Media Engineering and Technology Faculty

German University in Cairo

Deep Convolutinal NN for Liver

Tumor Segmentation in CT Scans

Bachelor Thesis

Author: Amir Zaghloul El-Sayed Gharib

Supervisors: Assoc. Prof. Mohammed Abdel-Megeed Salem
Submission Date: 26 May, 2019
 This is to certify that:

(i) the thesis comprises only my original work toward the Bachelor Degree

(ii) due acknowledgment has been made in the text to all other material used

Amir Zaghloul El-Sayed Gharib

26 May, 2019
Acknowledgments

First and foremost, I have to thank God for the strength, knowledge and ability he
provided me to finish this work without his work this work would never have been finished.

I would also like to thank my Supervisor Assoc. Prof. Mohammed Abdel-Megeed

Salem for his continuous support, guidance and patience for the past few months, his
guidance helped me overcome all the obstacles that faced me during this work.

I would have achieved nothing if not for my family and their moral and unwavering
support and for pushing me forward when i most needed the push, I can not deny the
help I got from my friends and colleagues whenever I was facing problems or any kind of
obstacles.

III
Abstract

Liver cancer has been as the second most fatal cancer to cause death for men and sixth
for women. Early diagnosis by Computed Tomography could lead to high recovery rate,
however going through all the CT slices for thousands or even millions of patients man-
ually by professionals is hard, tiresome, expensive, time-consuming and prone to errors.
Therefore, we needed a reliable, simple and accurate method to automate this process.
In this thesis we used CNNs to overcome all the aforementioned obstacles, we used a Re-
sUNet model on the 3D-IRCADb01 dataset which contains CT slices for patients along
with masks for liver, tumors and other body organs. ResUNet is a hybrid between the U-
Net and ResNet models where it uses Residual blocks rather than traditional convolution
blocks. We used 2 Cascaded CNNs one for segmenting the liver and extracting the ROI
and the second one use the extracted ROI from the first CNN and segment the tumors.
We achieved dice coefficient of up to 95% and True Value Accuracy of up to 99%.

IV
Contents

1 Introduction 1
1.1 Liver Anatomy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Liver Tumors and Stages . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Computed Tomography CT . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.4 Current State of Liver Tumor Segmentation . . . . . . . . . . . . . . . . 6
1.5 Problem Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.6 Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.7 Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

2 Background 8
2.1 Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.2 Concepts Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.2.1 Deep Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.2.2 Artificial Neural Networks . . . . . . . . . . . . . . . . . . . . . . 11
2.2.3 Convolutional Neural Network . . . . . . . . . . . . . . . . . . . . 11
2.2.4 Activation Functions . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2.5 ReLU . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.2.6 Vanishing Gradient Problem . . . . . . . . . . . . . . . . . . . . . 13
2.2.7 Max Pooling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.2.8 Batch Normalization . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.2.9 ResNet . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

3 Methodology 18
3.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
3.2 Data Preprocessing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.2.1 Hounsfield Windowing . . . . . . . . . . . . . . . . . . . . . . . . 19
3.2.2 Hisogram Equalization . . . . . . . . . . . . . . . . . . . . . . . . 20
3.3 Data Augmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.3.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.3.2 Merging Liver Tumors . . . . . . . . . . . . . . . . . . . . . . . . 22
3.3.3 Reflection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.3.4 Rotation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
3.4 Liver Segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.5 Tumor Segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

V
 3.6 U-Net . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.6.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.6.2 U-Net Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.6.3 U-Net Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.7 ResUNet . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.7.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.7.2 Residual Units . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.7.3 ResUNet Architecture . . . . . . . . . . . . . . . . . . . . . . . . 30

4 Results 32
4.1 Experiment setting and Dataset . . . . . . . . . . . . . . . . . . . . . . . 32
4.1.1 Tools used . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
4.1.2 Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
4.2 Experiments and Results . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
4.2.1 Evaluation Metrics . . . . . . . . . . . . . . . . . . . . . . . . . . 33
4.3 Liver Segmentation Results . . . . . . . . . . . . . . . . . . . . . . . . . 33
4.3.1 Liver Segmentation Model Evaluation . . . . . . . . . . . . . . . . 33
4.4 Tumor Segmentation Results . . . . . . . . . . . . . . . . . . . . . . . . . 39
4.4.1 Tumor Segmentation Model Evaluation . . . . . . . . . . . . . . . 39
4.4.2 ResUNet Model Limitations . . . . . . . . . . . . . . . . . . . . . 45

5 Conclusion and Future Work 47

5.1 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
5.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47

Appendix 48

A Dataset and Code 49

VI
List of Figures

1.1 Anatomy of the Liver . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.2 Stage 1A of Liver Tumors . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Stages 2A & 2B of Liver Tumors . . . . . . . . . . . . . . . . . . . . . . 3
1.4 Stages 3A & 3B of Liver Tumors . . . . . . . . . . . . . . . . . . . . . . 4
1.5 Stages 4A & 4B of Liver Tumors . . . . . . . . . . . . . . . . . . . . . . 5

2.1 Architecture of the CNN proposed by [10] . . . . . . . . . . . . . . . . . 8

2.2 Schematic representation of the network architecture prposed by [11] . . 10
2.3 Simple Neuron structure [19] . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.4 Sigmoid Activation Function [17] . . . . . . . . . . . . . . . . . . . . . . 12
2.5 Hyperbolic Tangent function- Tanh [17] . . . . . . . . . . . . . . . . . . . 13
2.6 Variations of ReLU Functions [17] . . . . . . . . . . . . . . . . . . . . . . 13
2.7 Sigmoid function and its derivative [22] . . . . . . . . . . . . . . . . . . . 14
2.8 Applying Max pool on a function . . . . . . . . . . . . . . . . . . . . . . 15
2.9 Max pooling effect on the 3-axes . . . . . . . . . . . . . . . . . . . . . . . 15
2.10 Residual Blocks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

3.1 Overview of the proposed workflow . . . . . . . . . . . . . . . . . . . . . 18

3.2 CT slice before and after HU windowing . . . . . . . . . . . . . . . . . . 20
3.3 CT slice not containing liver before and after HU windowing . . . . . . . 20
3.4 CT slice before and after HU windowing & Histogram Equalization, the
increase in contrast on the image on the right makes the tumors more
visible and thus easier to segment . . . . . . . . . . . . . . . . . . . . . . 21
3.5 CT slice not cotaining liver before and after HU windowing & Histogram
Equalization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.6 Mask for the first tumor . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.7 Mask for the second tumor . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.8 Final mask for both tumors merged together . . . . . . . . . . . . . . . . 23
3.9 Slice before & after reflection . . . . . . . . . . . . . . . . . . . . . . . . 23
3.10 Liver mask before & after reflection . . . . . . . . . . . . . . . . . . . . . 24
3.11 Tumor mask before & after reflection . . . . . . . . . . . . . . . . . . . . 24
3.12 Slice before & after rotation by 90◦ . . . . . . . . . . . . . . . . . . . . . 25
3.13 Liver mask before & after rotation by 90◦ . . . . . . . . . . . . . . . . . . 25
3.14 Tumor mask before & after rotation by 90◦ . . . . . . . . . . . . . . . . . 25
3.15 Liver Segmentation example . . . . . . . . . . . . . . . . . . . . . . . . . 26

VII
3.16 Liver Segmentation without liver . . . . . . . . . . . . . . . . . . . . . . 26
3.17 Tumor Segmentation example . . . . . . . . . . . . . . . . . . . . . . . . 27
3.18 Liver has no tumor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.19 Architecture of the U-Net CNN[16] . . . . . . . . . . . . . . . . . . . . . 28
3.20 Random sample #1 that shows the segmentation result of the U-Net
model. On the left is the True input value and on the right there is the
predicted label from the model . . . . . . . . . . . . . . . . . . . . . . . . 29
3.21 Random sample #2 that shows the segmentation result of a complex liver
shape using the U-Net model. On the left is the True input value and on
the right there is the predicted label from the model . . . . . . . . . . . . 29
3.22 ResUNet Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

4.1 Progress of training the ResUNet model for Liver segmentation on our
dataset for 20 epochs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
4.2 CT slice of the first random sample before liver segmentation . . . . . . . 35
4.3 Result of liver segmentation on the first random slice in figure 4.2. True
label on the left and predicted on the right, white is pixels with liver and
black is pixels with no liver . . . . . . . . . . . . . . . . . . . . . . . . . 35
4.4 Confusion Matrix from the predicted value of the first slice in figure 4.2
after liver segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
4.5 Another random sample of a CT slice before liver segmentation, as can
easily be observed the CT slice is augmented by rotation along with its mask 37
4.6 Result of liver segmentation on the second slice in figure 4.5. True label
on the left and predicted on the right, white is pixels with liver and black
is pixels with no liver . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
4.7 Confusion Matrix from the predicted value of the second slice in figure 4.5
after liver segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
4.8 Progress of training the ResUNet model on our dataset for 50 epochs . . 39
4.9 CT slice of the first random sample with the liver mask applied upon it
before tumor segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . 41
4.10 Result of tumor segmentation for the first random slice in figure 4.9. True
label on the left and predicted on the right, white is pixels with tumor and
black is pixels with no tumor . . . . . . . . . . . . . . . . . . . . . . . . 41
4.11 Confusion Matrix from the predicted value of the first slice in figure 4.9
after tumor segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . 42
4.12 CT slice of the second random slice with the liver mask applied upon it . 43
4.13 Result of tumor segmentation on the second random slice in figure 4.12.
True label on the left and predicted on the right, white is pixels with tumor
and black is pixels with no tumor . . . . . . . . . . . . . . . . . . . . . . 43
4.14 Confusion Matrix from the predicted value of the second slice in figure 4.12
after tumor segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
4.15 A sample CT slice with the liver mask applied upon it that contains a very
small tumor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

VIII
4.16 Result of tumor segmentation on slice in figure 4.15. True label on the
left and predicted on the right, white is pixels with tumor and black is
pixels with no tumor. It is obvious that the tumor was not predicted by
the model at all . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.17 Confusion Matrix from the predicted value of slice in figure 4.15 showing
that all the pixels containing tumors were misclassified . . . . . . . . . . 46

IX
List of Tables

3.1 Hounsfield Unit values for body organs [7] . . . . . . . . . . . . . . . . . 19

4.1 Dice coefficient results for training and validation data on the ResUNet
model for Liver Segmentation . . . . . . . . . . . . . . . . . . . . . . . . 34
4.2 Dice coefficient results for training and validation data on the ResUNet
model for Tumor Segmentation . . . . . . . . . . . . . . . . . . . . . . . 40

X
Chapter 1

Introduction

1.1 Liver Anatomy

The Liver is composed of 2 lobes figure 1.1 and it is one of the most essential organs in
the Digestive System. It carries alot of functions, one of the main function is processing
the absorbed nutrients by the small intestine. The Liver also secretes Bile juice into the
small intestine which aids in digesting fats. It also takes raw materials absorbed by the
intestine and makes the chemicals the body needs to function and detoxify the body from
harmful chemicals.[15]

Figure 1.1: Anatomy of the Liver

1
CHAPTER 1. INTRODUCTION 2

1.2 Liver Tumors and Stages

Liver cancer has been reported to be the second most frequent cancers to cause death in
men and sixth for women. About 750,000 people got diagnosed with liver cancer 696,000
of which died from the cancer in 2008. Globally the rate of infection of males is twice
than that of females. The highest infection rates are found in East and SouthEast Asia
and in Middle and Western Africa. Liver cancer incidence rate is increasing in many
parts in the world including United States and Central Europe, which is possible caused
by obesity and the increase in Hepatitis C Virus (HCV). [8].
Liver Tumors has many stages which require different treatment and diagnosis pro-
cesses figures 1.2, 1.3, 1.4 and 1.5 [20].

Figure 1.2: Stage 1A of Liver Tumors

1.3 Computed Tomography CT

Since the introduction of Computer Tomogrphy in the 1970s, CT scanning cause a revo-
lution in diagnostics. It led to improved cancer diagnosis and treatment, surgery, cardiac
conditions monitoring and injury treatment. It eliminated the need for exploratory surg-
eries. [14]
CT scanners use a motorized x-ray source on contrast with conventional x-ray ma-
chines which use a fixed x-ray tube. The motorized source rotates around the gantry. The
patient lies on a bed that moves through the gantry while the x-ray tube rotate around
them. CT scanners use digital x-ray detectors instead of films and then transmit the data
CHAPTER 1. INTRODUCTION 3

Figure 1.3: Stages 2A & 2B of Liver Tumors

CHAPTER 1. INTRODUCTION 4

Figure 1.4: Stages 3A & 3B of Liver Tumors

CHAPTER 1. INTRODUCTION 5

Figure 1.5: Stages 4A & 4B of Liver Tumors

CHAPTER 1. INTRODUCTION 6

to a computer where the computer use the received data from one rotation to build a 2D
image slice of the patient. The computer can also build a 3D model of the patient that
shows all organs scanned which aids the Medical Professionals to find the exact location
of the problem by rotating the model in space of view the slices in succession.[12]

1.4 Current State of Liver Tumor Segmentation

CT scans of the Liver are usually interpreted by manual or semi-manual techniques,
but these techniques are subjective, expensive, time-consuming and highly error-prone.
To overcome these obstacles and improve the quality of liver tumors’ diagnosis multiple
computer-aided methods have been developed. However these systems were not that great
at the segmentation of the liver and lesions due to multiple challenges like the low-contrast
between the liver and neighbouring organs and the liver and tumors, different contrast
levels in tumors, variation in the numbers and sizes of tumors, tissues’ abnormalities and
irregular tumor growth in response to medical treatment. Therefore a new approach must
be used to overcome these obstacles.

1.5 Problem Statement

We ought to try a new approach to overcome the obstacles we mentioned. One of the new
technologies that showed great results and great promise in Image Classification, Object
Detection, Face Recognition, Obstacle Avoidance, Material Inspection, Natural Language
Processing and many other applications is Convolutional Neural Networks (CNNs). Our
target is to build a robust CNN model to segment the Liver and detect the Region Of
Interest (ROI) from neighbouring organs using a CNN then use the ROI and detect liver
tumors using another CNN.

1.6 Objectives
The objective of our work in this thesis is to

1. Augment the 3D-IRCADb01 Dataset.

2. Build and Train CNN model to segment the Liver.

3. Build and Train a CNN model to segment the Tumors.

CHAPTER 1. INTRODUCTION 7

1.7 Outline
This Thesis is organized as follows:

• Chapter 1 gives a brief introduction to the liver anatomy, liver tumors and Com-
puted Tomography.

• Chapter 2 discusses the background and the related work done in liver tumors
segmentation and elaborate some concepts used in the thesis.

• Chapter 3 explains the data augmentation and preprocessing techniques, CNN mod-
els used in liver and tumor segmentation.

• Chapter 4 showcases the results we got from our CNN models in the segmentation
of the liver and its tumors.

• Chapter 5 summarizes the work we done in the thesis and future work that should
be done.

• Finally the references are found at the end of the thesis.

Chapter 2

Background

2.1 Literature Review

Grzegorz Chlebus et al.[2] discusses the Detection of Liver lesions using 2D Covolutional
Deep Neural Networks followed by a shape-based post-processing achieving an accuracy
of 77% and getting state-of-the-art results on the LiTS challenge.
Random-Forrest classifier was trained on the features produced from the Convolu-
tional Neural Network to filter False Positives and achieved an 87% accuracy.

Wen Li, Fucang Jia, Qingmao Hu[10] reduced the noise in CT images by the means of
applying Gaussian smoothing filter , then the result images were normalized and down-
sampled for faster training time.
The pre-processed images are then provided to the CNNs, 5 CNNs of different patch
size were created 13x13, 15x15, 17x17, 19x19 but the 17x17 outperformed all the other
CNNs and thus was the optimal choice

Figure 2.1: Architecture of the CNN proposed by [10]

8
CHAPTER 2. BACKGROUND 9

Avi Ben-Cohen et al.[1] used Fully Convolutional Network (FCN) in liver segmentation
and detection of metastases of Computed Tomography (CT) scans. Which proved to be
a powerful tool in the segmentation process. They tested on a small dataset with just 20
patients with an aggregate of 68 lesions and 43 livers in one slice and 20 different patients
with 3D liver segmentation. after cross-validation they achieved good and promising
results with true positive rate of 0.86 and 0.6 false positive for every patient.

After Data Augmentation they trained two networks one for lever segmentation which
isolated the neighbouring organs of the liver and the other for tumor and lesions segmen-
tation which operated on the result of the first network.

Patrick Ferdinand et al.[3] proposed a method to automatically segment the liver and
lesions from both MRI and CT abdominal images using 2 Cascaded Fully Covolutional
neural Networks (CFCNs) one of which is used in the segmentation of the Liver it self
and the other to detect lesions from the resulting Region Of Interest (ROI) of the first
CFCN.

In the Preprocessing phase of the CT images, First Hounsfield unit values were win-
dowed to a range of [-100, 400] to neglect neighbouring organs to the liver, then Histogram
Equalization was applied to the windowed image to increase contrast, lastly multiple data
augmentations steps were used to increase the data and teach the desired invariance prop-
erties like translation, rotation, elastic deformation and addition of Gaussian noise with
standard deviation.

Mohammed Goryawala et al.[4] discussed a new 3D-Segmentation method based on

the combination of a modified k-mean algorithm with a localized contouring algorithm.
During the Segmentation step this method it detects five seperate regions in the CT
images, In combination with the method and 3D-rendering the method proved to be fast
and accurate achieving a mean accuracy of 98.27%

Fausto Milletari, Nassir Navab, Seyed-Ahmad Ahmadi [11] proposed an approach to

medical image segmentation that utilized the power of Fully Convolutional Neural Net-
works (FCNNs), the FCN was trained end-to-end to process MRI Volumes. However
they opted for volumetric convolutions rather than preprocessing the input volumes in
slice-wise manner. They proposed a new objective function based on Dice coefficient
maximization. They showed accurate and fast results on the MRIs of the prostate.
CHAPTER 2. BACKGROUND 10

Figure 2.2: Schematic representation of the network architecture prposed by [11]

Zhengxin Zhang, Qingjie Liu and Yunhong Wang [23] worked on Road Extraction from
aerial images which is a fundamental task in remote sensing but its a hard task because
of noise pulled from aerial images , road extraction has a variety of applications like
unmanned vehicles navigation, plotting maps and updating geographical data.
After failing to achieve the optimal results with a U-Net model they came with a
new model that incorporates the advantages of a U-Net model with the advantages of
the ResNet model and called it ResUNet. ResUNet utilizes residual blocks which use
skip-connections instead of the traditional convolutions used by the traditional U-Net
thus resulting in faster training with less data.

2.2 Concepts Overview

2.2.1 Deep Learning

Deep Learning is part of the family of Machine Learning that is inspired by the biolog-
ical nervous system especially the part responsible of processing and communication of
Data. Deep Learning has many learning architectures, such as Deep Neural Networks
CHAPTER 2. BACKGROUND 11

and Recurrent Neural Networks. It can use supervised, semi-supervised or unsupervised

Learning. Its applications span multiple fields, such as computer vision, natural language
processing, speech recognition, analysis of medical images and many other applications.

2.2.2 Artificial Neural Networks

An Artificial Neural Network (ANN) is a paradigm for processing of information. It is

designed to mimic the work of the biological nervous system. An ANN is constructed
from a huge number of interconnected nodes called Neurons (refer to figure 2.3) that work
together in order to solve specific complex or simple problems. Like biological nervous
system, ANNs learn by example which is called Supervised Learning in which the ANN
is trained by receiving an input image along with the desired output image called Labels
[19].

Figure 2.3: Simple Neuron structure [19]

2.2.3 Convolutional Neural Network

Convolutional Neural Networks (CNNs) are similar to neural networks as they are made
up of neurons, where each neuron take several inputs take a weighted sum over them,
then feeds the sum to the activation function where it responds with the output. They
also have loss functions just like neural networks. Unlike neural networks CNNs operate
on volumes rather than vectors. They also contain convolution layers which are basically
of a set of independent filters that convolve on the input image.

2.2.4 Activation Functions

An Activation functions is crucial for Artificial Neural Networks in order for it to learn
and make sense of complicated, non-linear mapping of functions between inputs and
CHAPTER 2. BACKGROUND 12

response variables (the variable you are measuring). Activation Functions main purpose
is to convert a node’s input signals to output signals. They also introduce non-linearity
to the ANNs. Neurons apply the Sum of Products on its Inputs and Weights, applies the
Activation Function and then feed the Output to the next Neuron.[17]
Without Activation Functions the inputs and output of the Neurons will be mapped
linearly, which is easier to solve but lacks complexity and the ability to learn complex
functional mappings. An ANN without an Activation Function is merely a Linear Re-
gression Model, which does not have a good performance nor good power. Activation
Functions provide the ANNs with the ability to learn and model complicated kinds of
data like images, videos, audio and speech.[17]
Why is non-linearity so important? Non-linear functions have a higher degree than
one and they have curvatures. Neural Networks need to learn and represent complex
functions that maps inputs to outputs. As Universal Function Approximators Neural
Networks must be capable of representing any process as a functional computation. So
non-linear activation functions make non-linear mappings between inputs and outputs
possible. Examples of Activation Functions can be found in figures 2.4 and 2.5 [17]

Figure 2.4: Sigmoid Activation Function [17]

CHAPTER 2. BACKGROUND 13

Figure 2.5: Hyperbolic Tangent function- Tanh [17]

2.2.5 ReLU
Rectified Linear Units (ReLU) has become one of the most important activation func-
tions in Machine Learning and Deep Learning. It was proven that it showed sixfold the
improvement of conversion over the Tanh function with a very simple rule
R(x) = max(0, x)
. As we can observe from the mathematical rule the ReLU is so simplistic and it’s been
observed many times in Machine Learning and Deep Learning applications that the most
simplistic of techniques are preferred and lead to best results. On top of that ReLU avoids
the Vanishing Gradient problem. ReLU has many variations that solve its limitations
figure 2.6. [17]

Figure 2.6: Variations of ReLU Functions [17]

2.2.6 Vanishing Gradient Problem

Vanishing Gradient problem occurs with some activation functions when more layers are
added to the network as the loss function approaches zero. Functions like the sigmoid
CHAPTER 2. BACKGROUND 14

function fits a large input space into a small range between 0 and 1 which leads to small
changes in the output even with large changes in the input leading to a small derivative
figure 2.7.
As shown in figure 2.7 as the input of the sigmoid functions gets very large or very
small its derivative approaches zero. For Neural Networks with alot of layers a small
gradient cause by small derivative is a big issue for training effectiveness. Gradients of
Neural Networks are built by backpropagation which finds the derivatives of the network
by traversing the network layer by layer from the final layer to the initial layer. During
this traversal the derivative of each layer is multiplied with the derivative of the preceding
layer (traversing from final layer to initial layer) computing the derivative of the initial
layer. If we apply this to the sigmoid function we can see that the small derivatives of the
sigmoid function multiplied together yeild an exponentially smaller derivative and thus
smaller gradient. Small gradient minimizes the effectives of updating the weights and
biases of the initial layer which is important to recognizing core elements in the input
data thus can lead to inaccuracies in the whole network.[22]

Figure 2.7: Sigmoid function and its derivative [22]

2.2.7 Max Pooling

Max pooling down-samples the input representation whether it’s an image, audio, an
output from a hidden-layer, etc by sample-based discretization. Dimensionality reduc-
tion following the down-sampling allows for assumptions to be made about the features
contained in each sub-region.[21]
Max pooling minimizes computational costs using reduction of parameters to learn
and it also prevents over-fitting by abstracting the representation of data. Max pooling is
achieved by applying a max filter to contiguous sub-regions of the original matrix figures
2.8 and 2.9
CHAPTER 2. BACKGROUND 15

Figure 2.8: Applying Max pool on a function

Figure 2.9: Max pooling effect on the 3-axes

CHAPTER 2. BACKGROUND 16

2.2.8 Batch Normalization

Batch Normalization is used for improving the overall speed, performance and stability
of an ANN by normalizing the input layers and scaling the activation functions. It was
initially proposed to solve the Internal Covariate Shift problem, which occurs in networks
where the change of parameters in the input of the preceding layer lead to change of
distribution of the current layer and so on, so in deep networks the small changes are
amplified as they propagate along the networks layers resulting in a huge shift in the
deeper layers. Beside solving the Covariate Shift problem Batch Normalization leads
to better training rate without exploding gradients and it regularizes the network and
overcomes the need of using dropouts to avoid overfitting.[6]

2.2.9 ResNet

ResNet is like plain Artificial Neural Networks but it introduces a little modification to
the normal architecture called Residual Blocks. Residual Blocks are the main building
unit of ResNets and they utilize a very important concept called Skip Connections. The
main idea behind Skip Connections is to connect a layer with a successor layer of its
successor layer figure 2.10.[5]
Skipping layers simplify the network by using fewer layer in the initial training stages
and also accelerate learning by reducing the effect of the Vanishing Gradient problem due
to the fact that there are fewer layers to propagate through. Residual blocks reuse the
activation function from previous layers to learn the weight and then adapts to amplify
the skipped layer and mute the upstream layer.[5]
ResNets usually skip only one layer, However it can add a weight matrix to skip more
than one and then are called HighwayNets.[18]
CHAPTER 2. BACKGROUND 17

Figure 2.10: Residual Blocks

Chapter 3

Methodology

3.1 Overview
This section discusses the work we done in the bachelor project, it covers data pre-
processing, data augmentation and CNNs used to detect and segment liver tumors. Figure
3.1

Figure 3.1: Overview of the proposed workflow

18
CHAPTER 3. METHODOLOGY 19

3.2 Data Preprocessing

We Used 3D-IRCADb01 dataset and for every slice, we needed to preprocess every slice
in order to distinguish the liver from neighbouring Organs.

3.2.1 Hounsfield Windowing

Hounsfield Unit (HU) is the average of the attenuation values of a certain voxel compared
to the attenuation value of distilled water at standard temperature and pressure where
the HU of water is zero and air is -1000. It is encoded in 12 bits thus have 212 values which
is 4096 ranging from -1024 HU to 3071 HU table 3.1. It was named after the inventor of
CT-scanning Sir Godfrey Newbold Hounsfield, and it’s computed for any tissue as follows
where µ is the linear attenuation coefficient

1000 × (µtissue − µH2 O)

HU [9] =
µH2 O

The HU values assigned to each pixel are computed by assigning a gray-scale intensity
to each value, higher value mean brighter pixels.After slides are read in DICOM format,
Hounsfield Windowing was applied to ranges [-100, 400]. We can clearly see more details
after Hu windowing figures 3.2 and 3.3

HU values for body organs

Bone 1000
Liver 40 to 60
White Matter 46
Grey Matter 43
Blood 40
Muscle 10 to 40
Kidney 30
Cerebrospinal 15
Water 0
Fat -50 to -100
Air -1000

Table 3.1: Hounsfield Unit values for body organs [7]

CHAPTER 3. METHODOLOGY 20

Figure 3.2: CT slice before and after HU windowing

Figure 3.3: CT slice not containing liver before and after HU windowing

3.2.2 Hisogram Equalization

Histogram Equalization (HE) is a method that is used to enhance the contrast in images
especially when the contrast is in a narrow window and not distributed.[13].
Histogram Equalization was applied to the resulting image from the Windowing in
order to increase the contrast between the Liver and neighbouring organs figures 3.4 and
3.5, then the image is normalized to ranges [0,1].
CHAPTER 3. METHODOLOGY 21

Figure 3.4: CT slice before and after HU windowing & Histogram Equalization, the
increase in contrast on the image on the right makes the tumors more visible and thus
easier to segment

Figure 3.5: CT slice not cotaining liver before and after HU windowing & Histogram
Equalization

3.3 Data Augmentation

3.3.1 Overview

Since our dataset suffered from class imabalancing because the number of pixels not con-
taining tumors is much higher than the pixels containing tumors we had to augment the
dataset. Data Augmentation is a technique that is widely used in many Deep Learn-
ing and Machine Learning applications, to enlarge the training dataset in order to avoid
overfitting, fix class imbalancing and improve training accuracy. Data augmentation
CHAPTER 3. METHODOLOGY 22

incorporates different image transformation techniques, such as reflection, rotation, crop-

ping, translation and addition of noise. In this paper we used two techniques to augment
the data

3.3.2 Merging Liver Tumors

Since the 3D-IRCADb01 Dataset contains tumor masks for every tumor on its own,
we needed to merge all the different masks into 1 mask to facilitate the training and
Augmentation of data figures 3.6, 3.7 and 3.8

Figure 3.6: Mask for the first tumor

Figure 3.7: Mask for the second tumor

CHAPTER 3. METHODOLOGY 23

Figure 3.8: Final mask for both tumors merged together

3.3.3 Reflection

We reflect every slice that contains a tumor along the y-axis along with both the liver
mask and tumor mask in order to increase the number of infected slices figures 3.9, 3.10
and 3.11

Figure 3.9: Slice before & after reflection

CHAPTER 3. METHODOLOGY 24

Figure 3.10: Liver mask before & after reflection

Figure 3.11: Tumor mask before & after reflection

3.3.4 Rotation

We rotate every slice that contains a tumor along with both the liver mask and tumor
mask in order to increase the number of slices figures 3.12, 3.13 and 3.14
CHAPTER 3. METHODOLOGY 25

Figure 3.12: Slice before & after rotation by 90◦

Figure 3.13: Liver mask before & after rotation by 90◦

Figure 3.14: Tumor mask before & after rotation by 90◦

CHAPTER 3. METHODOLOGY 26

3.4 Liver Segmentation

We use a CNN (ResUNet) for the sole purpose of Liver Segmentation. It is trained on
CT scans and Liver masks to be able to detect the Region Of Interest (ROI) in order
to segment the liver and mask the neighbouring organs as we have no interest in them.
Examples in figures 3.15 and 3.16

Figure 3.15: Liver Segmentation example

Figure 3.16: Liver Segmentation without liver

3.5 Tumor Segmentation

At first we used a U-Net to segment the tumors in the liver but got bad results so we
tried using the ResUNet to segment the tumors. It is trained on CT scans of the Liver
CHAPTER 3. METHODOLOGY 27

after extracting the ROI from the first CNN along with the tumors’ masks. Examples in
figures 3.17 and 3.18

Figure 3.17: Tumor Segmentation example

Figure 3.18: Liver has no tumor

3.6 U-Net

3.6.1 Overview

U-Net is a Convolutional Neural Network (CNN) developed by the Computer Science

Department at the University of Freiburg, Germany specifically for the segmentation of
bio-medical images[16].
CHAPTER 3. METHODOLOGY 28

3.6.2 U-Net Architecture

The network architecture is shown in figure 3.19. It consists of a contracting path and
an expansive path and it consists of a total of 23 convolution layers[16].
Every step in the Contracting Path shares a typical architecture of CNNs which
consists of
• Repeated two 3x3 convolutions
• A Rectified Linear Unit (ReLU) refer to following every convolution. See 2.2.5 for
more details on ReLU
• 2x2 max pooling operation with a stride of 2. See 2.2.7 for more details on ReLU
• Downsampling where the number of feature channels is doubled
Every Step in the Expansive Path consists of
• Upsampling of the feature map
• 2x2 convolution which decrease the size of feature maps by half
• Concatenation with the cropped feature map
• Two 3x3 convolutions
• ReLU activation function

Figure 3.19: Architecture of the U-Net CNN[16]

CHAPTER 3. METHODOLOGY 29

3.6.3 U-Net Limitations

Although the U-Net produced really good results on the Liver Segmentation figures 3.20
and 3.21, it produced negative results on the Tumor Segmentation where we got empty
resulting mask. After searching and trying we found the Residual U-Net (ResUNet)
model which is a hybrid between the U-Net and ResNet models.

Figure 3.20: Random sample #1 that shows the segmentation result of the U-Net model.
On the left is the True input value and on the right there is the predicted label from the
model

Figure 3.21: Random sample #2 that shows the segmentation result of a complex liver
shape using the U-Net model. On the left is the True input value and on the right there
is the predicted label from the model
CHAPTER 3. METHODOLOGY 30

3.7 ResUNet

3.7.1 Overview

ResUNet is a hybrid between the traditional U-Net and the ResNet models, where it
replaces convolution blocks with residual blocks giving us the advantages of both models.
The residual unit will ease the training of the CNN and the skip connections between low
and high levels of the network and inside the residual unit eliminates degradation during
information propagation which leads to fewer training parameters,

3.7.2 Residual Units

Every Residual Block consists of

• Two 3x3 convolution blocks which consists of

– Batch Normalization Layer. See 2.2.8 for more details on ReLU

– ReLU activation layer
– Convolutional Layer

• Identity mapping

3.7.3 ResUNet Architecture

The ResUNet is constructed from three paths Encoding which encodes the input to a
compact representation, Decoding which is the opposite of Encoding and categorizes
the representation in a pixel-wise fashion and the Bridge which connects the two paths
together. Each path is built with Residual Units mentioned in 3.7.2 [23].
CHAPTER 3. METHODOLOGY 31

Figure 3.22: ResUNet Architecture

Chapter 4

Results

4.1 Experiment setting and Dataset

4.1.1 Tools used
We used Python and many tools in our project to be able to segment Liver Tumors, such
as:
• Anaconda, a distribution of the Python programming language which eases package
management and deployment.
• Numpy, a fundamental package for scientific computation in python, it adds support
for multidimensional lists and matrices and mathematical functions to work on these
matrices.
• Matplotlib, a plotting library for python which helps visualizing data and plotting
graphs.
• TensorFlow, an open source library developed by Google Brain Team for machine
learning and deep neural networks research.
• Keras, a High-Level API used in building and training Deep Learning Models. It
simplifies the process for non-experts in Machine Learning and Deep Learning and
run on multiple backend engines such as TensorFlow, CNTK and Theano without
the need for learning the syntax and how the engines works.

4.1.2 Dataset
We used the 3D-IRCADb01 dataset offered publicly by the IRCAD Research Institute
against digestive cancer, which is composed of 3D CT-scans of 10 women and 10 men
with Liver tumours in 75% of the cases. The 3D CT-scans are in the DICOM format and
are split into 2D-slices with a total of around 2,800 slices each slice has a mask for the
Liver, Tumors, Bones, Arteries, Kidneys and Lungs.

32
CHAPTER 4. RESULTS 33

4.2 Experiments and Results

4.2.1 Evaluation Metrics

We use some metrics to evaluate the performance of our models like:

• Accuracy: accuracy calculate how many pixels are classified correctly without any
regard to the class, which is not a very good metric as the true negatives value will
always be dominating by far and will always get 95%+

• Confusion Matrix: it calculates true positives, true negatives, false positives and
false negatives then draws a heat-map of the results

• Dice Coefficient: it calculates the overlap between the classes of the input A and
predicted label B and is calculated as follows

2 × |(A ∩ B)|
Dice[24] =
|A| + |B|

• True Value Accuracy: it calculates the accuracy of segmentation for tumors and
the liver and is calculated as follows
T rueP ositives
T rueV alueAccuracy = × 100
T rueP ositives + F alseN egatives

4.3 Liver Segmentation Results

4.3.1 Liver Segmentation Model Evaluation

After using the ResUNet model to segment the Liver we got the following results figure
4.1 and table 4.1
CHAPTER 4. RESULTS 34

Dice Coef, Val Dice Coef, Loss and Val

Dice Coef Validation Dice Coef Loss Validation Loss

0.9792 0.9549 0.9737 0.9752 0.9776

0.9398 0.9168 0.9332
0.8931 0.9053
0.75
0.843

0.5

0.25
0.1306
0.0991 0.0768 0.0662 0.0538 0.0428 0.0359 0.0313 0.0277 0.0246 0.0242

0
5 10 15 20

Epoch

Figure 4.1: Progress of training the ResUNet model for Liver segmentation on our dataset
for 20 epochs

Tumor training progress

Epoch Dice Coef Valid Dice Coef
1 0.8694 0.8931
2 0.9009 0.843
4 0.9232 0.9053
6 0.9338 0.0208
8 0.9462 0.9398
10 0.9572 0.9168
12 0.9641 0.9332
14 0.9687 0.9549
16 0.9723 0.9737
18 0.9754 0.9752
20 0.9758 0.9776

Table 4.1: Dice coefficient results for training and validation data on the ResUNet model
for Liver Segmentation

We took a random sample from the validation data and evaluated it on the model
CHAPTER 4. RESULTS 35

and here are the results

One of the random samples (fig: 4.2) achieved an Accuracy of 99.4% and a True Value
Accuracy of 98.1%. We can see the predicted liver in figure 4.3 and the Confusion Matrix
in figure 4.4

Figure 4.2: CT slice of the first random sample before liver segmentation

Figure 4.3: Result of liver segmentation on the first random slice in figure 4.2. True label
on the left and predicted on the right, white is pixels with liver and black is pixels with
no liver
CHAPTER 4. RESULTS 36

Figure 4.4: Confusion Matrix from the predicted value of the first slice in figure 4.2 after
liver segmentation
CHAPTER 4. RESULTS 37

Another random sample (fig: 4.5) from the liver validation data achieved an Accuracy
of 99.6% and a True Value Accuracy of 97.6%. We can see the predicted liver in figure
4.6 and the Confusion Matrix in figure 4.7

Figure 4.5: Another random sample of a CT slice before liver segmentation, as can easily
be observed the CT slice is augmented by rotation along with its mask

Figure 4.6: Result of liver segmentation on the second slice in figure 4.5. True label on
the left and predicted on the right, white is pixels with liver and black is pixels with no
liver
CHAPTER 4. RESULTS 38

Figure 4.7: Confusion Matrix from the predicted value of the second slice in figure 4.5
after liver segmentation
CHAPTER 4. RESULTS 39

4.4 Tumor Segmentation Results

4.4.1 Tumor Segmentation Model Evaluation

After using the ResUNet model to segment the Tumors we got the following results figure
4.8 and table 4.2

Dice Coef Val Dice Coef Loss Val Loss

0.8784 0.8838
0.8607 0.8683
0.8437 0.8523
0.8131 0.8268
0.788 0.7897
0.7619 0.8829 0.8815
0.85 0.8523
0.8191 0.8315
0.75
0.7885 0.7745 0.7751
0.6377
0.7365

0.5

0.2635
0.3623
0.2115 0.2255 0.2249
0.25 0.1809 0.1685
0.15 0.1477
0.2381 0.1171 0.1185
0.212 0.2103
0.1869 0.1732
0.1563 0.1477 0.1393 0.1317 0.1216 0.1162
0
10 20 30 40 50

Epochs

Figure 4.8: Progress of training the ResUNet model on our dataset for 50 epochs
CHAPTER 4. RESULTS 40

Tumor training progress

Epoch Dice Coef Valid Dice Coef
1 0.6377 0.212
5 0.7619 0.7885
10 0.7897 0.7365
15 0.8131 0.8191
20 0.8268 0.83151
25 0.8437 0.85
30 0.8523 0.8523
35 0.8607 0.7745
40 0.8683 0.7751
45 0.8784 0.8829
50 0.8838 0.8815

Table 4.2: Dice coefficient results for training and validation data on the ResUNet model
for Tumor Segmentation

We took a random sample from the validation data and evaluated it on the model
and here are the results
One of the random samples (fig: 4.9) achieved an Accuracy of 99.8%, a True Value
Accuracy of 93.8% and a Dice Coefficient of 95.2%. We can see the predicted tumor in
figure 4.10 and the Confusion Matrix in figure 4.11
CHAPTER 4. RESULTS 41

Figure 4.9: CT slice of the first random sample with the liver mask applied upon it before
tumor segmentation

Figure 4.10: Result of tumor segmentation for the first random slice in figure 4.9. True
label on the left and predicted on the right, white is pixels with tumor and black is pixels
with no tumor
CHAPTER 4. RESULTS 42

Figure 4.11: Confusion Matrix from the predicted value of the first slice in figure 4.9 after
tumor segmentation

Another random sample (fig: 4.12) achieved an Accuracy of 99.4%, a True Value
Accuracy of 94% and a Dice Coefficient of 92%. We can see the predicted tumor in figure
4.13 and the Confusion Matrix in figure 4.14
CHAPTER 4. RESULTS 43

Figure 4.12: CT slice of the second random slice with the liver mask applied upon it

Figure 4.13: Result of tumor segmentation on the second random slice in figure 4.12.
True label on the left and predicted on the right, white is pixels with tumor and black is
pixels with no tumor
CHAPTER 4. RESULTS 44

Figure 4.14: Confusion Matrix from the predicted value of the second slice in figure 4.12
after tumor segmentation
CHAPTER 4. RESULTS 45

4.4.2 ResUNet Model Limitations

The ResUNet showed some limitations as it failed to diagnose very small tumors.
One of the failing samples (fig: 4.15) achieved an Accuracy of 99.95% and a True
Value Accuracy of 0%. We can see the failure in figure 4.16 and the Confusion Matrix in
figure 4.17

Figure 4.15: A sample CT slice with the liver mask applied upon it that contains a very
small tumor

Figure 4.16: Result of tumor segmentation on slice in figure 4.15. True label on the left
and predicted on the right, white is pixels with tumor and black is pixels with no tumor.
It is obvious that the tumor was not predicted by the model at all
and also some tumors have less true value accuracy.
CHAPTER 4. RESULTS 46

Figure 4.17: Confusion Matrix from the predicted value of slice in figure 4.15 showing
that all the pixels containing tumors were misclassified
Chapter 5

Conclusion and Future Work

5.1 Conclusion
In this work we used the ResUNet model for pixel-wise automatic liver and liver tumors
segmentation using CT scans. As we can observe from the results in section 4.4, CNNs
helped us achieve our goals and are one of the best techniques that can be used in the
segmentation of liver tumors and should be tried with different tumors than liver tumors,
also the ResUNet showed very promising results.

5.2 Future Work

Although the ResUNet showed very promising results there are some limitations as men-
tioned in subsection 4.4.2. We may be able to overcome such limitations by training for
more epochs, using more data, using different datasets or using different preprocessing
techniques.

47
Appendix

48
Appendix A

Dataset and Code

The Link for the dataset we used in this work 3D-IRCADb01 can be found at: https:
//www.ircad.fr/research/3d-ircadb-01/
Our Code is on this public Repo on Github: https://github.com/AmirZaghloul/
Liver-Tumor-Segmentation-using-CNNs

49
Bibliography

[1] Avi Ben-Cohen, Idit Diamant, Eyal Klang, Michal Amitai, and Hayit Greenspan.
Fully convolutional network for liver segmentation and lesions detection. In Gustavo
Carneiro, Diana Mateus, Loı̈c Peter, Andrew Bradley, João Manuel R. S. Tavares,
Vasileios Belagiannis, João Paulo Papa, Jacinto C. Nascimento, Marco Loog, Zhi
Lu, Jaime S. Cardoso, and Julien Cornebise, editors, Deep Learning and Data La-
beling for Medical Applications, pages 77–85, Cham, 2016. Springer International
Publishing.

[2] Grzegorz Chlebus, Andrea Schenk, and Jan Hendrik Moltz. Deep learning based
automatic liver tumor segmentation in ct with shape-based post-processing. 2018.

[3] Patrick Ferdinand Christ, Florian Ettlinger, Felix Grün, Mohamed Ezzeldin A.
Elshaer, Jana Lipková, Sebastian Schlecht, Freba Ahmaddy, Sunil Tatavarty, Marc
Bickel, Patrick Bilic, Markus Rempfler, Felix Hofmann, Melvin D’Anastasi, Seyed-
Ahmad Ahmadi, Georgios A Kaissis, Julian Walter Holch, Wieland H. Sommer,
Rickmer F Braren, Volker Heinemann, and Bjoern H. Menze. Automatic liver and
tumor segmentation of ct and mri volumes using cascaded fully convolutional neural
networks. CoRR, abs/1702.05970, 2017.

[4] Mohammed Goryawala, Magno Guillen, Mercedes Cabrerizo, Armando Barreto, Seza
Gulec, Tushar Barot, Rekhaben Suthar, Ruchir N Bhatt, Anthony Mcgoron, and
Malek Adjouadi. A 3-d liver segmentation method with parallel computing for se-
lective internal radiation therapy. IEEE transactions on information technology in
biomedicine : a publication of the IEEE Engineering in Medicine and Biology Soci-
ety, 16:62–9, 01 2012.

[5] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning
for image recognition. CoRR, abs/1512.03385, 2015.

[6] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network
training by reducing internal covariate shift. CoRR, abs/1502.03167, 2015.

[7] Ir J De Backer, Ir W Vos, Olivier Vanderveken, Dr A Devolder, Marc Braem, Dirk

van dyck, and Wilfried Backer. Combining mimics and computational fluid dynamics
(cfd) to assess the efficiency of a mandibular advancement device (mad) to treat
obstructive sleep apnea (osa). 05 2019.

50
BIBLIOGRAPHY 51

[8] Bray F. Center M.M. Ferlay J. Ward E. Jemal, A. and Forman. Global cancer
statistics. CA: A Cancer Journal for Clinicians, 61:60–90, 2017.

[9] M.H. Lev and R.G. Gonzalez. 17 - ct angiography and ct perfusion imaging. In
Arthur W. Toga and John C. Mazziotta, editors, Brain Mapping: The Methods (Sec-
ond Edition), pages 427 – 484. Academic Press, San Diego, second edition edition,
2002.

[10] Wen Jung Li, Fucang Jia, and Qingmao Hu. Automatic segmentation of liver tumor
in ct images with deep convolutional neural networks. 2015.

[11] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully con-
volutional neural networks for volumetric medical image segmentation. CoRR,
abs/1606.04797, 2016.

[12] National Institute of Biomedical Imaging and Bioengineering (NIBIB). Com-

puted tomography (ct). https://www.nibib.nih.gov/science-education/
science-topics/computed-tomography-ct, 2019. Last accessed 18 May 2019.

[13] Stephen M. Pizer, E. Philip Amburn, John D. Austin, Robert Cromartie, Ari
Geselowitz, Trey Greer, Bart ter Haar Romeny, John B. Zimmerman, and Karel
Zuiderveld. Adaptive histogram equalization and its variations. Computer Vision,
Graphics, and Image Processing, 39(3):355 – 368, 1987.

[14] Stephen Power, Fiachra Moloney, Maria Twomey, Karl James, Owen O’Connor,
and Michael Maher. Computed tomography and patient risk: Facts, perceptions
and uncertainties. World Journal of Radiology, 8:902, 12 2016.

[15] Sherif R Z Abdel-Misih and Mark Bloomston. Liver anatomy. The Surgical clinics
of North America, 90:643–53, 08 2010.

[16] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional net-
works for biomedical image segmentation. CoRR, abs/1505.04597, 2015.

[17] Anish singh Walia. Activation functions and its types-

which is better? https://towardsdatascience.com/
activation-functions-and-its-types-which-is-better-a9a5310cc8f, 2017.
Last accessed 14 May 2019.

[18] Rupesh Kumar Srivastava, Klaus Greff, and Jürgen Schmidhuber. Highway net-
works. CoRR, abs/1505.00387, 2015.

[19] Christos Stergiou and Dimitrios Siganos. Neural networks. SURPRISE 96, 4, 1996.

[20] Cancer Research UK. Liver cancer stages, cancer research uk. https://www.
cancerresearchuk.org/about-cancer/liver-cancer/stages, 2019. Last ac-
cessed 17 May 2019.
BIBLIOGRAPHY 52

[21] Stanford University. Convolutional neural networks (cnns / convnets). http://

cs231n.github.io/convolutional-networks, 2019. Last accessed 15 May 2019.

[22] Chi-Feng Wang. The vanishing gradient problem the problem, its causes,
its significance, and its solutions. https://towardsdatascience.com/
the-vanishing-gradient-problem-69bf08b15484, 2019. Last accessed 15
May 2019.

[23] Zhengxin Zhang, Qingjie Liu, and Yunhong Wang. Road extraction by deep residual
u-net. CoRR, abs/1711.10684, 2017.

[24] Kelly Zou, Simon Warfield, Aditya Bharatha, Clare Tempany, Michael Kaus, Steven
J Haker, William Wells, Ferenc Jolesz, and Ron Kikinis. Statistical validation of
image segmentation quality based on a spatial overlap index. Academic radiology,
11:178–89, 02 2004.