Binary Particle Swarm Optimization (BPSO) based

Feature Selection from MRI Images for Automatic Brain

Tumor Detection

Presented By: Supervised By: Co-Supervised By:

Md. Mahmudul Hasan A. S. M. Shafi M. M. Imran Molla
Batch: 3rd , ID: 2016105050012 Lecturer Lecturer
Dept. of CSE, KYAU Dept. of CSE, KYAU Dept. of CSE, KYAU
01 Background Study
1. Bahadure, N. B., Ray, A. K., &Thethi, H. P.. Image analysis for MRI based brain tumor detection and feature
extraction using biologically inspired BWT and SVM. International journal of biomedical imaging, 2017.
2. https://braintumor.org/brain-tumor-information/brain-tumor-facts/
3. https://www.aans.org/en/Patients/Neurosurgical-Conditions-and-Treatments/BrainTumors
4. https://www.kaggle.com/navoneel/brain-mri-images-for-brain-tumor-detection.
5. https://en.wikipedia.org/wiki/Gray_level_size_zone_matrix.
6. https://pyradiomics.readthedocs.io/en/latest/features.html#module-radiomics.glcm.
7. https://pyradiomics.readthedocs.io/en/latest/features.html#module-radiomics.glrlm.
8. https://pyradiomics.readthedocs.io/en/latest/features.html#module-radiomics.glszm
9. https://pyradiomics.readthedocs.io/en/latest/features.html#module-radiomics.gldm.
10. B. Chakraborty, “Feature subset selection by particle swarm optimization with fuzzy fitness function,” in 3rd
International Conference on Intelligent System and Knowledge Engineering (ISKE’08), vol. 1, 2008, pp. 1038–
1042.
Overview
• Motivation
• Research Questions
• Objectives
• Clinical Background
• System Architecture
• Dataset
• Pre-processing
• Feature Extraction
• Feature Selection
• Classification
• Cross Validation
• Experimental Results
• Comparison with Existing Methods
• Discussion
• Conclusion
• Limitations
• Future Work
3 Motivation

 Brain Tumor is the second leading cause of cancer related death in male and fifth
leading in female.
 The incidence of the brain tumors has been increasing in all ages in recent decades.
 The standardization of age in different age in different countries is between 0.01 and
12.7 in males and 0.01 and 10.7 in women, per 1,00,000 people.
 The occurrence of brain tumors is numerously lower in East Asia, Southeast Asia, and
India.
 Most of the occurrence has been found in Europe, Canada, the USA, and Australia.
4 Research Questions
The specific problem statement for this thesis is:
“Can machine learning be used for brain tumor detection from MRI images? ”
However, the following research questions would facilitate the achievement of this
thesis:
 Is this approach to detect brain tumor properly?
 Is the system’s accuracy acceptable?
 Can the method help the clinical supervisors and radiologists?
5 Objectives

The primary objective of this thesis is :

 Effectively use machine learning to successfully detect brain tumor using MRI
images.
 Improve the accuracy by using feature selection and SVM classifier.
6 Clinical Background
Group of irregular cells in a brain is called ‘Brain Tumor’. There are two main types of
tumors : benign (Non Cancerous) tumor and Malignant (Cancerous) tumor.

Benign (Non
A malignant brain
Benign tumors tend to Malignant tumor is a cancerous
01 Cancerous)
tumor stay one place and do 02 (Cancerous)
growth in the brain.
tumor
not spread. It is slow Most malignant tumors
growing and fewer are secondary cancers,
dangerous. It will not which means they
usually come back if started in another part
all of the tumor can be of the body and spread
safely removed during to the brain.
surgery.
7 System Architecture
Input Images Input Images MRI Image
-------------------------------------------------------------------------------------------
Pre-processed
Preprocessing Preprocessing Image
--------------------------------------------------------------------------------------------
• GLCM
• GLRLM
Feature Extraction • GLSZM
Feature Extraction • DLDM  SVM: Support Vector Machine
-----------------------------------------------------------------  GLCM: Gray Level Co-occurrence
Matrix
Feature Selection BPSO
 GLRLM: Gray Level Run Length
----------------------------------------------------------------- Matrix
 GLSZM : Gray Level Size Zone
SVM
Classification Classification Matrix
-----------------------------------------------------------------
 GLDM : Gray Level Dependence
• Accuracy
Matrix
Assessment • Sensitivity
• Specificity  BPSO : Binary Particle Swarm
Figure 1: A framework of the proposed approach. Optimization
8 Proposed Algorithm
Algorithmic Steps:

1. Take MRI image as a input images.

2. Pre-process all input images into a NRRD format using 3D slicer.
3. For extracted the feature, execute texture feature algorithm on the pre-processed
image from step 2.
4. Use feature selection algorithm on the extracted feature from step 3.
5. Generate a Feature Vector (FV) for each training image.
6. Implement the classifiers (SVM) to test the train image for classification and
evaluate the performance of the results.
9 Dataset

“The proposed system utilizes 100 MRI images. 49 of them are normal images and 51
are brain tumor images”.
 Sources of database :
 KaggleDataset (
https://www.kaggle.com/navoneel/brain-mri-images-for-brain-tumor-detection).

(a) (b)

Figure 2: Sample dataset: (a) normal brain, (b) brain with tumor
10 Pre-processing
The original image is converted into Nearly Raw Raster Data (NRRD) format. NRRD
image is the standard file format for medical imaging and also NRRD files do not
contain sensitive patient information

(a) (b)

Figure 3: Preprocessing: (a) input image, and (b) pre-processed image .

11 Feature Extraction
 Feature extraction aims to reduce the number of features in a dataset by creating new
features from existing ones.
 We used 4 different types of extracting method. That are :
 Gray Level Co-occurrence Matrix
 Gray Level Run Length Matrix
 Gray Level Size Zone Matrix
 Gray Level Dependence Matrix
12 Gray Level Co-occurrence Matrix
The Gray Level Co-occurrence Matrix (GLCM) functions are used for finding texture
properties of an image by calculating the frequency of occurrence of pixel pairs with
specific values and in a specific spatial relationship. In our proposed system, we extract
16 features in GLCM. The formula for calculating GLCM features are :

Cluster
Autocorrelation
Tendency
μx = i−j)2p(i,j)
Joint Average Contrast

Cluster i+j−μx−μy)4 p(i,j)

Correlation
Prominence
Difference px−y(k)
i+j−μx−μy)3p(i,j)
Cluster Shade Average
13 Gray Level Co-occurrence Matrix (Cont.)
Difference x-y (k)log2(px−y(k)+ϵ)
IMSC 1
Entropy
-2(HXY2-HXY)
Difference k−DA)2px−y(k) IMSC 2
Variance
))2 IDM
Joint Energy
Q(i,j) =
Joint Entropy (i,j)log2(p(i,j)+ϵ) MCC

Where,  Ng is the number of gray levels used.

 μ is the mean value of P.
 μx, μy, ∂x and ∂y are the means and standard deviations
of Px and Py.
 Px(i) is the ith entry in the marginal matrix obtained by

summing rows of P(i, j).

14 Gray Level Run Length Matrix
For each grey level, the Gary Level Run Length Matrix (GLRLM) produces the size of
homogeneous runs. This matrix is calculated for each of the 11 texture indices derived
from this matrix. In our proposed system, we extract 10 features in GLRLM. The
formula for calculating GLRLM features are :

Gray Level Non-

Short Run Emphasis
Uniformity
(SRE)
Normalized (GLNN)

Long Run Emphasis Run Length Non-

(LRE) Uniformity (RLN):
Run Length Non-
Gray Level Non- Uniformity
Uniformity (GLN) Normalized (RLNN)
15 Gray Level Run Length Matrix (Cont.)

Run Percentage (RP) Run Variance (RV)

Gray Level Variance Run Entropy (RE)

(GLV)

Where,
is the number of pixels

denotes the total number of runs.

16 Gray Level Size Zone Matrix
The Gray Level Size Zone Matrix evaluate gray level zones in an image. A gray-level
zone is explained as a number of connected voxels that share the identical gray-level
intensity. In our proposed system, we extract 8 features in GLSZM . The formula for
calculating GLSZM features are :
Small Area Emphasis Size-Zone Non-
(SAE) Uniformity (SZN)
Large Area Emphasis Size-Zone Non-
(LAE) Uniformity
Normalized (SZNN)
Gray Level Non-
Uniformity (GLN) Zone Percentage (ZP)
Gray Level Non-
Uniformity Gray Level Variance
Normalized (GLNN) (GLV)
17 Gray Level Dependence Matrix
The Gray Level Dependence Matrix evaluate gray level constraint in an image. A gray-
level constraint is defined as a number of connected voxels within distance δ that are
dependent on the center voxel. In our proposed system, we extract 8 features in GLDM.
The formula for calculating GLDM features are :
Small Dependence Dependence Non-Uniformity
Emphasis (SDE) Normalized (DNN)

Large Dependence Gray Level Variance (GLV)

Emphasis (LDE)

Gray Level Non- Dependence Variance (DV)

Uniformity (GLN)
Low Gray Level Emphasis
Dependence Non- (LGLE)
Uniformity (DN)
18 Feature Selection

The feature selection method is the process of minimizing the number of input data. The
technique that helps :
 Minimizing a huge set of features.
 Select the best features from the original dataset.
 Reject the unnecessary features from the original dataset.
19 Binary Particle Swarm Optimization (BPSO)

 The BPSO algorithm is used in binary discrete search spaces.

 By putting a sigmoid transformation to the velocity component to force the velocities
into the value between 0 and 1.
 The position of the particles to be either 0 or 1.
 Kennedy and Eberhart have implemented the BPSO to search in binary discrete
search spaces.
20 Binary Particle Swarm Optimization (Cont.)

= W + C1r1(Pid - ) + C2r2(Pgd - ) (1)

S(Vid) = (2)

1, if rand <S )
Xid = (3)
0, Otherwise

Equation (1) is used to upgrade the velocity while (2) and (3) are used to upgrade the
position of each particle.
21 Binary Particle Swarm Optimization (Cont.)
Algorithm 1: Feature Selection algorithm using BPSO
1. begin
2. load the dataset into machine;
3. start the velocity of each particle and position;
4. while Maximum Iterations or the stopping criterion is not met do
5. for i=1 to P do
6. update the pbest of particle i;
7. update the gbest of particle i;
8. end
9. for i=1 to Population Size do
10. for d=1 to Dimensionality do
11. according to Equation 1, update the velocity of particle i ;
12. update the position of particle i according to Equations 2 and 3;
13. end
14. end
15. end
16. calculate the classification accuracy of the selected feature subset on the test set;
17. return the position of gbest (the selected feature subset);
18. return the training and test classification accuracies;
22 Classification
Support Vector Machine (SVM):
SVM is a binary classifier based on the concept of a hyperplane that defines
decision boundaries.

Figure 4: A linear support vector machine.

23 Cross Validation
K-Fold Cross Validation:
I. Shuffle the dataset randomly. III. For each unique group:
II. Split the dataset into k groups.  Take the group as a hold out
or test data set.
 Take the remaining groups
as a training data set.
 Fit a model on the training
set and evaluate it on the test
set.
 Retain the evaluation score
and discard the model.
 Summarize the skill of the
model using the sample of
model evaluation scores.
Figure 5: K-fold cross-validation process.
24 Experimental Results
To evaluate the performance of the proposed method, we have examined three metrics per
class: Sensitivity, Specificity and Accuracy
Table 1: Formula for evaluation scheme. Table 2: Confusion matrix for two-class
classification problem.

Sensitivity Predicted Class

TP / (TP + FN)
(Recall) Positive Negative

Specificity TN / (TN + FP) Positive TP FN

Actual
Class
Negative FP TN
Accuracy TP + TN / (TP + TN + FP + FN)

TP = True Positives; TN = True Negatives; FP = False Positives; FN = False Negatives

25
Experimental Results (cont’d)
89

88.2%
88
Table 3: Confusion matrix of the model 87.5% 87.5
without feature selection 87%
87

Predicted class
86

Normal Abnormal 85%

85
Actual class

Normal 42 7
84

Abnormal 6 45
83

Sensitivity Specificity Accuracy

Figure 6: Performance analysis of the model without feature selection.

26
Experimental Results (cont’d)
100
99
97.9% 98%
98
Table 4: Confusion matrix of the model 97%
97
with feature selection 96% 96.07%
96
95
94
Predicted class 93
92
Normal Abnormal 91
90
Actual class

Normal 48 1 89
88
87
Abnormal 2 49 86
85
Sensitivity Specificity Accuracy

Figure 7: Performance analysis of the model with feature selection.

27
Experimental Results (cont’d)
100
97.95%
97%
Table 5: Comparative analysis of the 96.035%

model 95

Average Average
Accuracy 90
Sensitivity Specificity
(%) 87.85%
(%) (%) 87%
85.75%
Model without feature
85.75 87.85 87.00 85
selection
Model with feature
97.95 96.035 97.00
selection 80

75

Sensitivity Specificity Accuracy

Figure 8: Performance metrics of the model with and without feature

selection
28 Comparison with Existing Methods

Table 6: Comparison among different methods.

Authors Methodology Accuracy

Bahadure, Ray, Thethi et al. [21] GA 92.03%

Roy, Sadhu, Kumar et al. [23] ANFIS 98.25%

Parasuraman, Vijay Kumar et al. [24] Ensemble Classifier 91.17%

Proposed Method BPSO + SVM 97%

 GA : Genetic Algorithm
 ANFIS : Adaptive Neuro Fuzzy Inference System
29 Discussion

 We incorporate both model without feature selection and with feature selection by
using SVM classifier.
 We consider the BPSO feature selection method because it gives us the best
Accuracy of 97.0%.
 The proposed system for brain tumor classification showed that the use of BPSO
feature selection method’s achieve average sensitivity, and specificity of 97.95%
and 96.035% respectively with SVM classifier.
30 Conclusion

The results of the experimental study have to lead us to the following clear conclusions:
 NRRD files are unacknowledged and do not contain sensitive patient information.
 Texture Feature provide information about the connected length of a particular pixel
in a definite direction.
 The BPSO feature selection technique helps in minimizing a huge set of features by
selecting the best features and rejecting the unnecessary features from the original
dataset.
 SVM classifier successfully detect the tumor by analyzing feature vectors and area
of the tumor.
31 Limitation

 Lack of publicly available MRI images.

 The high difference in various image qualities.
 Coding has developed as a top-down programming in each step, which caused the
processing to take long.
 We used only one classifier (SVM) to classify the brain tumor.
32 Future Work

 In the future, these methods will be tested on integrating a larger dataset with
high-resolution images.
 In future we will use some method that can identify exact type of the tumor.
33 Acceptance

Accepting Journal Names:

1. International Journal of Science and Research

(IJSR)
2. International Journal of Advance Research in
Engineering and Technology (IJARET)
3. International Journal of Innovative Research
in Engineering and Technology (IJIRSET)
35