08-16-2017, 09:17 PM
Brain tumor detection from MRI
A Project Report
Submitted in partial fulfilment of
the requirements for the award of the degree of
Master of Technology
in
Computer Science and Engineering
by
Irene Antony Tharayil
M105107
Department of Computer Science & Engineering
College of Engineering Trivandrum
Kerala - 695016
2010-11
A Project Report
Submitted in partial fulfilment of
the requirements for the award of the degree of
Master of Technology
in
Computer Science and Engineering
by
Irene Antony Tharayil
M105107
Department of Computer Science & Engineering
College of Engineering Trivandrum
Kerala - 695016
2010-11
Abstract
Brain tumor is one of the major causes for the increase in mortality among children and
adults.A tumor is a mass of tissue that grows out of control of the normal forces that regulates
growth.The proposed system enables Automatic detection of brain tumor through MRI.First
stage has preprocessing and enhancement. Second, feature extraction, feature selection, classi-
fication, and performance analysis are compared and studied. Preprocessing and enhancement
techniques are used to improve the detection of the suspicious regions in MRI.
The enhancement method consists of three processing steps: first, the MRI image is acquired.
Second, removal of film artificates such as labels and marks on the MRI image and finally the
high frequency components are removed. Segmentation describes separation of suspicious region
from background MRI images.The critical problem is finding the tumor location auto- matically
and later finding its boundary precisely. The objective of this work is to present an automated
unsupervised method for finding tumor.An important factor in detecting tumor from healthy
tissues is the difference in intensity level.
The aim of this work is to design an automated tool for brain tumor quantification using
MRI image data sets. This work is a small and modest part of a quite complex system. The
whole system will when completed visualize the inside of the human body, and make surgeons
able to perform operations inside a patient without open surgery.
[attachment=8650]
1 Introduction
The body is made up of many types of cells. Each type of cell has special functions. Most
cells in the body grow and then divide in an orderly way to form new cells as they are needed
to keep the body healthy and working properly. When cells lose the ability to control their
growth, they divide too often and without any order. The extra cells form a mass of tissue
called a tumor. Tumors are benign or malignant. There are three methods of segmentation.
The aim of this project is to design an automated tool for brain tumor quantification using
MRI image data sets. This work is a small and modest part of a quite complex system. The
whole system will when completed visualize the inside of the human body, and make surgeons
able to perform operations inside a patient without open surgery. More specifically the aim for
this work is to segment a tumor in a brain. This will make the surgeon able to see the tumor
and then ease the treatment. The instruments needed for this could be ultrasound, Computer
Tomography (CT Scan) and Magnetic Resonance Imaging (MRI). In this project, the technique
used is Magnetic Resonance Imaging (MRI).
Watershed segmentation uses the intensity as a parameter to segment the whole image data
set. Moreover, the additional complexity of estimation imposed to such algorithms causes
a tendency towards density dependent approaches.[2]. Three dimensional segmentation is a
reliable approach to achieve a proper estimation of tumor volume. Among all possible methods
for this purpose, watershed can be used as a powerful tool which implicitly extracts the tumor
surface. Watershed segmentation based algorithm has been used for detection of tumor in 2D
and in 3D. For detection of tumor in 2D the software required is MATLAB. But for detection
of tumor in 3D, the software required are MATLAB and 3D Slice.This work is a small and
modest part of a quite complex system.
2 Methodology
A conceptually simple supervised block-based and image-based (shape, texture, and content)
technique has been used to analyze MRI brain images with relatively lower computational
requirements. The process flow of our proposed methodology may be shown as figure 1.
The first section discusses how images are divided into regions using a block-based method.
The second section shows how each classified block is studied individually by calculating its
multiple parameter values. In this instance, the multiparameter features refer to the following
three specific features: the edges (E), gray values (G), and local contrast (H) of the pixels in
the block being analyzed.
Input Image The images we got from MRI are of three types: axial Images, saggital Images,
coronal Images. The numbers of images depend on the resolution of the movement of the MRI
magnets.
2.1 Preprocessing
The Preprocessing is used for loading the Input MRI images to the MATLAB Environment and
also it removes any kind of noise present in the input images. In preprocessing the first step is
to load the MRI image data set on to the MATLAB workspace and after loading they will be
processed in such a way that instead of processing 128 images in one direction a whole clip of
5
Figure 1: Methodology
128 images is processed by one command, otherwise it would have been very hectic situation
for processing each and every image independently. Thus after this processing there are only
three clips instead of 384 separate images, i.e. one clip for axial images, one clip for saggital
images and one clip for coronal images. After that all the clips are combined to get the single
clip for further processing. Then the noises are filtered out from MRI images using the Weiner
filter which is a type of linear filter. The MRI image after removal of noise is further used for
parameter calculation.
2.2 Multiparameter Calculations
Recent advances in medical image analysis often include processes for an image to be segmented
in terms of a few parameters and into smaller sizes or regions, to address the different aspects of
analyzing images into anatomically and pathologically meaningful regions. Classifying regions
using their multiparameter values makes the study of the regions of physiological and patho-
logical interest easier and more definable. Here, multiparameter features refer to the following
three specific values for the edges (E), gray values (G), and local contrast (H) of the Pixels
2.2.1 Edge (E) Parameter
Edge information is often used to determine the boundaries of an object. This is mainly used
for analysis to derive similarity criterion for a predetermined object. The incidences of cerebral
compression reduce the edge. Given this understanding, It can be use the Sobel edge detection
method to detect image edges (IE) is obtained by filtering an input image with two convolution
kernels concomitantly, one to detect changes in vertical contrast (hx) and the other to detect
horizontal contrast (hy ).
6
Figure 2: Watershed segmentation simplified to 2 Dimensions
2.2.2 Gray (G) Parameter
Gray parameter avoids the need to scale the data-tocolor mapping, which would be required if
we used a color map of a different size. The gray parameter (G) for each block of the brain is
accumulated, and controlled by a binary image (IT ) using the GD value as a threshold.
2.2.3 Contrast (H) Parameter
An intensity image is a data matrix, I, whose values represent intensities within some range.
MATLAB stores an intensity image as a single matrix, with each element of the matrix cor-
responding to one image pixel. The matrix can be of class double, uint8, or uint16. While
intensity images are rarely saved with a color map, MATLAB uses a color map to display
them. In essence, MATLAB handles intensity images as indexed images. Contrast (H) is often
used to characterize the extent of variation in pixel intensity. In the present technique, the
computational program analyses the differences, especially in instances of strong dissimilarity,
between entities or objects in an image I(x,y).x
2.2.4 Watershed Segmentation
By interpreting the gradient map of an intensity image as height values, we get lines which
appear to be ridges. If the contours were a terrain, falling rain would find the way from
the dividing lines towards the connected catchments basin. These dividing lines are called
watersheds. As illustrated in Fig. 5.1 steep edges cause high gradients which are watersheds .
Image segmentation by mathematical morphology is a methodology based upon the notations
of modification. The watershed transformation can be built up by flooding process on a gray
tone image and may be shown as shown in figure 2.
It has been found that among the segmentation methods investigated in this work, the
watershed segmentation, a classic in image segmentation, marked out as the most automatic
method of the three. As watershed segmentation technique segregates any image as different
intensity portions and also the tumerous cells have high proteinaceous fluid which has very
high density and hence very high intensity, therefore watershed segmentation is the best tool
to classify tumors and high intensity tissues of brain. Watershed segmentation can classify
the intensities with very small difference also, which is not possible with snake and level set
method. It has been found that the snake and the level set method were best initialized from
the inside of the tumor. The program needs to be extending to handle segmentation with a
probe through the tumor. This can be done by segmenting the probe or in combination with
tracking information which provides the position of the probe.
7
Figure 3: Axial Slices
3 Detailed Design
The different stages in this project are
3.1 Preprocessing
Load and View Axial Images in MATLAB Environment as shown in figure 3.
Load and View Saggital Images in MATLAB Environment as shown in figure 4.
Load and View Coronal Images in MATLAB Environment as shown in figure 5. Now all
the clips were combined together to produce a single clip and then any noise present in the
MRI images had been removed by using the algorithm which is based on Weiner filter.
3.2 Multiparameter Calculations
Recent advances in medical image analysis often include processes for an image to be segmented
in terms of a few parameters and into smaller sizes or regions, to address the different aspects of
analyzing images into anatomically and pathologically meaningful regions. Classifying regions
using their multiparameter values makes the study of the regions of physiological and patho-
logical interest easier and more definable. Here, multiparameter features refer to the following
three specific values for the edges (E), gray values (G), and local contrast (H) of the pixels.
Extract Axial, saggital and coronal slices and create Movie clip as shown in figure 6.
8
Figure 4: Saggital Slices
Figure 5: Coronal Slices
9
Figure 6: Movie clips of axial, saggital and coronal Slices
Figure 7: Edge Image of MRI Data Set
3.2.1 Edge Parameter Calculation (E)
We will be using Sobel edge detection for detecting edges as explained in the theoretical section.
This can be achieved by executing the algorithms and the result obtained may be shown as
figure 4.
3.2.2 Gray Parameter (G) Calculation
The gray parameter (G) for each block of the brain is accumulated, and controlled by a binary
image using the value as a threshold. Pixels intensity for each slice was calculated to establish
the threshold values and thus provide the basis for analysis of clinical MR images from patients
with brain tumors. This can be achieved by executing the algorithm and the result obtained
may be shown as figure 5.
10
Figure 8: Binary Image of MRI Data Set
3.2.3 Contrast Parameter (H) Calculation using morphological watersheds
Contrast (H) is often used to characterize the extent of variation in pixel intensity. A computa-
tional program analyses the differences, especially in instances of strong dissimilarity, between
entities or objects in an image using Watershed Segmentation. As we know that malignant
tumor cells contain highly proteinaceous fluid, which is represented as high signal intensity on
MRI images of the brain. Usually the watershed transformation is applied to a boundary map,
which is a gray scale function, derived from the input image, that has low values within the
regions and high values along region boundaries. The gradient magnitude of an intensity based
image is oftenly used as the boundary map, as well as higher order features such as curvature.
Watershed segmentation can be used for segregating the different intensity portions and this
can be achieved by executing the algorithm in MATLAB and the result obtained may be shown
as figure 6.
3.3 Tumor Block Detection and Visualization
3.3.1 Segmentation of brain tumor using Region of Interest (ROI) Command
As it has been seen from the above result that high density images have been separated from the
MRI images using Watershed Segmentation. Here main aim is to segment the tumor from the
MRI images. This can be done by using the ROI command in MATLAB. After the application
of the ROI command, the tumor may be segmented. This can be achieved by executing the
algorithm in MATLAB
3.3.2 Formation of 3D image of MRI Data Set using 3D Slicer
As the MRI Image date set is collection of 2D images. The tumor can not be segmented
in 3D unless and until we have 3D MRI image data set. Therefore, a software 3D SLICER
has been used to get a 3D image from a collection of 2D MRI data set of axial, saggital and
coronal images. Then, applying watershed segmentation (3D) in MATLAB to this 3D image,
the segmented tumor in 3D with all its dimensions can be obtained using 3D Slicer.
3.3.3 3D Watershed segmentation
The unbiased hierarchical queue The 3D watershed segmentation is a segmentation tech-
nique based on a physical concept of immersion. It is achieved in analogy to the way that
water fills a geographic surface. As the water floods its basins there will be points where flood
11
Figure 9: face neighborhood
regions meet. These points construct the watershed lines that divide the surface into distinct
regions. This method is based on the application of watershed segmentation to the faces of this
3D mesh. The immersion process is simulated from the heights of marking faces. The water
rises in each basin and when two basins meet, a watershed is created between them. The height
function of faces that we have used is the principal curvature.
The hierarchical queue is made up of several FIFO queues and each queue corresponds to a
curve level. Queues are sorted by level and each queue can be unstacked only if previous queues
are emptied. The faces initially marked are the first to be stacked by assigning each a different
label. The faces observed in the neighborhood of the faces treated throughout the immersion
process will be placed in the queue corresponding to their level of curvature
A watershed segmentation biase occurs when the neighbors of a treated face are also the
neighbors of the other faces not yet treated or in the case of the immersion of the flat regions
where the order of processing is arbitrary. To eliminate this biase, has proposed a technique to
assign a pair-value to a final label, and an odd-value to a temporary label. A face that is stacked
is analyzed: if it has two different neighbors of pair labels, it is a watershed. Otherwise it will
take its own associated pair-label. For each face of the final label, we explore its neighborhood
that we label on a temporary basis before placing it in the hierarchical queue (Fig.2). A simple
stack (called alpha) that has higher priority is added to manage the faces of the neighborhood
witch are of the same level.
We used the transformation algorithm based on the unbiased hierarchical queue for 2D wa-
tershed segmentation invented by S.Beucher. To perform the watershed transformation, we
define the concept of neighborhood: For each face, the neighborhood is defined by the faces
sharing with this face one or several common vertices(figure 9).
The hierarchical queue is made up of several FIFO queues and each queue corresponds to a
curve level. Queues are sorted by level and each queue can be unstacked only if previous queues
are emptied. The faces initially marked are the first to be stacked by assigning each a different
label. The faces observed in the neighborhood of the faces treated throughout the immersion
process will be placed in the queue corresponding to their level of curvature
12
Figure 10: Flow chart of Watershed Segmentation
A watershed segmentation biase occurs when the neighbors of a treated face are also the
neighbors of the other faces not yet treated or in the case of the immersion of the flat regions
where the order of processing is arbitrary. To eliminate this biase, has proposed a technique to
assign a pair-value to a final label, and an odd-value to a temporary label. A face that is stacked
is analyzed: if it has two different neighbors of pair labels, it is a watershed. Otherwise it will
take its own associated pair-label. For each face of the final label, we explore its neighborhood
that we label on a temporary basis before placing it in the hierarchical queue (Fig.11). A simple
stack (called alpha) that has higher priority is added to manage the faces of the neighborhood
witch are of the same level.
The application of watershed segmentation approach can be used to perform an initial seg-
mentation of a 3D object; however, it is necessary to merge some adjacent regions in order to
reach a given number of regions. Merging these regions is based on the depth of the Gaussian
13
Figure 11: Extraction cycle and hierarchical queue stacking
curvature (the depth of the watershed saddle point S between regions A and B corresponds to
the difference between the saddle point curvature and the minimum curvature of the regions A
and B). The Flowchart in Figure 10 shows the different steps of our segmentation process.
On applying watershed algorithm, it is easy to identify the 3d view of tumour. By merging
the generated regions based on guassian curvature, tumour is identified and specifications like
height, width, length etc can be calculater.
14
4 Conclusion
The project uses Watershed Segmentation and hope that it can successfully segment a tumor
provided the parameters are set properly in MATLAB environment. This project tries to prove
that methods aimed at general purpose segmentation tools in medical imaging can be used for
automatic segmentation of brain tumors.Our aim is to maintain quality of the segmentation to
be more actual than manual segmentation and will speed up segmentation in operative imaging.
It has only one limitation that the method is semi-automatic.