Unsupervised Method for 3D Brain Magnetic Resonance Image Segmentation

Full text

Unsupervised Method for 3D Brain Magnetic
Resonance Image Segmentation
Adi Setyo Nugroho
Department of Informatics
Institut Teknologi Sepuluh Nopember
Surabaya, Indonesia
adinugroho.206025@mhs.its.ac.id
Chastine Fatichah
Department of Informatics
Institut Teknologi Sepuluh Nopember
Surabaya, Indonesia
chastine@if.its.ac.id
Francisca Notopuro
Department of Radiology
National Hospital
Surabaya, Indonesia
franciscanotopuro@gmail.com
Aziz Fajar
Department of Informatics
Institut Teknologi Sepuluh Nopember
Surabaya, Indonesia
azizfajar519@gmail.com
Achmad Fahmi
Department of Neurosurgery
Universitas Airlangga
Surabaya, Indonesia
achmad.fahmi-13@fk.unair.ac.id
Riyanarto Sarno
Department of Informatics
Institut Teknologi Sepuluh Nopember
Surabaya, Indonesia
riyanarto@if.its.ac.id
Sri Andreani Utomo
Department of Radiology
Dr. Soetomo General Hospital
Department of Radiology
Universitas Airlangga
Surabaya, Indonesia
sriandreaniutomo@gmail.com
Abstract— Research at Digital Imaging and Communication
in Medicine (DICOM) is very useful research in the field of
health. In brain images, the problem encountered is when you
want to divide or segment each part of the brain. In previous
studies, some of research are still segmenting from 2-
dimensional images, where the results will be different for each
image slice. Therefore, in this research, we conducted the
Magnetic Resonance Image (MRI) segmentation of the brain
from the 3-dimensional plane to prevent the information
contained in the images from being lost. In the early stages, MRI
images will be converted to NifTi format to obtain 3-dimensional
volume. The pre-processing is added as a modification from
previous research, such as, convert image to grayscale, bias field
correction, and skull stripping method to remove the skull (non-
brain tissue) so that only brain tissue remains from the human
brain. The segmentation process is done using multi-otsu
thresholding. The experimental result shows that our method
has successfully got three different brain tissue named white
matter (WM), gray matter (GM), cerebrospinal fluid (CSF)
Keywords—DICOM, segmentation, multi-Otsu thresholding
I. INTRODUCTION
Research in the medical imagery field is one of the most
widely conducted research subjects [4]. And Magnetic
Resonance Imaging (MRI) images are one of the most
frequently performed. Digital Imaging and Communication in
Medicine (DICOM) MRI images are obtained from a machine
to obtain images of the human brain. The images of the human
brain obtained from MRI are divided into many slices, which
are divided into 3 planes [8]. Namely the axial (from bottom
to top), coronal (from front to back), and sagittal (from right
to left).
Previously, research that had been done a lot was on 2-
dimensional MRI brain. Namely by taking one of the slices to
be processed. An example is a research conducted by Atikah
et al [1]. This study segmented the MRI image by taking one
of the planes, namely Sagittal. The method used is filtering,
then the largest connected component is carried out to take the
brain image, then segmentation is carried out using the
unsupervised learning method, namely k-means. However,
the disadvantage of using 2-dimensional images is the value
carried by each slice is different. And the selection process
must be done manually regarding which slice will be used as
the initial dataset.
There are several unsupervised methods that can be used
for the segmentation process. One of the segmentation
methods is threshold-based segmentation [10]. And in
general, the goal of segmenting brain tissue is classified brain
images into three main tissue classes, which is, white matter
(WM), gray matter (GM), and cerebrospinal fluid (CSF) [7].
Sakib et al. [2] conducted research on the MRI brain using
the Insight Toolkit (ITK) to perform segmentation. The study
conducted comparisons of 3 segmentation methods, namely
Otsu thresholding, Bayesian classification, and Bayesian
Gaussian smoothing. The purpose of this study was to
segmented brain images into several parts (WM, GM, CSF).
However, in this study researcher said that he did not do any
further about preprocessing images. So that this is a weakness
of the algorithm of the ITK. Because MRI can be noisy
without any preprocessing method [15].
Based on these problems, this research was proposed to
segment the 3-dimensional MRI image using an unsupervised
approach. This research, it was done by taking the slice from
the axial side. Then the conversion is done in the NifTi format
to obtain 3D images. The pre-processing is added as a
modification from previous research, such as gray-scaling
images, bias field correction, and skull stripping method to
remove the head bone so that only brain tissue remains from
the human brain. The results of the skull stripping will be
segmented using multi-Otsu thresholding to get three different
brain tissue (WM, GM, CSF).

II. P
REVIOUS
S
TUDY
.Atikah et al, [1], using adaptive thresholding, K-means at
clustering phase, and morphological operation to segmenting
the brain in the corpus callosum. The adaptive threshold is
used in the preprocessing phase of brain images. The purpose
of adaptive thresholding is to make the skull stripping much
easier. After pre-processing and get the brain, k-means is
performed to divide the parts of the brain to be segmented,
namely the corpus callosum. Then the noise in that phase is
removed using mathematical morphology.
Sakib et al. [2] compare segmentation methods on brain
tissue (GM, WM, CSF) using 3 methods, namely Otsu
thresholding, Bayesian, Bayesian plus Gaussian smoothing.
The brain dataset was tested one by one on the methods
available at ITK. The results obtained are quite good. And
CSF is the important point. In Otsu thresholding, brain tissue
segmentation in CSF is superior to the other 2 methods.
Despotovic et al. [6] discuss various method to do
segmentation on brain images. They use preprocessing
techniques such as bias field correction, image registration,
removal non-brain tissue, or many researchers call it skull
stripping. And then, they also discuss MRI segmentation
methods such as manual segmentation that have been done by
experts, intensity-based methods (Threshold, region growing,
clustering, classification), atlas-based method, surface-based
method, and hybrid methods. For segmentation evaluation,
this paper also mentioned Tanimoto coefficient and dice
similarity coefficient.
III. R
ESEARCH
M
ETHODOLOGY
This research was conducted with the aim of segmenting
several parts of the human brain on MRI images. The data
used is DICOM data which consists of 500 slices from various
planes. The programming language used in this research is
Python. A general description of the methodology is shown in
Figure 1
.
Figure 1 Research Methodology
Figure 1 describes the steps in this study. the first step is
to take a image dataset. After that, the data was converted to
NifTi. Then the skull stripping is done to remove the skull and
leave only the tissue image of the brain. The brain images are
then segmented to obtain tissue which is divided into three
part (WM, GM, CSF). And CSF will be used as a dividing
border to separate brain structures.
A. Dataset DICOM
In this study, the data used were DICOM data obtained
from National Hospital Surabaya. We used 2 datasets which
were scanned from 2 people. Let us call it dataset A and
dataset B. In the dataset, there is an MRI scan containing
brain images from various planes, namely sagittal, coronal,
axial. The size of each slice is 256 x 256 pixels. On MRI scan,
there are several metadata used to build 3D volumes defined
in Table 1
.
T
ABLE
1
D
ICOM METADATA
Tag Name
(0028, 0010) Rows
(0028, 0011) Columns
(0028, 0030) Pixel Spacing
Rows and Column use to defined 2 dimensional images
for the X-axis and Y-axis, and Pixel Spacing for the Z-axis to
perform 3D image. However, what we will process is from the
axial side. An example of this dataset A and datsaet B is shown
in Figure 2.
From figure 2, it is known that there is an image that is
defined as the brain tissue, and there is an image that is
designated as bone. The next process is to take an image of the
brain with the skull stripping.
(a)
(b)
Figure 2 MRI dataset from axial plane, sagittal plane, and coronal
plane in dataset A (a), and dataset B (b)
B. Pre-Processing
Preprocessing is the first step to removing irrelevant data
[14] and also to improve the quality of the image to make
processing much easier. There are several steps to performing
pre-processing images such convert to grayscale, bias field
correction, and skull stripping.

Figure 3 3D U-Net Architecture
1) Convert to Grayscale
The first step of the pre-processing image is gray-scaling.
The main purpose of gray-scaling is to convert images into 1
channel so that it is easy to carry out to the next process. This
1 channel will only store pixel values from a range of 0 – 255.
2) Bias Field Correction
Bias field correction is the step to removing intensity
inhomogeneity that coming from the magnetic field,
sensitivity variation of the coil, and interaction between
human and magnetic field [9]. Bias field correction also
makes the difference of brain tissue clearer.
3) Skull Stripping
Skull stripping is performed to remove bones from the
head. The purpose of the skull stripping is to simplify the next
segmentation process, leaving only brain tissue [8]. At this
stage, it is done using the open-source python code called
deep-brain from PyPI online python repository. It uses a
convolutional neural network based on 3D U-Net. The 3D U-
Net architecture can be seen in figure 3.
Figure 3 explains that in the 3D U-Net architecture there
are several stages. There are 4 types of arrows (orange arrow,
red arrow, yellow arrow, and green arrow), each of which
represents the convolution process, max pooling, upsampling,
and concatenation from each step. The convolution process is
carried out with a kernel size of 5 and using the activation
function named Rectified Linear Unit (ReLU). The process of
convolution itself will increase the depth of the images. So the
image is getting thicker. Max pooling, reduces image size by
half because it taking the important feature of the image. In
the upsampling phase, the image size is expanded and
concatenation is then carried out from the previous phase.
This model trained with already segmented brain images
(CC359 dataset, NFBS dataset, and ADNI dataset). This
training process produces weight and bias value that saved on
pb file format. So, this model performed segmentation of brain
tissue using the probability of each image voxel to generate
brain masking. The voxel is selected based on probability
which larger than same p (0.5 as default)
C. Segmentation
Thresholding is commonly used in image segmentation,
both in medical [1,12] or non-medical [10]. And the popular
one is Otsu thresholding. In this segmentation section, Brain
data that previously passed the skull stripping stage will be
segmented using multi-Otsu thresholding. Otsu [11] mention
that multi-Otsu thresholding is a renewal of thresholding that
can accommodate more than 1 threshold. This method,
maximizing variance between classes to obtain an optimal
threshold [13]. Multilevel threshold is good to segment the
image brightness into several regions [10], which correspond
to several objects and background. The method works very
well for objects of various colors. On which bi-level
thresholding fails to produce satisfactory results. For
example, in the case of 4 class, it is assumed we have 3
thresholds: 1 <= t
1
<= t
2
<= t
3
< L, to separate 4 classes, C0
for [1, ..., t
1
], C1 for [t
1
+ 1, ..., t
2
], C2 for [t
2
+ 1, ..., t
3
], and
C3 for [t
3
+ 1, …, L]. Then the interclass variance is
maximized by using the equation (1).





∗
, 

∗
, 

∗


  











, 

 (1)
Where :
 = variant of each class
L = level of each pixel
t = level of a specified threshold.
Table 2 show the basic algorithm of multi-Otsu
thresholding.
This multi-Otsu thresholding will be used to segment the
brain tissue to obtain WM, GM, and CSF. The goal is to get
the brain structure (WM and GM), and the border (CSF) that
is used as the basis of segmentation.

T
ABLE
2
M
ULTI
-O
TSU
T
HRESHOLDING
A
LGORITM
Multi-Otsu Thresholding
Input :
gray_images (a,b,c)
Output : image with 4 main class (background, WM, GM, CSF)
If
gray_images (a, b, c) < t
1
gray_images (a, b, c) = 0
If gray_images (a, b, c) < t
2
and If gray_images (a, b, c) > t
1
gray_images (a, b, c) = 1
If gray_images (a, b, c) < t
3
and If gray_images (a, b, c) > t
2
gray_images (a, b, c) = 2
Else
gray_images (a, b, c) = 3
D. Evaluation
Evaluation of segmentation process is done using dice
similarity coefficient (DSC). DSC, also known as the Dice
coefficient, is a tool with based on statistical method that
measures the similarity from two sets of data [4]. This the
most popular tool to validate the algorithms of image
segmentation. Not only the image, but this also can be applied
with several dataset and application including natural
language processing. The DSC formula is defined (2).
∗ | ∩ |
|| ||
(2)
Where :
X = set 1
Y = set 2.
A set with vertical bars either side refers to the cardinality
of the set, i.e. the number of elements in that set, e.g. |X|
means the number of elements in set X. ∩ is used to represent
the intersection of two sets, and means the elements that are
common to both sets.
IV. R
ESULT AND
A
NALYSIS
A. Pre-Processing
In the first stage, images are converted into grayscale with
the range of pixel values between 0 – 255. This will make the
image easy to process. The second step is performing bias
field correction. Bias field correction is done to remove the
magnetic field to make brain tissue clearer. The bias field
result can be seen in figure 4.
After that, skull stripping was performed on the dataset.
the purpose of this process is to get a brain image for further
processing. The skull striping is done using 3D U-Net from
the previous section. the results of skull stripping can be seen
in the figure 5.
(a) (b)
Figure 4 Brain image with bias field (a), and without bias field
Figure 5 Brain Masking
The figure 5 shows the result of the skull stripping. From
this process, we get a mask whose size matches the size of the
brain in the DICOM dataset. This mask contains a Boolean
value, True and False. White region on the brain has True
Boolean value, and the black region has False Boolean value.
Furthermore, the mask will be fit to the original brain images.
The results of applying masking to the brain can be seen in the
Figure 6.
From figure 6, we can see that the brain mask is fits with
the brain image to get only the brain images. After this step,
non-brain tissue is removed from the brain image, only the
brain tissue remain.
(a)
(b)
Figure 6 Applying Brain Masking to dataset A (a), and Dataset B
(b)
B. Segmentation
In general, segmentation of the MRI brain image is done
to obtain brain tissue. Brain tissue is divided into 3 main
parts, namely gray matter (GM), white matter (WM),
cerebrospinal fluid (CSF). at this stage, it is done by using
multi-Otsu thresholding. The results of this stage can be seen
in the Figure 7.
(a)

(b)
Figure 7 Result of WM, GM, CSF dataset A (a), and dataset B (b)
Figure 7 shows that the brain has been divided into several
regions based on the threshold results. the yellow region is
WM, the blue region is GM and CSF.
V. C
ONCLUSION
This paper uses an unsupervised approach to segment the
brain images from MRI data. Three parts of brain tissue such
as WM, CSF, GM were successfully segmented by this
method. The preprocessing process is done by grayscaling to
make the segmentation process easier by changing the pixel
value to 0 - 255. Skull stripping using 3D U-Net is used to
eliminate non-brain tissue to prevent segmentation error. And
then, multi-Otsu thresholding is used to separate the three
brain tissue classes.
A
CKNOWLEDGEMENT
This research was funded by AUN/SEED-Net Under Special
Program for Research Against COVID-19 (SPRAC) and the
Ministry of Research and Technology/National Research and
Innovation Agency Republic of Indonesia (Ristek-BRIN)
and the Indonesian Ministry of Education and Culture under
RISPRO Invitation Program managed by Lembaga Pengelola
Dana Penelitian (LPDP) and under Penelitian Terapan
Unggulan Perguruan Tinggi (PTUPT) Program managed by
Institut Teknologi Sepuluh Nopember (ITS).
R
EFERENCES
[1]
L. Atikah, N. A. Hasanah, R. Sarno, A. Fajar, D. Rahmawati.
“Brain Segmentation using Adaptive Thresholding, K-Means
Clustering and Mathematical Morphology in MRI Data,” 2020
International Seminar on Application for Technology of
Information and Communication (iSemantic), Semarang, pp.
161-167, 2020
[2]
S. Sakib, M. A. B. Siddique, “Unsupervised Segmentation
Algorithms' Implementation in ITK for Tissue Classification
via Human Head MRI Scans”, arXiv, 1902.11131, 2020
[3]
L. Wang, C. Xie, N. Zeng, “RP-Net: A 3D Convolutional
Neural Network for Brain Segmentation from Magnetic
Resonance Imaging”, IEEE Access, vol. 7, pp. 39670 – 39679,
2019
[4]
S. Kumar, A. Negi, J. N. Singh, H. Verma. “A Deep Learning
for Brain Tumor Segmentation MRI images Semantic
Segmentation Using FCN,” 2018 4
th
International Conferences
on Computing Communication and Automation (ICCCA),
Greater Noida, pp. 1-4, 2018
[5]
H. M. Moftah, A. E. hassanien, M. Shoman, “3D Brain Tumor
Segmentation Scheme Using K-Means Clustering and
Connected Component Labeling Algoritm,” 2010 10
th
International Confrence on Intelligent System Design and
Application, Cairo, pp. 320-324, 2010
[6]
I. Despotovic, B. Goossens, W. Phillips, “MRI Segmantation
of the Human Brain: Challenges, Methods, and Application,”
Computational and Mathematical Methods in Medicine, vol.
2015, Article ID 450341, 23 pages, 2015
[7]
Kamarujjaman, M. Maitra, “3D unsupervised modified spatial
fuzzy c-means method for segmentation of 3D brain MR
image,” Pattern Anal Applic 22, pp. 1561–1571, 2019.
[8]
J. G. Park, T. Jeong, C. Lee, “Automated Brain Segmentation
Algoritm for 3D Magnetic Resonance Brain Images,” 2007 2
nd
International Workshop on Soft Computing Application,
Oradea, pp. 57-61, 2007
[9]
N. Yamanakkanavar, J. Y. Choi, B. Lee, “MRI Segmentation
and Classification of Human Brain Using Deep Learning for
Diagnosis of Alzheimer’s Disease: A Survey,” Sensors, 20,
3243, 2020.
[10]
S. Arora, J. Acharya, A. Verma, Prasanta K. Panigrahi,
“Multilevel thresholding for image segmentation through a fast
statistical recursive algorithm,” Pattern Recognition Letters,
vol. 29, pp. 119-125, 2008
[11]
N. Otsu, “A Threshold Selection Method from Gray Level
Histograms”, IEEE Transaction On System, Man, and
Cybernetics, vol. 9, pp. 62-66, 1979
[12]
K. N. Aisyah, C. Fatichah, R. Sarno, “Multilevel Thresholding
and Morphological Relationship Approach for Automatic
Detection of Anterior and Posterior Commissure in Mid-sagital
Brain MRI,” International Journal of Intelligent Engineering
and System, vol. 13, no. 5, pp. 368-378, 2020
[13]
H. Tariq, A. Muqeet, A. Burney, M. A. Hamid, H. Azam,
“Otsu’s Segmentation: Review, Visualization and Analysis In
Context of Axial Brain MR Slices”, Journal of Theoretical and
Applied Information Technology, vol. 95, no. 22, 2017
[14]
P. Damayanti, D. Yuniasri, R. Sarno, A. Fajar and D.
Rahmawati, "Corpus Callosum Segmentation from Brain MRI
Images Based on Level Set Method," 2020 International
Seminar on Application for Technology of Information and
Communication (iSemantic), pp. 155-160, Semarang,
Indonesia, 2020
[15]
L. Lazli, M. Boukadoum, “Improvement of CSF, WM, and GM
Tissue Segmentation by Hybrid Fuzzy - Possibilistic Clustering
Model based on Genetic Optimization Case Study on Brain
Tissues of Patients with Alzheimer's Disease,” International
Journal of Networked and Distributed Computing, vol. 6, no.
2, 2018