CHAPTER 1

1.1 BACKGROUND INTRODUCTION

Over the years, several researchers from various backgrounds have relied on image
recognition techniques for the identification of brain tumor cells (Amin et al., 2021) To
get the optimum performance, they have used a variety of machine learning techniques
to detect cancerous cells. Advanced neural networks and deep learning techniques are
also utilized. For instance, advanced neural networks, graph-based CNN, and CNN are
employed to improve the detection of malignant lesions in breast mammograms (Zhang
et al., 2021b). A convolutional neural network with exponential linear units and rank-
based weighted pooling is implemented for the early diagnosis of optimal therapeutic
intervention (Zhang et al., 2021a).
(Wadhwa et al., 2019) examined various methods for tumor identification and proposed
that combining Conditional Random Field (CRF) with FCNN and CRF with Deep Medic
or Ensemble offers better performance than the other approaches for tumor
segmentation. In (Özyurt et al.,2019), for segmentation, (Zyurt et al., 2021) proposed
the use of the neutrosophic set expert maximum fuzzy-sure entropy (NS-EMFSE)
process, and SVM and KNN classifiers were used to remove segmented functionality
from the CNN architecture. Recently, CNN has been employed by many researchers for
image classification in the domain of medical sciences (Ayadi et al., 2021), (Jin et al.,
2020) (Kalaiselvi et al., 2022), (Mohsen et al., 2018), (Murthy et al., 2022), (Rehman et
al., 2021), (Suganthe et al., 2020). Good performance results are reported using
advanced neural network models for MRI scan classification (Liu et al., 2018).
Automated classification is very useful in computer-aided diagnosis systems. Ensemble
models combining SVMs and neural networks are also implemented for the design of
medical diagnosis systems (Deepak et al., 2021). Soft computing techniques like fuzzy
logic are also incorporated for better results (Jayachandran et al., 2013). Advanced
fuzzy methods like adaptive fuzzy C-means clustering are used for segmentation.
Results are further improved by the deer hunting optimization algorithm (Murthy et al.,
2022).
(Subhaninia et al. 2018) suggested that medical image recognition relies heavily on
image segmentation, because medical photographs are too diverse, and used MRI and
CT scan images to segment the brain tumor. The most common use of MRI is for brain
tumor segmentation and classification. They proposed the use of fuzzy C-Means
clustering for tumor segmentation, which can reliably model tumor cells. After
segmentation, classical classifiers and CNNs were used to classify the data. They
implemented and compared the effects of various conventional classifiers such as K-
Nearest neighbor, logistic regression, multilayer perceptron, nave bayes, random forest,
and support vector machine in the traditional classifier section. SVM had the best
precision of 92.42 percent among these conventional ones. They also introduced CNN,
which yielded 97.87 percent accuracy with a split ratio of 80:20 of 217 photographs and
suggested experimenting with 3D brain images in the future to accomplish more
effective brain tumor segmentation. Working with a wider dataset would be more difficult
in this regard, and they aspired to create a dataset that emphasizes the abstract about
their region, which will help them expand the reach of their research.

1.2 INTRODUCTION
These days cancer has become so common that any individual can be prone to it. A
brain tumor is one kind of tumor. But early detection and diagnosis can help to heal the
cancer. Today in the medical field there are various techniques to detect brain tumors.
Segmentation of brain tumors from MRI using deep learning is one of the methods used
to detect brain tumors. This method undergoes brain tumor detection by learning the
properties of the cancerous tissues and diagnosing the tumor. It will also note whether
the condition of the brain is normal or abnormal. As there is a problem in the
segmentation of brain tumors for multi-modal images due to not predicting the shapes
and sizes of tumors in the brain. Sometimes there are many variations in tumor
structures and representations and also the position where the tumor is present, hence
we require automatic methods for brain tumor segmentation. Some of the operations on
this image can be performed by using image processing
Abnormal growth of cells or tissues in the brain can lead to a brain tumor. Neither the
exact symptoms of a brain tumor nor the reasons that cause brain tumors are known
today. Thus, people may be suffering from brain tumors without realizing the gravity of
the situation. It is of paramount importance to detect and extract the tumors at their
early stages to save the patient’s life. The MRI is an important tool for the detection,
diagnosis, and monitoring of brain tumors. However, examining MRI scans is a
dexterous, time-consuming, and difficult process. Further, it is very difficult to detect
tumors manually, and the results may vary from one clinical expert to another based on
their experience. Effective classification and segmentation of MRI images are quite
challenging. The rationale is to build an expert system that would assist in the effective
diagnosis of cancerous cells in MRI scans of the brain.
Deep learning techniques have revolutionized medical image analysis, particularly in the
field of brain tumor segmentation. These advanced algorithms leverage the power of
artificial neural networks to automatically detect and delineate tumor regions from
medical imaging data. Deep learning has emerged as a powerful technique for medical
image analysis, particularly in the field of brain tumor segmentation. With the ability to
automatically learn intricate patterns and features from vast amounts of data, deep
learning algorithms have shown promising results in accurately delineating brain tumor
regions from medical images (Bauer et al., 2013)
The rate of brain tumor cases among all age groups is increasing at an alarming rate all
over the world. Gliomas (Hanif et al., 2017) are the most common form of brain tumors
found but they are comparatively low grade i.e. the life expectancy of the patient
suffering from glioma is several years. On the other hand, more aggressive tumors
reduce the life expectancy of the patient to a maximum of two years. In the earlier
stages of cancer, the growth of the tumor can be slowed down using containment
measures like chemotherapy, radiation, etc. MRI scans give optical images of the brain
which help the doctors to identify these tumors and these images also enable the
doctors to predict the growth of the tumor at early stages and help to reduce the growth
and impact of the tumor on the patient’s life. Different tumors like gliomas or
glioblastomas or meningiomas all present different levels of difficulties in identification
like gliomas and glioblastomas are much more difficult to localize. Such tumors (along
with the neighboring edema) have diffused and poorly contrasted shapes and properties
and exhibit structures like tentacles that make them difficult to segment (Havaei et al.,
2017).
Brain tumor segmentation is a critical task in medical image analysis, playing a crucial
role in the diagnosis, treatment planning, and monitoring of brain tumors. In recent
years, the advancement of deep learning techniques has shown great promise in
achieving accurate and automated segmentation of brain tumor regions.
Deep learning models, particularly convolutional neural networks (CNNs), have
demonstrated remarkable capabilities in learning intricate patterns and spatial
relationships from medical images. By training on large datasets of annotated brain
scans, these models can automatically extract relevant features and discern subtle
differences between tumor and non-tumor regions. The learned representations enable
precise delineation of tumor boundaries, aiding clinicians in accurate diagnosis and
treatment decisions.
Numerous studies have highlighted the effectiveness of deep learning segmentation
approaches in brain tumor analysis. For instance, (Havaei et al., 2017) conducted a
comprehensive survey on MRI-based medical image analysis for brain tumor studies.
Their review discussed the utilization of deep neural networks for brain tumor
segmentation, emphasizing the significant advancements and improved performance
achieved through these techniques. Deep learning models have shown superior
accuracy and robustness, enabling accurate identification and localization of tumor
regions, even in the presence of challenging imaging conditions.
Brain tumor segmentation consists of extracting the tumor region from healthy brain
tissues; the existence of brain tumors can often be detectable. However, accurate and
effective segmentation of tumors remains a challenging task, since the tumors can
occur in different regions of the brain and their shapes and sizes are almost always
different. These tumors are very complex in shape and do not exhibit any rigid shape or
size and they have very different appearance properties.
Magnetic Resonance Imaging (MRI) is a preferred technique widely employed by
radiologists for the evaluation and assessment of brain tumors (Bauer et al., 2013). It
provides several complimentary 3D MRI modalities acquired based on the degree of
excitation and repetition times, i.e. T1-weighted, post-contrast T1-weighted (T1ce), T2-
weighted, and Fluid Attenuated Inversion Recovery (FLAIR). The highlighted sub-
regions of the tumor across different intensities of these sequences (Bakas et al., 2018),
such as the whole tumor (the entire tumor inclusive of infiltrative edema), are more
prominent in FLAIR and T2 modalities. In contrast, T1 and T1ce images show the tumor
core exclusive of peritumoral oedema (Menze et al., 2010). It allows for the combinative
use of these scans and the complementary information they deliver toward the detection
of different tumor sub regions.

1.3 STATEMENT OF THE PROBLEM

Deep learning has shown promise in correctly segmenting brain tumors from medical
imaging. However, certain obstacles must be solved in order for these models to be
more successful. To begin, the scarcity of different training data offers a substantial
challenge, as getting a big annotated dataset is difficult due to the rarity of particular
tumor forms and the necessity for expert annotations. Brain tumors can vary greatly in
terms of size, shape, location, and appearance. Deep learning models must be strong
enough to manage these variances and segment tumors effectively across patients and
imaging modalities. Furthermore, the issue of class imbalance develops, in which the
quantity of tumor pixels is considerably lower than that of background pixels, resulting in
biased models that fail to perform well.
Furthermore, deep learning models' interpretability and trustworthiness are critical for cli
nical adoption. Explainable models are required to offer light on the decision-
making process and allow doctors to have confidence in the segmentation outcomes.
Finally, models trained on a single dataset may not perform well on fresh data from vari
ous sources or obtained using different imaging techniques, therefore guaranteeing gen
eralization to unknown data is critical.
To overcome these problems and obtain reliable and accurate brain tumor segmentatio
n for enhanced diagnosis, treatment planning, and patient monitoring in real-world clinic
al situations, adaptable and resilient models are required.

1.4 RESEARCH GAP

The problems encountered when creating deep learning segmentation of brain tumors
(Using segmentation techniques), and potential solutions are as follows:
i. The lack of annotated brain tumor photos makes training reliable deep learning
models challenging. To address this, one technique is data augmentation, which
involves augmenting accessible data with synthetic instances or altered copies of
existing photos. Collaborations with medical institutes and the implementation of
data-sharing programs can also aid in the collection of bigger and more
diversified datasets for training.
ii. Deep learning models have difficulties in generalizing due to the varied
properties of brain tumors. Transfer learning may be used on large-scale
datasets by using pre-trained models. The network may learn key properties and
respond to unique tumor changes by fine-tuning these models using brain tumor
pictures.
 iii. Deep learning models are frequently difficult to comprehend, making it difficult for
physicians to trust the segmentation findings. Methods such as attention
mechanisms or post-hoc interpretability approach such as Grad-CAM can be
used to improve interpretability. These strategies emphasize the parts of input
photos that are most important to the model's judgment, revealing information
about the model's segmentation process.
iv. Deep learning models may struggle to generalize to unseen data due to domain
shifts or variations in imaging protocols. Domain adaptation techniques, such as
domain adversarial training or domain-specific normalization, can be applied to
bridge the gap between different data sources. This helps the model adapt to
unseen data and perform well in real-world clinical settings.

1.5 RESEARCH QUESTIONS

i. What effect does network design and hyperparameter selection have on t
he accuracy and efficiency of brain tumor segmentation?
In terms of data availability, data quality, and interpretability, what are the 
problems and limits of deep learning segmentation?
ii. Can deep learning models handle multi-modal medical imaging data well, 
such as merging MRI and PET images for better brain tumor segmentatio
n?
iii. What are the deep learning methodologies for dealing with class imbalanc
e and short sample numbers in brain tumor segmentation tasks?
iv. Deep learning models trained on one dataset to segment brain tumors in u
nknown datasets are how generalizable?
v. What are the most recent cutting-edge deep learning techniques for brain
tumor segmentation
vi. How deep learning do based segmentation approaches fare in
comparison to older methods utilzed in clinical practice?

1.6 RESEARCH OBJECTIVE

i. To find out the current state-of-the-art deep learning approaches for brain
tumor segmentation
ii. To examine the performance of deep learning-based segmentation
methods compare to traditional methods used in clinical practice
iii. To analyze the choice of network architecture and hyperparameters affect
the accuracy and efficiency of brain tumor segmentation
iv. To determine the challenges and limitations of deep learning
segmentation in terms of data availability, data quality, and interpretability
v. To find out the deep learning models effectively handle multi-modal
medical imaging data, such as combining MRI and PET scans, for
improved brain tumor segmentation
 vi. To examine the strategies for addressing class imbalance and small
sample sizes in brain tumor segmentation tasks using deep learning
vii. To find out how generalizable are deep learning models trained on one
dataset to segment brain tumors in unseen datasets
viii. To find out if deep learning models assist in predicting tumor malignancy
or other clinically relevant tumor characteristics based on the segmented
regions

1.7 DATA SOURCE

i. The Cancer Imaging Archive (TCIA): TCIA provides a wide range of
medical imaging data, including brain tumor datasets.
ii. BraTS (Multimodal Brain Tumor Segmentation Challenge): BraTS is a
well-known dataset specifically designed for brain tumor segmentation
research. It contains multimodal MRI scans of brain tumor patients.
iii. MICCAI Brain Tumor Segmentation (BRATS) Challenge: MICCAI BRATS
is an annual challenge that provides datasets for brain tumor
segmentation.
iv. Hospital and Research Institution Data: Many hospitals and research
institutions have their own databases of medical images, including brain
tumor data. These datasets might require permission or collaboration with
the respective institutions.
v. Data Sharing Platforms: Researchers often share their datasets through
platforms such as GitHub, Kaggle, or Zenodo. These platforms can be a
good source for finding publicly available brain tumor segmentation
datasets shared by the research community.
vi. Data Annotation Services: Some companies and organizations specialize
in providing annotated medical imaging datasets for research purposes.
These datasets are typically labeled by experts, which can be helpful for
training deep learning models.

1.8 RESEARCH METHODOLOGY

i.

1.9 EXPECTED OUTCOME