Introduction

CHAPTER 1
INTRODUCTION

One of the most deadly conditions that can impact anyone is a brain tumour. without regard to
age, gender, or race. Unusual brain growth is referred to as a brain tumour either normal or
cancerous tissue [1]. Non-cancerous benign tumours develop from tissues connected to the brain.
Tumors that are malignant are cancerous, rapidly expanding, and fatal.Thus, to improve patient
survival, early diagnosis of brain cancers is crucial [7].

The expertise of neurologists is required for accurate tumour detection. The dramatic rise in
brain tumour occurrences in recent years has forced researchers to automate their work of a brain
tumour finding. Neurologists benefit from computer-aided diagnosis in numerous ways methods
that include.

1. It reduces human intervention and helps early detection of tumors for better medication.

2. It assets the physician as a second opinion while making the final decision .

The crucial step in brain tumour diagnosis is brain tumour classification. hence, researchers are
driven to put a brain pathology detecting system into place. Pathology identification system
divides the provided brain picture into normal and pathological categories. Nayak and others
have suggested quick curvelet entropy characteristics for detecting diseased brains. Siyuan and
others developed a pre-trained AlexNet with transfer learning for detecting diseased brains.

A deep neural network based on a deep wavelet autoencoder has been proposed by Mallick et al.
for detecting cancer. Kaur et al. used parameter-free BAT and fisher criteria. categorization of
brain tumour optimization algorithm. Hemanth et al.'s modification identification of brain
abnormalities using a genetic algorithm.

Multi-class classification, as opposed to binary classification, concentrates on the type of brain

tumour.concomitant pathology detection. More than 120 different forms of brain tumours exist,
and the Meningioma (35%), glioma (16%), and pituitary (14%) are the three most prevalent
forms of brain tumours. Magnetic resonance (MR) images with contrast-enhanced T1 weighting
of each tumour, the green line highlighting the tumour location. To determine the type of tumour,
there are a number of criteria to consider, including intensity, texture, location, size, and shape.
Meningioma and pituitary tumours are shown in Figure 1a and c, respectively. These Tumors
have similar textures and homogenous intensities. Although these Tumors can be identified by
their location, form, and size. pituitary meningioma The pituitary gland and the area around the
skull are where cancers grow, respectively. The glioma tumour is different intensities[32].

The human expert can predict the sort of tumour by looking at these qualities. Due to the
complexity of the computer-aided tumour classification system many features of brain tumours.
This research addresses multi-tumor classification using a residual network with transfer
learning. The suggested model makes use of To address the overfitting issue, use global average
pooling.

A brain tumour, which is regarded as one of the most serious illnesses of the nervous system, is
an unexpected and uncontrollable development of brain cells. The National Brain Tumor Society
(NBTS) estimates that around 400,000 people worldwide are affected with brain tumours every
year, and that 120,000 people die from them. These numbers are rising annually. Gliomas, the
most prevalent primary brain tumour in adults, severely harm the central nervous system. Low-
grade glioma (LGG) and high-grade glioma (HGG) are the two common classifications of
glioma, which starts in the brain's glial cells [32].

The HGG is more malignant than LGG because it spreads quickly and patients with HGG often
have a life expectancy of two years.

Brain tumours can now be visualised using a variety of magnetic resonance imaging (MR)
sequences, such as T1-weighted, T2-weighted, T1-weighted with contrast enhancement (T1c),
and fluid-attenuated inversion recovery (FLAIR). For medical diagnosis, treatment planning, and
surgical planning, the precise segmentation of brain tumours is extremely important.

However, due to the shape, size, and placement of brain tumours, which differ significantly
between patients, automatically segmenting brain tumours is a difficult task. Due to hazy borders
between nearby structures and smooth intensity gradients, segmenting the core and enhanced
tumours is very difficult.
When it comes to DL-based techniques, 3D convolutional neural networks (CNNs) are thought
to be better suited for volumetric segmentation tasks than 2D CNNs because they require more
computing power. In contrast, 2D-CNN takes less computing resources and training data and
explores the existence of tumours in each slice in a 2D manner. However, 2D-CNNs are unable
to interpret the 3D sequential data, which is necessary for volumetric segmentation, which
negatively impacts segmentation performance [32]. In order to take advantage of both forms of
architecture, we used a hybrid technique that makes use of a trade-off and can take advantage of
inter-slice sequential information while using far less memory than 3D CNNs.

1.2 Objective

A brain tumor is a mass of aberrant and unneeded cells that are developing in the brain and
is a condition that poses a threat to life. Therefore, it is more important to segment and finds such
tumors via Magnetic Resonance Imaging (MRI) at an early stage to preserve life. When it comes
to finding people with brain cancer, MRI is so effective that its detection rate is slightly greater
than that of other imaging modalities. The detection of brain tumors is very difficult to work in
the system of medical imaging because of the fluctuations in size, shape, and appearance. As a
result, the suggested Improved Invasive Bat (IIB)-based Deep Residual network model is used to
create an effective brain tumor detection technique. The Improved Invasive Weed Optimization
(IWO) and Bat algorithm (BA), respectively, are therefore incorporated into the proposed IIB
algorithm. Brain tumor early detection is greatly impacted by the segmentation of tumors with
MR images. The deep learning-based approach successfully produced improved detection
outcomes with MR images. Features are obtained from the tumor regions using the segmentation
findings, and these features are then applied to the Deep Residual network detection procedure.

1.3 Overview

The goal of machine learning is to automatically spot important data patterns and solve
problems that are difficult (or impossible) to represent with explicit processes. Deep learning
(DL), a subfield of machine learning, derives ideas like object categories directly from sensory
data. The proliferation of low-cost, multi-processor graphics cards is closely tied to the success
of DL. CNN can be used to detect COVID-19 in patients based on chest CT or X-ray scans and
MRI because of its capacity to extract information from visual cues. Its fundamental composition
is shown as a series of Convolution
Convolutional—Pooling—Fully
Fully Connected Layers(Fg 1.1)
1 with
additional intermediary layers for dropout and normali
normalization.

Figure. 1.1Generic
Generic two-dimensional
two 31]
CNN for COVID-19 detection.[31

Medical duties include the diagnosis of retinopathy, pneumonia, cardiomegaly, as well as some
cancer types, and have effectively used it. The majority of papers discussed in the literature
classify CXR images in "no-finding,"
finding," pneumonia, and COVID-19
COVID using transfer learning. Even
though the authors looked at two different sets of photos, only one database was used for each
class, which might have influenced the model that was developed. It is anticipated that using an
training and testing could lead to erroneously better model
enhanced rendition of a picture for training
performance.

based models that don't rely on previously trained networks and are built
There aren't enough AI-based
on open databases with radiological images from various sources (e.g., ResNet, VGG, DeTraC,
DeTraC
SqueezeNet). The authors only employed two databases, each of which contained a certain
finding" or COVID19 related), after using transfer learning to refine their
image type (either "no-finding"
VGG 19 and ResNet-50
model. Their model outperformed the widely used VGG-19 ResNet models. Improper
ation may lead to incorrect diagnostic conclusions, which may ultimately lead health
data utilization
professionals to make the wrong management decisions for patients [31].

1.4 Convolutional Neural Networks

Online learning can result in "zig zagging" behavior, in which the gradient estimate from a single
"zig-zagging"
data point changes course and does not directly approach the minimum. Rectified linear units are
a simple and effective approach to building a non-linear network (ReLu). Convolutional layer's
objective is to teach input feature representations. Translation invariance is ensured through
parameter sharing that results from convolution. The effect of the pooling layer is to decrease the
dimensions of the feature maps and improve the robustness of feature extraction.

By stacking many convolutional layers and pooling layers, we may extract high-level properties
from inputs. It is helpful to reduce the size of the activation map in the deep end of the network
to make the network easier to maintain for classification. The additional benefit of pooling is that
it increases the translation invariance of the final network. Information regarding the spatial
relationships between the pattern's component pieces may be lost during pooling. Using
convolutional layers with a higher stride value allows pooling layers to frequently be eliminated
without sacrificing accuracy, according to research[3].

Typically, data is divided into three sets: training, validation, and test data. The data used to train
the network is known as the training data set. Images that were not included in either the
network's training or validation make up the test set. It shows how effectively the network
generalizes to unknowable data. To determine the most accuracy, flow architecture and the
backpropagation algorithm have been very successful. To train a neural network to approach
target outputs from known inputs, the weights of each neuron are chosen. The back-propagation
technique offers a straightforward and efficient way to solve the weights[8].

1.5 Deep Learning

In this chapter, one of the subfields of artificial intelligence, we examine neural networks
and deep learning (AI). Neural networks, also known as artificial neural networks (ANNs) or
simulated neural networks, are the foundation of deep learning techniques. They mimic the way
that neurons communicate by modeling themselves after the human brain. Natural language
processing, computer vision, and other fields make use of convolutional neural networks.
Convolution neural networks have a three-dimensional arrangement of neurons rather than the
typical two-dimensional array.
They are made up of input layers, hidden layers, and output layers. The employment of multiple
layers of non-linear, hierarchically structured information processing stages is what defines deep
learning. These methods have increased the state-of-the-art in voice recognition, visual object
identification, object detection, and other areas in numerous other fields, such as drug
development and genomics. Deep architectures with built-in discrimination are intended to
provide direct discrimination for classification tasks. They describe the distributions of the
various classes given the available information. These models are used by unsupervised machine
learning to perform tasks including likelihood and probability estimations, modeling of data
points, and class separation[1].

It is common to practice using Sigmond's derivative on the output rather than the input. The local
gradient is crucial for backpropagation because, if it is too little, the gradient would be killed and
the network would be unable to learn. The gradient of the Tanh function is significantly steeper
than that of the sigmoid curve. The fact that the ReLU function does not simultaneously activate
all of the neurons is its main advantage over other activation functions. The weights and biases
of some neurons are not updated as a result of "dying ReLU," which is a problem with this
system[1].

Dead neurons that never fire as a result of this are possible. The most popular child among
optimizers is Gradient Descent. Gradient descent works well for convex functions, but it cannot
decide how far to descend the gradient for nonconvex functions. We now comprehend the
potential drawbacks of using gradient descent on sizable data sets. Gradient descent takes a
considerably quieter path than stochastic gradient descent (SGD). To get the perfect minimum,
additional iterations are required, which makes the process slower. Consider using a high
momentum and learning rate to hasten the process even more.

Compared to stochastic or batch gradient descent procedures, the mini-batch gradient descent
method is quicker and more dependable. The adaptive gradient descent technique used by
Adagrad uses different learning rates for each iteration, doing away with the requirement for
manual learning rate adjustment. AdaGrad's optimizer has some shortcomings. Gradient descent
and the AdaGrad variant of gradient descent are both used in RMSProp. Utilizing a decreasing
average of partial gradients, it adjusts the step size for each parameter.
The basic objective of RMSProp is to reduce the number of function evaluations required to find
local minima. In contrast to SGD training, which maintains a single learning rate, Adam
optimization alters the learning rate for each network weight separately. It has been modified to
serve as a benchmark for deep learning articles and is suggested as a typical optimization
strategy. The adam optimizer is frequently used due to all of its benefits.

1.6 Type of Region-Based CNN Algorithms

A fully linked layer in a convolution neural network (CNN) cannot handle multiple items
and frequency of recurrence. In 2013, Ross Girshick et al. proposed the R-CNN (Region-based
CNN) architecture for object detection[21].

1.6.1 R-CNNs

Ross Girshick et al. suggested a way where we utilize selective search to extract just 2000
regions from the image to get over the issue of selecting a large number of regions. This squared-
off representation of the 2000 candidate region proposals is input into a convolutional neural
network. There are many different types of lung disorders, including lung nodules and diffuse
lung diseases. The state-of-the-art in pattern recognition in the areas of voice and vision has been
significantly enhanced by deep learning techniques.

Utilizing convolutional neural networks, we created an image-based CADe for the detection of
lung anomalies (CNN).372 patients, with or without widespread lung abnormalities on CT scans,
were used in this study. These imaging cases included 167 emphysema cases, 121 ground-glass
opacity cases, 56 consolidation cases, 56 honeycombing cases, and 55 normal cases (NOR). R-
CNN is an object detection framework that uses a CNN to categorize different image regions. R-
CNN uses a selective search to produce bounding boxes or region suggestions. It applies a CNN
model after warping the area to a regular square size[12].

1.6.2 Fast R-CNN

The task of object detection is more difficult and necessitates the use of more sophisticated
techniques. Current methods include slow, cumbersome multi-stage pipelines for training
models. We provide a single-stage training technique that simultaneously gains the ability to
categorize object suggestions and improve their spatial positions. It was suggested that spatial
pyramid pooling networks (SPPnets) could accelerate R-CNN by sharing computation. Using a
feature vector that was taken from the common feature map, SPPnets identifies each object
proposition after creating a convolutional feature map for the full input image.

A whole image as well as a list of potential objects are fed into a FastR-CNN network. The
network uses many convolutional (conv) and max pooling layers to process the entire image. A
region of interest (RoI) pooling layer extracts a fixed-length feature vector for each object
proposition. With a multi-task loss, the architecture is trained from beginning to end. Each RoI is
combined into a fixed-size feature map, which fully connected layers then map to a feature
vector. Each RoI in the network has two output vectors: bounding-box regression offsets and
softmax probabilities[13].

Weights below the spatial pyramid pooling layer cannot be updated using SPPnet. When each
training sample (i.e. RoI) originates from a separate image, back-propagation through the SPP
layer is extremely inefficient. Stochastic gradient descent (SGD) micro batches are sampled
hierarchically during Fast RCNN training. We sample 64 ROIs from each image using R = 128
mini-batches. Images have a 0.5.3 likelihood of being horizontally flipped[13].

Derivatives are transmitted by backpropagation through the RoI pooling layer. N = 2 evenly
picked random images make up each SGD mini-batch. During both training and testing, every
image is processed at a specific pixel size. Through an image pyramid, the multi-scale technique
gives the network approximative scale invariance. Due to GPU memory restrictions, we only test
multi-scale training on smaller networks[13].

Compressing them makes it simple to accelerate large, fully connected layers. For whole-image
classification, truncated SVD lowers the parameter count from UV to t(u + v). When there are
many ROIs, this straightforward compression technique provides good speedups[13].

1.6.3 Faster R-CNN

Ross Girshick, Shaoqing Ren, Kaiming He, and Jian Sun introduced the Faster RCNN object
identification architecture in 2015. To identify things, it employs convolution neural networks
like YOLO (You Look Only Once) and SSD (Single Shot Detector). Ross B. Girshick proposed
a faster R-CNN in 2016. The performance as a whole has significantly improved, particularly in
terms of detecting speed. It generates the proposed box using the convolution network and
decreases the number of recommended frames from roughly 2000 to about 300[16].

A series of fundamental Conv with reflux and pooling layers by Faster R-CNN is used to extract
image feature maps. After 3*3 convolution, the image size becomes M*N. The output length and
width are reduced to half of the input via pooling layers. The typical sliding window and SS
technique are dropped in favor of a faster R-CNN that employs RPN to directly produce
detection boxes. The softmax classification anchor provides the foreground and background for
the first branch. Classification layers employ bounding box regression to determine the final
precise position of the checkbox and proposal feature maps to determine the proposal's class. A
portion of the network structure of the classification layers is illustrated in Fig. 1.2 to create a
more precise rectangle box.

Figure. 1.2 Part of the network structure of the classification layer[10]

1.6.4 Mask R-CNN

Pixel-to-pixel alignment in Mask R-CNN fills in the primary gap in Fast/Faster. The
additional mask output requires extraction of an object's considerably more precise spatial
arrangement and is unique from the class and box outputs. In contrast, the majority of modern
systems rely on mask predictions for classification. The spatial layout of an input object is
encoded by a mask. Using an FCN, we forecast an m m mask from each RoI.
This prevents the explicit m-m object spatial layout from being collapsed into a vector
representation that lacks spatial dimensions in any of the mask branch's layers. A common
procedure known as RoIPool is used to extract a small feature map (such as a 7 7) from each
RoI. Misalignments between the RoI and the retrieved features are introduced by quantization.
We suggest a RoIAlign layer to mitigate RoIPool's harsh quantization. Mask R-CNN feature
extraction with a ResNet-FPN backbone results in good accuracy and speed improvements.
Since res5 is already present in the FPN backbone, a more effective head with fewer filters is
possible. Although more intricate designs have the potential to perform better, they are not the
main emphasis of this work[27].

If an image has an IoU with a ground-truth box of at least 0.5, it is deemed positive; otherwise, it
is considered negative. Only positive ROIs are used to determine mask loss Leask. Images are
resized to have an 800-pixel scale (shorter edge). We add RoIAlign functionality to two already
available Faster R-CNN heads. We employ ReLU in hidden layers, and the output conv is 11.
The decons are 22 with stride 2. Any coordinates utilized by the RoI or its bins are not subjected
to any quantization[27].