. The segmented tumor mask utilized for classifying the tumorous MRI volumes as LG or
‘MB(p,q)’ obtained in tumor segmentation process has been HG.
multiplied with the respective bias-corrected T2-weighted E. Feature Classification
slices of MR images (refer Fig.2c). This volume of the The support vector machine (SVM) classifier was
output image is called as ‘MI(p,q)’ and is used for the utilized in proposed study to classify the input brain MRI
feature extraction process. volumes as low grade (benign) or high grade (malignant).
To extract texture features, gray-scale images of In SVM classifier the input data is separated into two
segmented tumor volume ‘MI (p,q)’ were further different classes by identifying an optimal hyper-plane.
decomposed into binary images by implementing This hyper-plane distinguishes data points of one class from
consecutive thresholding operations. In this study, another with maximum distance. The decision rule for
thresholding values were obtained by applying the multi- separating the classes was obtained from the training set
level Otsu algorithm, so as to reduce the intra-class variance whose classes are known in prior. The hyperplane obtained
in the input image. This algorithm was implemented until from decision rule was then applied for identifying the class
the chosen number of thresholds (a user-defined of data in the testing set [20]. SVM is suitable for the small-
constraint) was reached. Let these thresholds be = sample dataset as it can effectively solve local optimal and
, ,…… . over-fitting problem. SVM basically maps the input vectors
into some high dimensional feature space using certain
Next, from these thresholds, pairs of lower ( ) and functions known as kernels. A linear function is used as a
upper ( ) thresholds were selected for generation of kernel in this study to predict the grade of brain tumor in
binary images using the following equation: input MRI volumes.

1 < ( , ) ≤ The K-fold cross-validation method was applied in

( , )= (2) proposed study to divide the input dataset into testing and
0 ℎ
training set [20]. It divides the entire input dataset into k
For these binary images, thresholds were selected from different subsets. The classifier was tested on one of the k-
two sets and , as described in (3). subsets and trained on the remaining k-1 subsets. This
procedure has been reiterated k-times to ensure that each
data point is in the test data set at least once and in the
= ∪ = , (3) training data set at k-1 times. In this way, K-fold validation
procedure for feature classification makes full use of the
where ∈ and represent the maximum gray level in the limited sample dataset. It also ascertains that for training
input image. In this way, 2 binary images were generated and validation of classifier whole dataset is considered.
using two threshold decomposition algorithm. The main advantage of implementing the K-fold cross-
Feature vector was extracted from the generated binary validation process for classifying tumor is that the accuracy
images which comprise of the following parameters: mean and other parameters for evaluating classifier performance
value, the area of the binarized image, and fractal were obtained by averaging them across all k trials.
dimension (FD) of their boundaries. To determine FD for III. PERFORMANCE METRICS
the boundary of binarized images box-counting technique
was used [19], which gives information about the The performance evaluation metrics used for the
complexity of segmented tumor region. quantitative assessment of the proposed brain tumor
detection technique are described in this section. To
In this study, was considered as three, resulting in the compute evaluation parameters, the confusion matrix was
generation of six binary images. Each of these binary determined from k-fold SVM classifier output as shown in
images was characterized by three above mentioned Table 1. The parameters, TP and TN in Table 1 represent
features. In this way, 18 features were produced for each the quantity of positive and negative cases identified
image using the SFTA technique. correctly by the classifier. Furthermore, the misclassified
samples are represented as FP and FN variables in Table 1.
2) Shape Measures: Eccentricity and perimeter were
measured from segmented tumor volume to characterize • Accuracy – It gives the percentage of correct cases
the tumor shape. predicted by classifier over the total cases
considered for classification.

95

Authorized licensed use limited to: ANNA UNIVERSITY. Downloaded on May 27,2020 at 11:20:18 UTC from IEEE Xplore. Restrictions apply.
 Accuracy= × 100 (4) and tumor grade classification (LG /HG) was performed.
The results obtained for various performance metrics are
shown in Table 2.
• Sensitivity- It gives the percentage of true positives
the total positive cases for the given dataset. The performance of feature set extracted in this study
was compared with the method presented by Kumar et
Sensitivity = × 100 (5) al.[12]. The discrete wavelet transforms (DWT) and GLCM
( )
based features were extracted in method presented by
Kumar et al. The classification results obtained for this
• Specificity- It gives the percentage of true negatives method using SVM classifier are summarized in Table
to the total negative cases for the given dataset. 2.The feature set extracted in the proposed method
comprising of SFTA features and shape parameters
Specificity = × 100 (6) produced an accuracy of 87%. In addition, the proposed
( )
model achieved a sensitivity of 88% and a specificity of
• Positive Predicted value (PPV)- It gives a fraction of 86%. These results were significantly better when
positive predictions that were correct. compared to GLCM based features considered by Kumar et
al. [12] that achieved accuracy of 73%. The GLCM
approach used in method presented by Kumar et al. [12] is
PPV = (7)
( ) computationally expensive and generates redundant
features. The improved results from proposed method is
• Negative Predicted value (NPV)- It gives a fraction obtained by use of the binarized images generated using the
of negative predictions that were correct. SFTA approach that assess the texture differences in LG
(benign) and HG (malignant) tumors effectively. Moreover,
NPV = (8) the hybrid feature set including texture and shape measures
( )
extracted in the proposed study provides comprehensive
information on texture heterogeneity and morphology of
In this study, positive represents HG (malignant) tumor
tumors and hence improves its classification efficiency. The
patients and negative represents LG (benign) tumor results obtained demonstrate that the features extracted for
patients. brain tumor diagnosis can efficiently classify high grade
and low grade tumors and can help the radiologists to
TABLE I. CONFUSION MATRIX
quantitatively assess brain tumors non-invasively for their
diagnosis.
Actual Condition
from Predicted tumor class by Classifier
TABLE II. PERFORMANCE METRICS USING 5-FOLD CROSS-
Ground Truth
VALIDATION BASED SVM CLASSIFIER

Positive Negative
Method

True Positive False Negative S.No. Parameters

Positive Proposed Method Kumar et al.[12]
(TP) (FN)
(SFTA texture+ (DWT+
shape measures) GLCM features)
False Positive True Negative
Negative
(FP) (TN) 1 Accuracy 87 % 73 %

2 Sensitivity 88 % 75 %

IV. RESULTS AND DISCUSSION

3 Specificity 86 % 71 %
This section summarizes the results obtained from the
proposed model for brain tumor diagnosis. MRI images
4 PPV 0.88 0.75
(T2-weighted images) of 15 brain tumor patients from
MICCAI 2012 Challenge database were used in this
proposed study. This includes 8 HG tumor patients and 7 5 NPV 0.86 0.71
LG tumor patients.

V. CONCLUSION
The proposed feature set containing texture and shape
measures was calculated over the complete segmented This study provides a novel non-invasive quantitative
tumor volume, for all the tumor patients using MATLAB feature set for brain tumor diagnosis comprising of fractal
2013a. To assess the efficacy of brain tumor diagnosis texture features (SFTA based) and shape measures. T2-
model presented in this study, the extracted features were weighted brain MR images were utilized to segment tumor
fed to k-fold cross validation (k=5) based SVM classifier, region and classify them as LG or HG. Our method is

96

Authorized licensed use limited to: ANNA UNIVERSITY. Downloaded on May 27,2020 at 11:20:18 UTC from IEEE Xplore. Restrictions apply.
giving significantly better results in comparison to other based segmentation and texture based ensemble classification of
brain tumor,” Applied Soft Computing, vol. 21, pp. 330–340, Aug.
state of art method from literature. The use of hybrid
2014.
feature set (texture and shape) in our study enhances the [10] A. Singh, “Detection of Brain Tumor in MRI Images , using
classification efficiency of the proposed system for brain Combination of Fuzzy C-Means and SVM,” in 2nd IEEE
tumor diagnosis. The proposed feature set can assist the International Conference on Signal Processing and Integrated
Networks (SPIN), Feb 19, 2015, IEEE, pp. 98–102.
radiologists and clinicians in non-invasive detection of
[11] S. Kumar, C. Dabas, and S. Godara, “Classification of brain MRI
brain tumor. tumor images: A hybrid approach,” Procedia computer science, vol.
122, pp. 510–517, Jan. 2017.
ACKNOWLEDGMENT [12] P. S. Kumar and S. Chatteijee, “Computer aided diagnostic for
cancer detection using MRI images of brain (Brain tumor detection
The authors would thank Director, Sreenidhi Institute of and classification system),” in Proceedings of the IEEE Annual
Science and Technology, Hyderabad for providing research India Conference (INDICON), 2016, pp. 1–6.
facilities. [13] A. Khademi and S. Krishnan, “Shift-invariant discrete wavelet
transform analysis for retinal image classification,” Medical &
biological engineering & computing, vol. 45, no. 12, pp. 1211–
1222, 2007.
REFERENCES [14] B. H. Menze et al., “The multimodal brain tumor image
[1] H. Sontheimer, “A role for glutamate in growth and invasion of segmentation benchmark (BRATS),” IEEE transactions on medical
primary brain,” Journal of neurochemistry, vol. 105, no. 2, pp. 287– imaging, vol. 34, no. 10, pp. 1993-2024, Dec 2015.
295, 2008. [15] P. Y. Lau, F. C. Voon, and S. Ozawa, “The detection and
[2] P. L. Larrubia et al., “Magnetic Resonance Imaging of Gliomas,” in visualization of brain tumors on T2-weighted MRI images using
Advances in the Biology, Imaging and Therapies for Glioblastoma, multiparameter feature blocks,” in 27th Annual International
Prof. Clark Chen, Eds. InTech, China, 2011, pp. 225-250. Conference of the Engineering in Medicine and Biology Society,
[3] A. Munshi, “Central nervous system tumors: Spotlight on India,” IEEE-EMBS, 2006, 5104-5107.
South Asian journal of cancer, vol. 5, no. 3, pp. 146–147, 2016. [16] S. M. Smith et al.,“Advances in functional and structural MR image
[4] “Brain Tumor Statistics - American Brain Tumor Association,” analysis and implementation as FSL,” Neuroimage, vol. 23, pp.
[Online]. Available: http://www.abta.org/about-us/news/brain- S208–S219, Jan. 2004.
tumor-statistics/. [17] M. Gupta, V. Rajagopalan, E.P. Pioro, and B.P. Rao, “Volumetric
[5] S. A. Resmi and T. Tessamma, “Texture description of low grade analysis of MR images for glioma classification and their effect on
and high grade glioma using statistical features in brain MRIs,” brain tissues,” Signal, Image and Video Processing, vol. 11, no. 7,
International Journal of Recent Trends in Engineering and pp. 1337-1345, 2017.
Technoogy, vol. 4, no. 3, pp. 27–33, Nov. 2010. [18] A. F. Costa, G. Humpire-Mamani, and A. J. M. Traina, “An
[6] E. S. A. El-Dahshan, H. M. Mohsen, K. Revett, and A. B. M. Efficient Algorithm for Fractal Analysis of Textures,” in
Salem, “Computer-aided diagnosis of human brain tumor through Proceedings of IEEE 25th SIBGRAPI Conference on Graphics,
MRI: A survey and a new algorithm,” Expert systems with Patterns and Images, 2012, pp. 39–46.
Applications ,vol. 41, no. 11, pp. 5526–5545, Sep. 2014. [19] B. S. Raghavendra and N. D. Dutt, “Computing fractal dimension of
[7] F. Davnall et al., “Assessment of tumor heterogeneity : an emerging signals using multiresolution box-counting method,” International
imaging tool for clinical practice ?,” Insights into Imaging, vol. 3, Journal of Information and Mathematical Sciences, vol. 6, no. 1, pp.
no. 6, pp. 573–589, Dec. 2012. 50-65, 2010.
[8] S. Jafarpour, Z. Sedghi, and M. C. Amirani, “A Robust Brain MRI [20] A. A. Hasseim, R. Sudirman, and P. I. Khalid, “Handwriting
Classification with GLCM Features,” International Journal of classification based on support vector machine with cross
Computer Application, vol. 37, no. 12, pp. 1–5, 2012. validation,” Engineering, vol. 5 , no. 5, pp. 84-87, 2013.
[9] Q. Ain, M. A. Jaffar, and T. S. Choi, “Fuzzy anisotropic diffusion

97

Authorized licensed use limited to: ANNA UNIVERSITY. Downloaded on May 27,2020 at 11:20:18 UTC from IEEE Xplore. Restrictions apply.