Brain Tumor Segmentation and Classification

Likhitha S. (IMT2014030)
Nikunj Gupta (IMT2014037)
Tummala Armitha (IMT2014060)
December 10, 2017

1 Introduction
Brain tumor segmentation seeks to separate healthy tissue from tumorous re-
gions.This is an essential step in diagnosis and treatment planning in order to
maximize the likelihood of successful treatment. Due to the slow and tedious
nature of manual segmentation, computer algorithms that do it faster and ac-
curately are required. Because of the unpredictable appearance and shape of a
brain, segmenting brain tumors from imaging data is one of the most challenging
tasks in medical image analysis.

2 Goal Statement
Segmentation of the brain tumour and classify it as either benign or malignant
tumor

3 Background
Brain is a complex organ since it contains more than 10 billion working brain
cells. Primary brain tumors can be either malignant (contain cancer cells) or
benign (do not contain cancer cells). A primary brain tumor is a tumor which
begins in the brain tissue. If a cancerous tumor starts elsewhere in the body,
it can spread cancer cells, which grow in the brain. These type of tumors are
called secondary or metastatic brain tumors. The malignant tumor tends to
grow and spread in a rapid and uncontrolled way that can cause death and the
Tumor are graded according to how aggressive. They are as

• LGG: Low Grade Tumor (Benign stage)

• HGG: High Grade Tumor (Malignant stage)

1
4 Related work
A research paper by Dr.A.R. Kavitha [1], talks about using Genetic Algorithm
to segment the MRI brain tumor images. Pre-processing was done using Wiener
Filter (a 2D adaptive noise removal filter and it uses pixel-wise adaptive wien-
wer method). GLCM features are extracted from segmented images and given
to the SVM Classifier which gets trained and ready for classifying test images.

Another research paper by Alan Jose [2] did Brain Tumor Segmentation Us-
ing K-Means Clustering And Fuzzy C-Means Algorithms And Its Area Calcu-
lation. After the segmentation, (which is done through k-means clustering and
fuzzy c-means algorithms) the brain tumor is detected and its exact location is
identified.

5 Data
we have collected our data from the BRATS 2015 Challenge [3].

5.1 Data set Folder Structure:

• Training
– HGG - This folder contains brain images of 220 Patients.There is a
different folder for each patient. There are 5 different MRI images
for each Patient. The 5 different images are T1,T2,T1C,FLAIR and
OT(Ground truth of tumor Segmentation ) . All these image files
are stored in .mha format.
– LGG - This folder contains brain images of 54 Patients.There is a
different folder for each patient. There are 5 different MRI images
for each Patient. The 5 different images are T1,T2,T1C,FLAIR and
OT(Ground truth of tumor Segmentation ) . All these image files
are stored in .mha format.
• Testing
– HGG LGG - This folder contains brain images of 110 Patients.There
is a different folder for each patient. There are 4 different MRI im-
ages for each Patient. The 4 different images are T1,T2,T1C and
FLAIR.All these image files are stored in .mha format.

After gaining insights in domain and lot of research, we have decided to use
only FLAIR images for our work.

2
6 Summary of Implementation
We start with taking each image from the training set and segment the tumor
from the image using watershed clustering (region-growing) technique. Then,
we extract GLCM features on the segmented image. We repeat the above pro-
cess for every image in the data set and use them to train our model - random
forest Classifier. Then, for every testing image, we segment the tumor using the
same method (Watershed) and then classify the tumor as benign or malignant
using our model.

Figure 1: Flow Chart

7 Implementation
The whole project is basically divided into for major sections, namely,
• Data Processing

3
 • Segmentation
• Feature Extraction
• Classification

7.1 Data Processing

Gaussian filtering is highly effective in removing Gaussian noise from the image.
The images we got from BRaTS were very noisy and had to be de-noised to be
used for segmentation.

Gaussian Blur [4]: In this approach, a Gaussian kernel is used. It is done

with the function, cv2.GaussianBlur(). We should specify the width and height
of the kernel which should be positive and odd. We also should specify the
standard deviation in the X and Y directions, sigmaX and sigmaY respectively.
If only sigmaX is specified, sigmaY is taken as equal to sigmaX. If both are
given as zeros, they are calculated from the kernel size.
If you want, you can create a Gaussian kernel with the function,
cv2.getGaussianKernel().

7.2 Segmentation
The data set has brain images with tumor which we segmented to provide it
to the feature extraction module. After segmentation we obtained two major
regions in the image- tumor and non-tumor. The segmentation was done using
Watershed Algorithm.

Watershed Algorithm [5]: OpenCV implemented a marker-based water-

shed algorithm which is an interactive image segmentation method. It gives
different labels for the object we know. Label the region which we are sure of
being the foreground or object with one color (or intensity), label the region
which we are sure of being background or non-object with another color and
finally the region which we are not sure of anything, label it with 0. That is our
marker. Then apply watershed algorithm. Then our marker will be updated
with the labels we gave, and the boundaries of objects will have a value of -1.
This is basically a region-growing approach.

7.3 Feature Extraction

Now we need to extract some features from images as we need to do a binary
classification of them using a classifier which needs these features to get trained
on. We chose to extract GLCM (texture-based features).

GLCM Features [6]: After obtaining the segmented images form the above
subsection, we extract GLCM features and store them. GLCM stands for

4
Gray-Level Co-occurence Matrix. Texture Analysis Using the Gray-Level Co-
Occurrence Matrix (GLCM) is a statistical method of examining texture that
considers the spatial relationship of pixels.

7.4 Classification
We used Random Forest for classification. The classifier was trained on the
GLCM features obtained from the segmented images and used to classify be-
nign (LGG) and malignant (HGG).

Random Forests [7]: When the training set for the current tree is drawn
by sampling with replacement, about one-third of the cases are left out of the
sample. This oob (out-of-bag) data is used to get a running unbiased estimate
of the classification error as trees are added to the forest. It is also used to get
estimates of variable importance.
After each tree is built, all of the data are run down the tree, and proximities
are computed for each pair of cases. If two cases occupy the same terminal node,
their proximity is increased by one. At the end of the run, the proximities are
normalized by dividing by the number of trees. Proximities are used in replac-
ing missing data, locating outliers, and producing illuminating low-dimensional
views of the data.

8 Tools
We have implemented our project in Python
• Read .mha images: SimpleITK
• Gaussian Blur: OpenCV
• Watershed algorithm: SciPy, scikit-image, and OpenCV
• GLCM : Package mahotas
• Random forest: RandomForestClassifier from sklearn.ensemble

9 Various other approaches explored

We have attempted several approaches for the classification of tumor in brain
images.

• Data Preprocessing: Removing noise and enhancing the image.

– Median filter
– Gaussian blur

5
 The results of segmentation were not good on images that are filtered
using Median Filter.The outline of brain is also segmented as tumor on
images which are filtered using Median Filter.So, we have decided to use
Gaussian Filter which gave good results for segmentation (The outline of
brain is not segmented as tumor).

• Segmentation: Identifying the tumored region in the brain image. We

have tried using the following techniques:

– Binary thresholding
– Watershed algorithm

Binary thresholding can be used to convert a gray scale image to binary

image based on the selected threshold values. The problems associated
with such approach are that binary image results in loss of texture and
the threshold value is coming out be different for different images. Hence,
we are looking for a more advanced segmentation algorithm, the watershed
algorithm.

• Classification: Classify whether the tumored part belongs to malignent or

benign
– Support Vector Machine
– Random forest

The accuracy after training the model using SVM is only 40 percent.So,
we have decided to use Random Forest, which gave better accuracy.

10 Results
Since, we do not have ground truth labels in our testing dataset, we have decided
to use only training dataset for both training and testing. As metioned earlier,
there are only 54 images in LGG and there are 220 images in HGG, we have
decided to consider all 54 images in LGG and only 54 out of 220 images from
the HGG.Now, we have 54 LGG images and 54 HGG images.We have trained
our model using 50 HGG and 50 LGG images. The remaining 4 HGG and 4
LGG are used for testing.The results are shown in below figure.

6
 Figure 2: Results

Form the above figure , we can see that 6 testing images are classified correct
and other 2 images are classified incorrectly .
For evaluating our model,we have calculated measures like Accuracy, Precision,
recall and F1-Score.The results for 6 different training datasets and correspond-
ing testing datasets can be found below.

Figure 3: Evaluation Measures

The accuracy of our model is nearly 71 percent.

7
11 Conclusions
Since, we are automatically segmenting the tumor and classifying it as Benign
or Malignant,it significantly reduces the amount of time that a doctor takes to
identify and classify the tumor. It also to some extent reduces the error made
while segmenting and classifying tumor.

12 Constraints
As explained in earlier sections, due to imbalance in number of images in LGG
and HGG classes,we have trained our model with only 50 LGG amd 50 HGG
images.Due to some Constraints,we have used only classical ML techniques like
SVM, Random Forest etc instead of Deep Learning Techniques like CNN ,which
could have given more accuracy.

13 Future work
• Improving the accuracy of our model by training the model with more
number of images.
• Usage of Deep Learning techniques like Convolution Neural Networks to
make our model better.
• Identifying or labelling the tumor sub-regions i.e specifying which sub-
region is edema, non-enhancing solid core, necrotic/cystic core and en-
hancing core.

14 References
[1] Link for Data Set
[2] ”Brain Tumor Segmentation using Genetic Algorithm with SVM Classifier”
by Dr.A.R. Kavitha, L.Chitra, R.kanaga
[3] ”Brain Tumor Segmentation Using K-Means Clustering And Fuzzy C-Means
Algorithms And Its Area Calculation” by Alan Jose, S.Ravi, M.Sambath.
[4] https://docs.opencv.org/3.1.0/d3/db4/tutorialp yw atershed.html
[5]https : //www.smir.ch/
[6]https : //in.mathworks.com/help/images/texture−analysis−using−the−
gray − level − co − occurrence − matrix − glcm.html
[7]https : //www.stat.berkeley.edu/ breiman/RandomF orests/cc home.htmoverview
[8]https : //pyscience.wordpress.com/2014/10/19/image − segmentation −
with − python − and − simpleitk/