Brain tumor segmentation
March 16, 2020
Abstract
Brain tumor is a severe disease which may lead to the end of life if not
attended and treated in its early stage. To know the condition of the tu-
mor various conventional and recently developed deep learning based seg-
mentation and classification methods have been proposed by researchers.
Deep learning is found to be efficient and robust for classification and
segmentation as it finds the fine-to-corase information about the tumors.
The main component of deep learning is layered neural network architec-
ture popularly known as convolutional neural network. Depending upon
architecture distinct information from the brain images can be captured
and further analysed. Still more research is required in this area to get
high segmentation and classification accuracy. In this paper various CNN
architecture used by researchers in the literature has been discussed, com-
pared and analysed. This architecture performs better along with some
pre-processing and postprocessing stages. The researches used various
datasets for their study such as BRATS, Mild Traumatic Brain injury
outcome prediction (mTOP), Multiple sclerosis segmentation (MSSEG),
Neonatal brain segmentation (NeoBrainS12), Ischemic stroke lesion seg-
mentation. The qualitative segmentation and classification was evaluated
by various performance measures (Dice similarity coefficient, sensitivity,
specificity, positive predictive value and accuracy).
1 Introduction
In human body normal cells grow old or get damaged and die in due course
of time and new cells take their place. Sometimes this process goes wrong and
the old damaged cells do not die as they should, this leads to the formation of
new cells, which are not required by the body. The build-up of extra cells often
form a mass of tissue called a growth or tumor. The uncontrolled, unnatural
growth and division of cells in the body is called cancer. Skull, which encloses
the brain is very rigid and any growth inside is restricted. The growing tumor
inside the brain increases intra cranial pressure, which can cause brain damage
and may be even life threatening. The mortality rate due to brain tumor is
the highest among all other types of diseases. Still people do have a higher
chance of survival if the tumor can be detected in the early stages. Based on
the origin of the tumors it can be considered as primary brain tumors and
1
secondary brain tumors. Primary brain tumors are originated from the brain
tissue cells, whereas in secondary ones cells become cancerous at any part of
the body and spread into the brain. The tumors of any type can be benign or
malignant. Benign tumors are slow growing, non-cancerous and do not spread
to the surrounding tissue. Benign tumors have obvious border or edge and with
time these tumors can become malignant. Malignant tumors are cancerous, fast
growing and aggressive. These tumors can invade nearby tissue. According to
World Health Organization (WHO), the tumors are classified into four types
based on their microscopic characteristics. Grade I tumor’s tissue is benign
and it looks like normal brain cells, which grow slowly. In Grade II and Grade
III tumors, tissues are malignant and they look like normal cells. Grade IV
tumors are known as glioblastomas multiforme (GBM). In this case, tissue is
malignant and cells look abnormal and tend to grow rapidly. Gliomas are the
most common and life threatening brain tumors having the highest mortality
rate due to their quick progression. These are infilterative in nature and mostly
escalate near the white matter fiber, but can spread to any part of the brain
making them very difficult to detect. Grade I and Grade II tumors refer to the
low grade gliomas, whereas Grade III and Grade IV are high grade gliomas.
Grades of brain tumors are investigated by magnetic resonance imaging
(MRI), computed tomography (CT), positron emission tomography (PET), and
single-photon emission computed tomography (SPECT). Out of the above men-
tioned imaging techniques, MRI technique is preferred for the diagnosis and
treatment, due to its non invasive nature and higher contrast for soft tissues.
MRI provides various modalities, such as T1-weighted (T1), T1-weighted with
contrast enhancement (T1c), T2-weighted (T2), and fluid attenuated inversion
recovery (FLAIR) . Healthy tissues are distinguished by T1 images, whereas
the edema region is highlighted by T2 images. In T1-Gd images, the tumor
border can easily be distinguished by the bright signal of the accumulated con-
trast agent (gadolinium ions) in the active cell region of the tumor tissue. The
necrotic cells do not interact with the contrast agent so, they can be observed
by hyper intense part of the tumor core. FLAIR images help to distinguish the
edema region from the cerebrospinal fluid (CSF).
The probability of appearance of tumor near to white matter is higher com-
pared to grey matter and cerebro spinal fluid. The segmentation of tumor from
the MR brain images is challenging due its shape, size, position, appearance
among the patients, contrast with the surrounding tissue and distortion. The
above limitation can be overcomed by proper image segmentation technique,
which divides the image into different segments to analyse various diagnostic
information. Analysing the image by the physicians is time consuming, tedious
and sometimes prone to errors. So, automatic segmentation has been used by
the researchers for computer aided surgery. In recent years, researchers have
proposed deep learning techniques using convolutional neural networks (CNN)
architecture to determine fine and detail features about the brain tumor. Still
the architectures can be improved to increase Dice score for the accurate segmen-
tation of the brain regions. I would like to further investigate and understand
the problem to develop sensitive methods for segmenting brain tumors .
