Intelligent Detection and Classification of Lung Nodules

i

Intelligent Detection and Classification of Lung

Nodules from CT Images

By

Syed Muhammad Naqi

CIIT/FA14-PCS-001/WAH

PhD Thesis

In

Computer Science

COMSATS University Islamabad

Wah Campus, Pakistan

Spring, 2019

ii

COMSATS University Islamabad



A Thesis Presented to

COMSATS University Islamabad, Wah Campus

In partial fulfillment

of the requirements for the degree of

PhD (Computer Science)

By

Syed Muhammad Naqi


Spring, 2019

iii



A Post Graduate Thesis submitted to the Department of Computer Science as

partial fulfillment of the requirements for the award of Degree of Ph.D. in

Computer Science.

Name Registration No.

Syed Muhammad Naqi CIIT/FA14-PCS-001/WAH

Supervisor

Dr. Muhammad Sharif

Associate Professor

Department of Computer Science

COMSATS University Islamabad (CUI)

Wah Campus, Pakistan

viii

DEDICATION

To My Parents, Wife and Children

ix

ACKNOWLEDGEMENTS

I have completed this research thesis with the help of Almighty Allah who always

helped me whenever I was in trouble. I would like to acknowledge my advisor Dr.

Muhammad Sharif for his kind guidance and supervision that enabled me to complete

this research work. His continuous motivation, research ideas, and wonderful

knowledge helped me in a great deal to produce quality research.

I am thankful to the Head of the Computer Science Department, CUI Wah, Dr.

Muhammad Wasif Nisar who always encouraged me during the whole period of my

Ph.D. studies. I am also grateful to supervisory committee members and faculty

members of the CS department for their guidance. I am also thankful to the Higher

Education Commission (HEC), Pakistan for financial support under the Indigenous

PhD scholarship program.

I would like to extend my gratitude to the faculty of Quaid-i-Azam University. I wish

to extend my thanks to Dr. Khalid Saleem, Dr. Muhammad Tauqeer, Dr. Muddassar

Azam Sindhu, Dr. Muhammad Usman, Dr. Ayaz Hussain and Dr. Arfan Jaffar for

their valuable comments and guidance on my research.

I am especially indebted to Dr. Zahid Noman additional medical superintendent, Dr.

Naushaba Malik senior radiologist, and Dr. Maryam Rauf radiologist, at Social

Security Hospital, Islamabad; for their assistance and feedback.

Finally, a special word of thanks to my father, whose prayers not only enabled me to

complete this thesis but have been a source of illumination all my life. I want to pay

my superior appreciations to my wife, my lovely kids, and siblings for their special

prayers encouragements and patience. Thanks to every person who helped me directly

or indirectly during my PhD.

Syed Muhammad Naqi


x

ABSTRACT

Intelligent Detection and Classification of Lung Nodules from

CT Images

Lung cancer is considered as the most probable cancer in human beings. Moreover,

the survival rate of its patients is also very low. However, its early detection increases

the chances of survival of its patients. Medical imaging helps in early diagnosis, and

the most effective method is Computed Tomography (CT). The initial appearance of

lung cancer is the presence of a nodule in the lungs. On a CT scan, a lung nodule

appears as a round object, either independent (isolated), pleura-attached (juxta-

pleural) or vessel-attached (juxta-vascular). Computer-based systems help the

radiologists in the diagnosis process by providing the second opinion. However, there

is a significant challenge of sensitivity and the number of false positives in the

existing techniques.

The main objective of this research is to develop algorithms/methods which detect

and classify lung nodules precisely. Four methods are proposed, involving lung

volume extraction, nodule candidate detection, feature extraction, and classification.

Lung volume extraction is computed by using the optimal threshold. Fractional order

Darwinian particle swarm optimization is applied for threshold selection, which

significantly improves the segmentation of lung volume. Nodule candidate detection

is performed by applying 3D image analysis. Precise boundary correction and 3D

segmentation are proposed for the detection of juxta-pleural and juxta-vascular

nodules, respectively. Nodule candidates are refined using their shape properties. A

set of 2D, as well as 3D features, including geometric, texture, and gradient features,

are extracted with variations. Hybrid feature vectors are created by the combination

and selection of relevant features to improve classification. The initial classification is

performed by k-NN, SVM, Naïve Bayes and AdaBoost classifiers. In addition, an

SVM-ensemble classifier is implemented. Furthermore, a stacked autoencoder and

softmax is applied to reduce the false positives.

The proposed methods are evaluated over publicly available LIDC dataset, and the

largest benchmark dataset LIDC-IDRI. The first method achieved an accuracy of

xi

98.8%, 97.8% sensitivity and 3.7 False Positives (FP) per scan. The second method

achieved an accuracy of 99.2% with 98.3% sensitivity, 98.0% specificity and 3.3

FPs/scan. The third method achieved 99.0% accuracy, 98.6% sensitivity, 98.2%

specificity with 3.4 FPs/scan. The fourth method achieved 96.9% accuracy, 95.6%

sensitivity, 97.0% specificity and very low false positives of 2.8/scan. The results are

compared with the existing methods over standard parameters, including sensitivity,

specificity, accuracy and FPs/scan. The comparison is also extended to other

parameters, including the number of scans, number of nodules and size of nodules

used for experimentation. The comparison points to the significance of this research.

The methods presented in this thesis will be useful for the radiologists in automated

diagnosis of lung nodules at the earlier stage.

xii

TABLE OF CONTENTS

Introduction ................................................................................. 1 Chapter 1

1.1 Introduction ........................................................................................ 2

1.2 Lung Cancer ....................................................................................... 2

1.3 Lung Nodule ....................................................................................... 5

1.4 Computed Tomography ..................................................................... 5

1.5 CT Datasets ........................................................................................ 7

1.5.1 Lung Image Database Consortium (LIDC) ........................... 7

1.5.2 Lung Image Database Consortium-Image Database Resource

Initiative (LIDC-IDRI) .......................................................... 8

1.6 Evaluation Parameters ........................................................................ 8

1.7 Problem Statement ............................................................................. 9

1.8 Motivation and Objectives ................................................................. 9

1.9 Research Challenges ........................................................................ 10

1.10 Contributions .................................................................................... 10

1.11 Thesis Organization ......................................................................... 12

Literature Review ...................................................................... 14 Chapter 2

2.1 Introduction ...................................................................................... 15

2.1.1 Lung Region Segmentation ................................................. 16

2.1.2 Nodule Candidate Detection ............................................... 16

2.1.3 Feature Extraction ............................................................... 17

2.1.4 Classification ....................................................................... 17

2.2 Recent Developments in Lung Nodule Diagnosis ........................... 17

2.2.1 Supervised Learning Methods ............................................. 22

xiii

2.2.2 Unsupervised Learning Methods ......................................... 25

2.2.3 Diagnosis Based on Template Matching ............................. 26

2.2.4 Diagnosis Based on Fuzzy Logic ........................................ 27

2.2.5 Diagnosis Based on Thresholding ....................................... 27

2.2.6 Diagnosis Based on Mathematical Morphology ................. 28

2.2.7 Multiple Techniques/Hybrid Approaches ........................... 29

2.2.8 Deep Learning Based Methods ........................................... 31

2.3 Discussion and Analysis .................................................................. 32

2.4 Summary .......................................................................................... 35

Proposed Methods ..................................................................... 37 Chapter 3

3.1 Introduction ...................................................................................... 38

3.2 Method I: Lung Nodule Detection using Polygon Approximation

and Hybrid Features from Lung CT Images .................................... 40

3.2.1 Lung Region Extraction using Optimal Threshold ............. 41

3.2.2 Nodule Candidate Selection using Polygon Approximation

............................................................................................. 43

3.2.3 Nodule Candidate Enhancement ......................................... 44

3.2.4 False Positive Reduction using Nodule Properties.............. 44

3.2.5 Hybrid Intensity-Geometric and Gradient Feature Extraction

............................................................................................. 45

3.2.6 Support Vector Machine for Classification ......................... 48

3.3 Method II: 3D Nodule Candidate Detection Supported by Hybrid

Features to Reduce False Positives .................................................. 48

3.3.1 Lung Volume Extraction using Optimal Threshold ............ 50

3.3.2 Nodule Candidate Detection ............................................... 52

xiv


3.3.4 Classification ....................................................................... 60

3.4 Method III: Multistage Segmentation Model and SVM-Ensemble for

Precise Lung Nodule Detection ....................................................... 61

3.4.1 Nodule Segmentation .......................................................... 62

3.4.2 Geometric Texture Feature Descriptor (GTFD) .................. 68

3.4.3 Ensemble Classification ...................................................... 69

3.5 Method IV: Lung Nodule Detection and Classification based on

Geometric Fit in Parametric Form and Deep Learning ................... 70

3.5.1 Lung Volume Segmentation ................................................ 72

3.5.2 Nodule Candidate Detection using Geometric Fit in

Parametric Form .................................................................. 79


3.5.4 Stacked Sparse Autoencoder and Softmax Based Feature

Optimization and Classification .......................................... 85

3.6 Summary .......................................................................................... 88

Results and Discussion .............................................................. 89 Chapter 4

4.1 Introduction ...................................................................................... 90


and Hybrid Features from Lung CT Images .................................... 91

4.2.1 Experimentation and Results ............................................... 91

4.2.2 Comparison of Proposed Technique with Existing

Techniques ........................................................................... 93

4.3 Method II: 3D Nodule Candidate Detection Supported by Hybrid

Features to Reduce False Positives .................................................. 96

4.3.1 Dataset and Evaluation Parameters ..................................... 96

xv

4.3.2 Experimentation .................................................................. 97

4.3.3 Comparison of Proposed Method with Existing Techniques

........................................................................................... 104

4.4 Method III: Multistage Segmentation Model and SVM-Ensemble for

Precise Lung Nodule Detection ..................................................... 106

4.4.1 Experimentation and Results ............................................. 107

4.4.2 Comparative Analysis ....................................................... 111


Geometric Fit in Parametric Form and Deep Learning ................. 113

4.5.1 Dataset Selection ............................................................... 114

4.5.2 Segmentation Results ........................................................ 114

4.5.3 Initial Candidate Detection Results ................................... 119

4.5.4 Classification Results ........................................................ 119

4.5.5 Comparative Analysis ....................................................... 123

4.6 Summary ........................................................................................ 125

Conclusion and Future Work ................................................ 126 Chapter 5

5.1 Conclusion ...................................................................................... 127

5.2 Future Work ................................................................................... 130

References ................................................................................................... 131

List of Publications ........................................................................................ 155

Appendix: WHO Cancer Statistics .............................................................. 156

xvi

LIST OF TABLES

Table 2.1: Categorization of lung nodule detection methods ...................................... 18

Table 2.2: Limitations in recently developed methods ................................................ 33

Table 3.1: Threshold values for nodule candidate set refinement ............................... 81

Table 4.1: Results produced by SVM linear kernel ..................................................... 91

Table 4.2: Results produced by SVM polynomial kernel ............................................ 92

Table 4.3: Results produced by SVM-RBF ................................................................. 92

Table 4.4: Best results by SVM classifier with different kernels ................................ 92

Table 4.5: Comparison of the proposed method with the results of existing techniques

...................................................................................................................................... 94

Table 4.6: Results with the geometric feature vector .................................................. 97

Table 4.7: Results with the texture feature vector ....................................................... 98

Table 4.8: Results with HOG-PCA feature vector....................................................... 98

Table 4.9: Results with texture + geometric feature vector ......................................... 99

Table 4.10: Results with texture + HOG-PCA feature vector ..................................... 99

Table 4.11: Results with geometric + HOG-PCA feature vector .............................. 100

Table 4.12: Results with geometric + texture + HOG-PCA feature vector ............... 100

Table 4.13: Results comparison of the proposed method with existing techniques .. 105

Table 4.14: Results with geometric features .............................................................. 107

Table 4.15: Results with texture features ................................................................... 108

Table 4.16: Results with GTFD ................................................................................. 108

Table 4.17: Comparison of datasets used in existing and proposed systems............. 111

Table 4.18: Comparison of results with existing methods ......................................... 113

xvii

Table 4.19: FODPSO initialization parameters ......................................................... 114

Table 4.20: Comparison of segmentation results ....................................................... 116

Table 4.21: Results with different variations of deep net .......................................... 121

Table 4.22: Comparison of classification results ....................................................... 121

Table 4.23: Comparative analysis in terms of dataset and results ............................. 124

xviii

LIST OF FIGURES

Fig 1.1: WHO worldwide cancer statistics 2018 [11] (a) number of new cases (b)

number of deaths ............................................................................................................ 3

Fig 1.2: WHO cancer statistics of Pakistan [12] (a) 2018 new cases (b) number of

deaths ............................................................................................................................. 4

Fig 1.3: Normal vs. nodule cells [14] (a) Normal cells (b) Cells forming a nodule ...... 5

Fig 1.4: Lung CT images with nodules (a) isolated (b) juxta-pleural (c) juxta-vascular

........................................................................................................................................ 6

Fig 1.5: Block diagram of thesis organization ............................................................. 13

Fig 2.1: Basic steps of a diagnostic system.................................................................. 15

Fig 2.2: Labeled CT slice ............................................................................................. 16

Fig 3.1: Block diagram of proposed methods .............................................................. 39

Fig 3.2: Major steps of the proposed method .............................................................. 40

Fig 3.3: Lung region extraction process ...................................................................... 42

Fig 3.4: Gray level distribution (a) input CT images (b) histogram of images (c)

threshold images .......................................................................................................... 42

Fig 3.5: Nodule candidate selection (a) objects in ROI (b) marked by the polygon ... 43

Fig 3.6: ROI and nodule segmentation (a) input CT image (b) lung mask (c) ROI with

vessels and nodules (d) nodule candidate region ......................................................... 45

Fig 3.7: Gradient orientation (a) nodule candidates (b) HOG descriptor .................... 47

Fig 3.8: Major steps of the proposed method .............................................................. 50

Fig 3.9: Lung region extraction (a) input CT images (b) histogram of images (c) final

lung masks of the images ............................................................................................. 51

Fig 3.10: Nodule candidate detection process ............................................................. 52

xix

Fig 3.11: Nodule candidate detection (a) segmented ROIs, (b) corresponding contours

of the objects within the ROI‘s, (c) the final selected candidates. ............................... 55

Fig 3.12: Feature extraction and selection ................................................................... 59

Fig 3.13: Workflow of the proposed method ............................................................... 62

Fig 3.14: Gray level distribution (a) input CT images (b) their respective histograms

...................................................................................................................................... 64

Fig 3.15: Juxta-pleural nodule extraction (a) input images with juxta-pleural nodules,

(b) corresponding lung masks missing pleural nodules, (c) boundary correction using

morphology, (d) extracted lungs with juxta-pleural nodules ....................................... 65

Fig 3.16: Lung region extraction (a) input image, (b) background removed, (c) lung

mask, (d) extracted lung ............................................................................................... 66

Fig 3.17: 3D connectivity of candidate pixels ............................................................. 67

Fig 3.18: Nodule candidate refinement (a) objects in ROI (b) nodule candidate regions

...................................................................................................................................... 68

Fig 3.19: Extracted nodule candidates ......................................................................... 68

Fig 3.20: Workflow of SVM-ensemble for nodule classification................................ 70

Fig 3.21: Block diagram of the proposed method........................................................ 72

Fig 3.22: Proposed lungs segmentation method .......................................................... 73

Fig 3.23: Lung region extraction (a) input CT images (b) threshold images (c)

extracted lung regions. ................................................................................................. 77

Fig 3.24: Juxta-pleural nodule extraction (a) Images with pleura nodule, (b) initial

lung mask, (c) after boundary correction, (d) final ROI .............................................. 78

Fig 3.25: Nodule candidate refinement (a) segmented ROIs, (b) the final candidate

regions .......................................................................................................................... 82

Fig 3.26: Nodule candidates (a) nodules (b) non-nodules ........................................... 82

xx

Fig 3.27: Texture special information (a) pixels with neighborhood directions (b) 3D

neighborhoods .............................................................................................................. 83

Fig 3.28: Feature fusion for the hybrid feature vector ................................................. 85

Fig 3.29: Stacked autoencoder for feature optimization .............................................. 87

Fig 4.1: Experimentation statistics of all proposed methods ....................................... 90

Fig 4.2: Comparison charts of classification results .................................................... 93

Fig 4.3: Comparison chart of proposed and existing system results............................ 95

Fig 4.4: ROC analysis (a) AdaBoost results with different number of iterations over

V7 (b) SVM results over hybrid feature vector V7 ................................................... 102

Fig 4.5: ROC analysis (a) results using feature vector V7 (b) results using feature

vector V5 .................................................................................................................... 103

Fig 4.6: ROC curve with GTFD feature vector ......................................................... 109

Fig 4.7: FROC analysis (a) FROC curve shows false positive reduction by each

classifier, (b) FROC curve illustrates the sensitivity comparison of each classifier at

3.4 FPs/scan ............................................................................................................... 110

Fig 4.8: 3D lung volume (a) left view and (b) right view .......................................... 117

Fig 4.9: Extracted 3D ROI with nodules and vessels (a) normal view of the complete

lung (b) a portion containing juxta-vascular nodule ................................................. 118

Fig 4.10: The first layer of autoencoder..................................................................... 120

Fig 4.11: Stacked autoencoder and softmax for feature reduction and classification 120

Fig 4.12: FROC analysis (a) FROC curve showing the performance of each classifier,

(b) FROC curve comparing the performance of classifiers at 2.8 FPs/scan .............. 122

xxi

LIST OF ABBREVIATIONS

ACM Active Contour Models

AHE Adaptive Histogram Equalization

ANN Artificial Neural Network

AUC Area Under Curve

CLAHE Contrast Limited Adaptive Histogram Equalization

CNN Convolutional Neural Network

CT Computed Tomography

CV Chan–Vese

DCT Discrete Cosine Transformation

DPSO Darwinian Particle Swarm Optimization

DTCNN Discrete Time Cellular Neural Network

FC Fuzzy Connectedness

FCM Fuzzy C-Means

FLD Fisher Linear Discriminant

FN False Negative

FNR False Negative Rate

FODPSO Fractional Order Darwinian Particle Swarm Optimization

FOPSO Fractional Order Particle Swarm Optimization

FP False Positive

FROC Free-Response Operator Characteristic

FPR False Positive Rate

GA Genetic Algorithm

GATM Genetic Algorithm Template Matching

G-CNN Genetic Cellular Neural Networks

GLCM Gray Level Co-occurrence Matrix

xxii

GLMR Generalized Linear Model Regression

GMM Gaussian Mixture Model

GPU Graphics Processing Unit

GTFD Geometric Texture Features Descriptor

GTSDM Gray Tone Spatial Dependence Matrices

HGTF Hybrid Geometric Texture Feature

HOG Histograms of Oriented Gradients

HU Hounsfield Unit

KNN K-Nearest Neighbor

kV Kilo Voltage

LDA Linear Discriminant Analysis

LDCT Low Dose Computed Tomography

LIDC Lung Image Database Consortium-Image

LIDC-IDRI Lung Image Database Consortium-Image Database Resource

Initiative

LoB Laplacian of Bilateral

LWTM Lung Wall Template Matching

mA.s Milli-Ampere Seconds

MRI Magnetic Resonance Imaging

MSM Mass–Spring Model

NBIA National Biomedical Imaging Archive

PCA Principle Component Analysis

PET Positron Emission Tomography

PRI Probabilistic Random Index

PSO Particle Swarm Optimization

RBF Radial Base Function

xxiii

ROC Receiver Operator Characteristic

ROI Region of Interest

SA Stacked Autoencoder

SOM Self-Organization Map

SVM Support Vector Machine

TN True Negative

TP True Positive

VoI Variation of Information

WHO World Health Organization

XML Extensible Markup Language

xxiv

LIST OF SYMBOLS

Threshold

Coordinate x-axis

Coordinate y-axis

Previous bias change

Momentum constant

Derivative of the performance in the weight

Attribute in the feature vector

Target class

Width of the polygon surrounding an object

Length of the polygon

ø Empty set

Set of objects in ROI

An object in set S

Probability of class

Probability of a gray level in gray level co-occurrence matrix

Mean

Standard deviation

Skewness

E Energy

H Entropy

xxv

Kurtosis

Maximum gray level in an image

Mean of class

Area

Diameter

Radius

Set of voxels in the segmented object

Voxel

Position of snake

Data vector of example in the data

Class label of example in data

Bias

Regularization parameter in SVM

Slack variable to control the margin in SVM

Weight Vector

Degree of SVM polynomial kernel

Gaussian of standard deviation

Ratio threshold

Lower threshold of area

Higher threshold of area

Circularity threshold

xxvi

Standard deviation with x-axis

Standard deviation with y-axis

Gradient magnitude with x-axis

Gradient magnitude with y-axis

Lowest gray level in an image

Highest gray level in an image

Geometric features

Texture features

HOG features

Serial fusion

Target class

Euclidian Distance

Segment length of lung boundary

Union

Probability distribution

Number of gray levels

Histogram of gray level g

Gray level in an image

Source image

Binary image

Gradient operator

xxvii

The scalar to control elastic energy in ACM

Trust factor

The scalar to control bending energy in ACM

The fractional coefficient in FODPSO

Learning rate

Additional learning rate

Partial derivative

Global weight controller

Local weight controller

The penalty associated with the probability of a category

Maximum threshold for the diameter of the nodule

Total iterations in FODPSO

Total swarms

Minimum swarms

Total particles in each swarm

Minimum particles in each swarm

Maximum particles in each swarm

Maximum swarms

Total pixels

1

Chapter 1

Introduction

2

1.1 Introduction

This thesis focus focuses on the detection and classification of lung nodules. Lung

cancer detection at initial stage increases the survival rate of patients. Automatic

detection of lung nodules facilitates radiologists during the diagnosis process.

However, there is a challenge of false positives in automated systems which may lead

to wrong findings. The primary focus of this study is the reduction of false positives

by preserving sensitivity.

1.2 Lung Cancer

In the 20th

century, more than 100 million deaths were reported due to tobacco use. It

is expected that the deaths might reach 1 billion in the 21st century. About one-third of

these deaths are due to lung cancer [1, 2]. Moreover, the lung cancer mortality rate in

comparison to other types of cancers is very high [3]. According to the statistics, 1.76

million deaths and 2.09 million new cases are reported in 2018, which is the highest

number in all types of cancers [4] as shown in Fig 1.1. In Pakistan, lung cancer is the

third largest cancer. According to WHO, there are 9771 new cases reported in

Pakistan during 2018, as shown in Fig 1.2. The initial appearance of lung cancer is the

presence of a nodule in the lungs. Different types of tests are usually advised by

doctors to identify lung cancer. More important of those tests is screening using

radiography, which includes Computed Tomography (CT) [5], Magnetic Resonance

Imaging (MRI) [6], and Positron Emission Tomography (PET) [7]. The CT scan is the

most common and cost-effective method to identify lung cancer [8]. Abnormalities

identified through CT scan imply the possibility of lung cancer in an individual.

During the CT scan of a patient, more than one hundred images are generated.

Radiologists have to examine all the images to identify cancer, which is very time-

consuming and error-prone. To overcome this issue, it is much needed to have some

automated or semi-automated systems which help the radiologists to distinguish

cancerous and non-cancerous regions clearly. In medical imaging, automated decision

support systems have a key role in early diagnosis [9, 10]. These systems facilitate the

radiologists by providing the second opinion. Radiologists use such decision support

systems to identify cancerous regions in the CT scan images. These systems comprise

multiple computer vision algorithms.

3

(a)

(b)

Fig 1.1: WHO worldwide cancer statistics 2018 [11] (a) number of new cases (b)

number of deaths

4

(a)

(b)

Fig 1.2: WHO cancer statistics of Pakistan [12] (a) 2018 new cases (b) number of deaths

5

1.3 Lung Nodule

Human cells grow in a well-controlled manner. In normal growth, the cells follow a

regular pattern of dividing and dying to produce a few new cells. The uncontrolled

growth of these cells leads to cancer. These abnormal cells grow very fast in number

and form a lump inside the lungs, also called lung nodule [13]. Fig 1.3 shows the

normal and abnormal cell pattern.

Fig 1.3: Normal vs. nodule cells [14] (a) Normal cells (b) Cells forming a nodule

A small size lung nodule is also called pulmonary nodule [15]. It is a round shaped

object within the lungs with a diameter less than 30mm. A nodule can be malignant or

benign. Benign nodules are non-cancerous, whereas malignant nodules are cancerous

[16]. Malignant nodules are harmful, and they grow with the passage of time.

Treatment of the patient becomes much harder if size of the nodule becomes large. In

addition, a nodule is also categorized depending upon its position within the lung. It

can be isolated [17], attached to the boundary of the lung called juxta-pleural nodule

[18] and vessel attached (juxta-vascular) nodules [19]. Different types of nodules in a

CT image are shown in Fig 1.4. Larger nodule size can disturb the normal functioning

of the lungs and can result in death in a very short span of time [20]. If the size of a

nodule becomes larger, then treatment of the patient becomes harder [21]. Therefore,

it is more desirable and highly recommended that cancer should be identified at its

initial stage.

1.4 Computed Tomography

Computed tomography (CT) was invented by a Nobel laureate Dr. G. N. Hounsfield

in 1971 [22]. CT is one of the most common procedures to achieve a non-invasive

visualization of the body's internal structure. CT images are produced by combining

(a) (b)

6

X-ray technology and a computer to help doctors to visualize the body in depth and

examine the organs. An X-ray tube slowly rotates 360° around the human body and

takes pictures. These images are called slices. If a suspicious structure is found during

a normal radiograph or ultrasound scan, the doctor may suggest a CT scan of that part

of the body. It is a method to investigate the internal structure of the human body in

detail; it enables doctors to check for any internal bleeding, mass, brain cancer, or

lung nodule. Over the last two decades, it has become a standard imaging modality.

CT produces more precise results for bone structures. However, it is also a preferred

technique for brain and lung cancer diagnosis [23]. Sometimes a CT scan performed

on the basis of other symptoms can result in detection of smaller nodules such as in

the case of known carcinoma of lung or anywhere else. The physician looking at a

metastatic disease can result in the detection of a nodule.

(a)

(b)

(c)

Fig 1.4: Lung CT images with nodules (a) isolated (b) juxta-pleural (c) juxta-vascular

7

The actual scan time for CT is usually less than for an MRI or PET scan, and it is less

sensitive to patient movement. Moreover, it is more cost effective than both [24]. It is

extremely sensitive, detecting nodules as small as 3mm [25, 26], that are invisible on

a conventional X-ray examination of the lungs. In fact, recent research on CT

screening reveals that most cases of lung cancer detected in CT are invisible on a

conventional chest X-ray carried out at the same time as the CT scan. However, the

recommendation is to use Low Dose Computed Tomography (LDCT) as even this can

identify nodules smaller than can conventional chest radiography due to

advancements in CT [27, 28].

1.5 CT Datasets

A well-characterized repository plays a vital role in the performance evaluation of a

diagnosis system. A diagnosis system produces the best performance when it is

trained over a large amount of well-documented data. Some researchers used clinical

data from different hospitals, which is a very time-consuming and intensive process.

Extra work is required to normalize these images. Therefore, it is necessary to

develop a standard repository that needs a lot of human resources and time [29]. In

recent years, various research groups have been working to resolve this issue for the

betterment of the diagnosis system performance, which can reduce the risk of false

positives [30].

1.5.1 Lung Image Database Consortium (LIDC)

LIDC is a standard dataset of lung CT images that is publicly available to the

researchers for the performance evaluation of their proposed methods. It can be

accessed through the National Cancer Institute's Imaging Archive [31-33]. This

database of thoracic CT images is developed with the cooperation of five medical

imaging research groups. The motivation behind the development of LIDC was to

develop a database of CT images as a web-accessible resource for the evaluation of

diagnosis systems [34]. Four experienced radiologists of different institutions

performed the annotations of the nodules with reference to their position and size. The

nodules in this dataset are of sizes ranging from 3-30mm [35].

8

1.5.2 Lung Image Database Consortium-Image Database Resource

Initiative (LIDC-IDRI)

LIDC-IDRI is the extended version of LIDC dataset. Initially, the LIDC consisted of

84 scans containing 58 nodules. In 2004 it was combined with Image Database

Resource Initiative (IDRI). With the passage of time size of the database has been

extended up to 1012 scans to form LIDC-IDRI. It is the largest and benchmark

database that is publicly available to support researchers in the evaluation of diagnosis

systems for pulmonary nodules. Moreover, along with the image data, another

repository of XML files is also provided for the convenience of researchers to get

annotation details about the nodules present in the dataset. In recent years, this dataset

has been considered as a benchmark dataset and is the most widely used in the

performance evaluation of diagnosis methods [36].

1.6 Evaluation Parameters

A diagnosis system classifies the images into diseased and normal categories. The

confusion matrix [37] is used to evaluate the performance of diagnosis methods,

which consists of four possible outcomes. If a positive instance of test data is

correctly classified by the diagnosis method, then it is called True Positive (TP),

otherwise False Negative (FN). Similarly, if a negative instance is classified as

negative, then it is referred to as True Negative (TN) otherwise False Positive (FP).

The specificity, sensitivity [38] and accuracy are used as performance measures [14]

of the proposed diagnosis methods. These measures are given in Eq. 1.1 to Eq. 1.3.

(1.1)

(1.2)

(1.3)

The sensitivity represents the true positive rate, and specificity shows the true

negative rate. In addition to these parameters, Receiver Operator Characteristic (ROC)

analysis is important and is a widely used measure [37, 38]. It is created by plotting

9

false positive rate versus sensitivity, and the Area Under the ROC Curve (AUC) is

computed. Its value ranges from 0 to 1, where 1 represents the ideal performance. In

addition to the ROC curve, another important measure named Free-Response

Operating Characteristic (FROC) analysis [39] is often used to validate the

performance of lung nodule diagnosis methods. It is created by plotting sensitivity

versus the average number of FPs/scan. In a single CT scan, some images have more

than one nodule, and some may not have any. Therefore, FROC measure is an

important measure that shows the relationship of sensitivity with the number of FPs

per scan. Although, the most recent studies have presented various lung nodule

detection and classification methods, but to overcome the number of false positives

remains a challenge.

1.7 Problem Statement

Existing methods lack both in achieving high accuracy and low false positive rate. It

is due to limitations in handling the complex structure of the lungs, and a large

number of CT images generated during a CT scan. Existing methods also lack in

handling different kinds of nodules, including isolated, juxta-pleural and juxta-

vascular nodules, especially of smaller sizes ranging from 3-30mm. The proposed

research aims to address these problems by providing novel methods for detection and

classification of nodules using image processing and machine learning techniques.

1.8 Motivation and Objectives

The rapid increase in lung cancer is an alarming situation. According to the statistics

of WHO, lung cancer is the leading type of cancer for the last two years. Therefore, it

is very important to diagnose lung cancer at its earliest stage. Lung nodule diagnosis

methods still have a challenge of false positives. The major focus of this research is

on lung CT image processing. The main objective is to develop

algorithms/techniques:

To automatically identify nodular regions in CT images.

To view CT images of patients and improve the performance of the diagnostic

methods in terms of sensitivity and the number of false positives.

To ensure the inclusion of juxta-pleural nodules in the ROI during the

segmentation. It improves the accuracy of the diagnosis process.

10

To separate the juxta-vascular nodules from vessels precisely for better

performance of the diagnosis method.

To segregate actual nodules from potential nodules by using relevant feature

extraction and robust classification.

To facilitate the radiologists during the diagnosis process by providing the

second opinion.

1.9 Research Challenges

As discussed earlier, the existing diagnosis methods lack both in high accuracy and

low false positive rate. Major contributory factors are limitations in the handling of

different types of nodules. Major challenges in automated diagnosis of lung nodules

are listed below:

Precise segmentation of juxta-pleural nodules is a challenging problem.

During the diagnosis process, a fundamental step is the extraction of the lung

portion. During this process, juxta-pleural nodules are fully or partially

removed from ROI due to their presence on the boundary of lung.

The separation of juxta-vascular nodules from vessels involves another

challenge. Current diagnosis methods usually miss juxta-vascular nodules

during vessel removal, and this results in a wrong diagnosis.

Detection of small nodules is a challenge as it contributes to false positive rate.

Instead of using 2D slices, a 3D approach can produce better results. However,

dealing with all slices of a scan during the process of diagnosis is another

challenge.

A major challenge for this study is to diagnose abnormal areas with fewer

false positives. To do so, there is a need to classify nodules and non-nodules

with greater classification accuracy using novel segmentation and

classification techniques.

1.10 Contributions

The major contributions of this research, which added to the existing body of

knowledge in this area are as follows:

11

Nodule candidate detection is a critical step that impacts the performance of a

diagnosis method. Lung nodules are of three types with reference to their

presence in the lung area, including juxta-pleural, juxta-vascular and isolated

nodules. These nodules cannot be detected with a single straight forward

method. To address this challenge, a multistage segmentation method is

proposed in this thesis to extract the lung volume and nodule candidates

precisely, which was lacking in the existing systems. It improved the detection

of actual nodules and reduced the inclusion of false positives in the initial

candidate set. The nodule candidate set is further refined by applying

geometric fit in parametric form

The segmentation of juxta-vascular nodules impacts the performance of a

diagnosis method. It is challenging to separate these nodules due to their

attachment to the complex structure of vessels. In this thesis, a novel 3D

segmentation of juxta-vascular nodules from vessels is proposed. Existing

systems perform slice-level segmentation and then reconstruct the 3D nodule

structure by analyzing the information. In contrast, this thesis proposes a scan

level 3D segmentation method, which reduces the number of false positives

and improves overall performance.

Lung region extraction is an important phase of the diagnosis process. A fixed

threshold cannot perform precise segmentation of all slices of a CT scan due

to heterogeneity of the slices in a CT scan. It results in missing some juxta-

pleural nodules and the inclusion of some extra regions with the lung area,

which results in the wrong diagnosis. To overcome this challenge, there is a

need to compute the optimal threshold with some strong optimization

algorithm. In meeting this bar, a FODPSO based lung volume extraction

method is proposed for accurate segmentation of the lung area.

SVM-ensemble classification based on three base classifiers linear,

polynomial and RBF is developed for better classification of nodules and non-

nodules. Instead of using the same type of base classifiers that may follow a

similar pattern, heterogeneous base classifiers are combined to improve the

classifier performance.

A stacked autoencoder and softmax based deep learning approach is proposed

for feature reduction and classification. We have designed a two-level stacked

12

autoencoder for feature optimization. It leads to an improved classification of

nodules.

Detailed texture and shape analysis is performed to improve the classification

performance. Furthermore, for texture analysis, Gray Tone Spatial

Dependence Matrices (GTSDM) is computed based on different angular ratios

and distances among the neighborhood pixels in 2D as well as 3D. The

analysis is extended by computing the mean and range of each feature.

Intensive experimentation is performed by creating different combinations of

the feature vectors. Such detailed level analysis was lacking in the existing

methods. As a result, the proposed diagnosis methods outperformed the

existing methods.

1.11 Thesis Organization

This thesis consists of five chapters, including this chapter. Fig 1.5 depicts the

organization of the thesis hierarchy. The chapters are organized as follows:

Chapter 1 covers a brief introduction to lung cancer, its statistics and the need for

automated diagnosis systems. It also describes the key objectives, motivation and

research challenges.

Chapter 2 reviews the recently developed lung nodule diagnosis methods proposed in

recent years. This review provides details about various image analysis and machine

intelligence methods used in the literature for lung nodule detection and classification.

Chapter 3 gives the details of the proposed methods designed to overcome the

limitations discussed in the previous chapter. Four methods are presented for the

precise detection and classification of nodules. Their main focus is to improve the

diagnosis process by reducing the number of false positives while retaining high

sensitivity.

Chapter 4 illustrates the details of experimentations, datasets, and results. It also

provides a detailed comparative analysis of the proposed methods with existing

methods reported in the literature.

Chapter 5 concludes the thesis by providing a summary of the thesis contributions

and future directions of research.

13

Fig 1.5: Block diagram of thesis organization

14

Chapter 2

Literature Review

15

2.1 Introduction

During the last decade, many researchers performed baseline studies for the

development of modern IT systems to identify and characterize nodules by processing

the radiographs based on computer vision and machine intelligent algorithms. These

kinds of systems are more supportive to the radiologists in terms of time and

correctness [40].

Fig 2.1: Basic steps of a diagnostic system

Computer-based systems were initiated in 1980 at the University of Chicago [41]. The

basic aim for such applications was to provide a computer-based system with

improved reliability and consistency to the radiologist when interpreting CT images.

The method must have also been time-efficient. It was a key issue from the early

1960s, but the success rate was very low in the development of an automated

computer-based system to assist the radiologists during the examination of a patient‘s

images [42, 43]. Fig 2.1 shows the fundamental steps of a diagnostic system, which is

a four-step process. In the first step, the input image is divided into a set of regions

called segments that lead to separation of lung parenchyma as well as internal details

of the lung volume, also known as Region of Interest (ROI) [44]. The ROI extraction

is a very useful step because rather than processing the whole image, it is better to

process only those sections where actual nodule may be found. After lung region

segmentation, the next step is to extract the objects from ROI, which have similar

properties as of nodule. These objects are known as nodule candidates. A set of

candidate objects is created, including actual nodules and non-nodules where nodules

are diseased objects. In the third step, features are extracted from candidate objects.

Classification is the next step that uses image features to identify the nodules, and it

provides the second opinion to radiologists in diagnoses. Fig 2.2 shows a labeled CT

image that provides information about different regions within a slice.

16

Aorta

Air

Bronchi

Ribs

Middle lobe

Branching

Vertebral body

Upper lobe

Chest muscles

Parenchyma

Airway

Lower lobe

Fig 2.2: Labeled CT slice

2.1.1 Lung Region Segmentation

The precise segmentation is essential and crucial. It facilitates the determination of the

exact location of a nodule within a CT image that impacts the detection and treatment

of lung cancer [45, 46]. Segmentation partitions the image into different regions that

have homogeneous regions inside and heterogeneous to each other. Proper

segmentation makes life easier for radiologists [47]. Several segmentation methods

have been proposed including thresholding [48], template matching [49, 50], 3D

template matching [51], morphological operation [52] and watershed [53].

Thresholding is a simple and popular technique used for segmentation. Different

variants of thresholding techniques are reported in the literature, including simple

thresholding, optimal thresholding [54], Otsu thresholding [55], thresholding

combined with morphological operations [56], fuzzy thresholding [57] and multiple

thresholds [58-61]. Some researchers also used combinations of different

segmentation methods to improve the segmentation quality. The accuracy of

segmentation directly affects the subsequent steps, i.e., feature extraction and

classification of pulmonary nodules [62, 63].

2.1.2 Nodule Candidate Detection

The nodule candidate detection in the automated diagnosis process has prime

importance. This step requires the consideration of key properties of nodule

candidates to detect them accurately. The nodules are round shape circular objects of

different sizes ranging from 3mm to 30mm [64]. In a CT scan, the nodules appear to

be the brighter object with maximum intensity at the center and mostly within dark

surroundings. The main target of this stage is to perform initial candidate detection

17

with 100% sensitivity, which reflects the inclusion of all nodules in the candidate set.

However, a major issue of this step is the inclusion of a large number of non-nodules

in the candidate set, which leads to high false positives. The higher sensitivity with

low false positives depicts the strength of a detection method.

2.1.3 Feature Extraction

Image features play a vital role in object classification. An image consists of a large

set of pixel values, each having some useful and related features. A feature is a

numerical representation of an image or object within the image. The extraction of

these values is called feature extraction, and from those features identification of most

relevant values, which can better represent an image, is called feature selection. On

the basis of selected features classification of nodules and non-nodules is performed,

various machine learning classifiers are used for this purpose [65]. A poor selection of

features can lead to the wrong diagnosis. The right feature selection improves the

diagnosis system performance, which shows the effectiveness of the feature selection

and extraction process [66].

2.1.4 Classification

Classification is the final step of the diagnosis process, which distinguishes nodules

and non-nodules. A classifier is trained on labeled examples. In the testing phase,

unlabeled data is provided to the trained classifier, which classifies it to a relevant

class [67]. SVM [68], Bayesian [69], and Artificial Neural Network (ANN) [70, 71]

are widely used classifiers in the automated diagnostic systems. Many researchers

have also used multiple classifiers by combining their results to form an ensemble

classifier [72]. The classifiers are trained and tested with the features extracted from

the candidate nodules.

2.2 Recent Developments in Lung Nodule Diagnosis

Automated diagnostic systems for lung cancer have remained a key area of research

in recent years. There is a primary issue of sensitivity and false positives in these

diagnosis systems. Lung nodule diagnosis methods are mainly comprised of

segmentation, feature extraction and classification [73]. The segmentation is further

divided into lung region extraction and nodule candidate detection. Furthermore, to

evaluate the performance of a diagnosis system, a standard dataset also plays a vital

18

role. Table 2.1 gives valuable details of different methods used in automated lung

nodule diagnosis. The methods proposed in these articles are categorized by novelty

area, in this section.

Table 2.1: Categorization of lung nodule detection methods

Author Year Feature Types Classification/Detection Methods Category

Armato, et

al. [74]

1999 Geometric and gray

level features

Gray-level thresholding Linear

Discriminant Analysis (LDA)

classifier

Thresholding

Armato III,

et al. [59]

2001 Morphological

and gray level

features

Multiple gray-level thresholds Thresholding

Lee, et al.

[75]

2001 Shape and texture Genetic algorithm template

matching (GATM)

Template

Matching

Armato, et

al. [76]

2002 Geometric features Gray-level thresholding, cascaded

and rule-based classifiers

Supervised

Learning

Gurcan, et

al. [77]

2002 Morphological

features

K-means, rule-based classification,

LDA

Unsupervised

Learning

Lin and Yan

[78]

2002 Circularity, size of

the area, and mean

brightness

Neural fuzzy model Fuzzy Logic

Zhao [79] 2003 Size and compact

shape

Local density maximum algorithm

based on thresholding

Thresholding

Kostis, et al.

[80]

2003 Morphological

features

Mathematical morphology and

gray level thresholding

Morphology

Based

Bae, et al.

[81]

2005 Morphological

features

3D and 2D morphological

matching

Morphology

Based

Lin, et al.

[82]

2005 Number of pixels,

gray value, Size of

area of the region

Neural network-based fuzzy model Fuzzy Logic

Zhang, et al.

[83]

2005 Local shape features Discrete-time cellular neural

network (DTCNN), genetic

algorithm (GA)

Supervised

Learning

Boroczky, et

al. [84]

2006 Geometric features

optimized using GA

GA for feature reduction and SVM

for classification

Supervised

Learning

Dehmeshki,

et al. [49]

2007 3D geometric Shape-based GATM Template

Matching

19


Osman, et

al. [85]

2007 3D shape features 3-D template-matching Template

Matching

Retico, et al.

[86]

2008 Morphological

features

Dot enhancement filter and Neural

Network

Unsupervised

Learning

Ozekes, et

al. [51]

2008 Shape properties 3D template matching, fuzzy

thresholding, and G-CNN

Template

Matching

Ye, et al.

[57]

2009 3-D local geometry

and statistical

intensity features

Adaptive fuzzy thresholding

method for segmentation, SVM, a

rule-based classifier

Hybrid

Method

Murphy, et

al. [87]

2009 Local shape features kNN classifier Supervised

Learning

da Silva

Sousa, et al.

[88]

2010 Geometric features,

Texture features

Mathematical morphology,

thresholding, SVM

Hybrid

Method

Messay, et

al. [89]

2010 Geometric, intensity

and gradient features.

Fisher Linear Discriminant (FLD)

and quadratic classifiers

Supervised

Learning

Lee, et al.

[90]

2010 Gray level values Random forests based

classification

Supervised

Learning

Lee, et al.

[91]

2010 Geometric and

texture features

Ensemble classification Supervised

Learning

Chen, et al.

[92]

2011 Texture, geometric Neural network ensemble Supervised

Learning

Riccardi, et

al. [56]

2011 Geometrical features,

maximum intensity

projection

Histogram thresholding,

mathematical morphology, radial

filtering, SVM

Hybrid

Method

Magalhães

Barros

Netto, et al.

[93]

2012 Shape and texture SVM classifier Supervised

Learning

Cascio, et al.

[94]

2012 Geometric and

intensity

3D Mass-Spring Model Supervised

Learning

Chen, et al.

[95]

2012 Morphological

features

Neural network Supervised

Learning

Choi and

Choi [96]

2012 Geometric, intensity,

DCT, wavelet,

texture features

Genetic programming based

classifier

Supervised

Learning

20


Elizabeth, et

al. [60]

2012 Shape and texture Thresholding Thresholding

Teramoto

and Fujita

[97]

2013 Shape and intensity Thresholding, cylindrical nodule

enhancement filter

Hybrid

Method

Choi and

Choi [98]

2013 Geometric and

texture features

Thresholding method for

segmentation and SVM

Supervised

Learning

Wang, et al.

[99]

2013 Texture and shape

features

SVM based on three-dimensional

matrix patterns

Supervised

Learning

Tartar, et al.

[100]

2013 Morphological

features and patient

information

Decision tree Supervised

Learning

Keshani, et

al. [101]

2013 2D stochastic and 3D

anatomical features

Adaptive fuzzy thresholding and

active contour models (ACM)

Supervised

Learning

Jang, et al.

[102]

2013 2D and 3D features Fuzzy clustering and genetic

algorithm

Fuzzy Logic

Choi and

Choi [54]

2014 Three-dimensional

shape-based feature

Dot enhancement filtering/SVM Supervised

Learning

Kuruvilla

and

Gunavathi

[103]

2014 Statistical parameters Artificial neural network Supervised

Learning

Cao, et al.

[104]

2014 Intensity, shape, and

gradient features

Ensemble learning Supervised

Learning

de Carvalho

Filho, et al.

[105]

2014 Shape and texture Thresholding based segmentation Hybrid

Method

Badura and

Pietka [106]

2014 Texture features Fuzzy connectedness and the

evolutionary computation

Hybrid

Method

Brown, et al.

[107]

2014 Shape features Watershed, intensity thresholding

and Euclidean Distance

Transformation (EDT). A vector

quantization (VQ)

Hybrid

Method

Taşcı and

Uğur [108]

2015 Texture features Otsu thresholding, morphological

operations, generalized linear

model regression (GLMR)

classifier.

Thresholding

21


Akram, et

al. [109]

2015 Geometric and

intensity based

statistical features

Artificial neural network Supervised

Learning

Wang, et al.

[110]

2015 Shape features Spherical shape enhancement

filter, Chan–Vese (CV) model

Supervised

Learning

Kaya and

Can [111]

2015 Shape, size, and

texture features

Ensemble classification Supervised

Learning

Shi, et al.

[112]

2015 Gray values ROI extraction based on hessian

matrix and Laplacian of Bilateral

(LoB)

Unsupervised

Learning

Dai, et al.

[113]

2015 No features Graph cuts algorithm with

Gaussian mixture models (GMMs)

Unsupervised

Learning

Han, et al.

[114]

2015 Geometric, intensity,

gradient, and hessian

features

Different levels of vector

quantization for lung and nodule

segmentation.

Unsupervised

Learning

Shen, et al.

[115]

2015 Contextual features Bidirectional chain Code, SVM Hybrid

Method

Hua, et al.

[116] 2015 Texture and

morphological

features

Deep belief network Deep

Learning

Firmino, et

al. [117]

2016 HOG features Region growing, SVM and rule

base classifiers

Supervised

Learning

Mukhopadh

yay [58]

2016 Texture and intensity Multiple thresholds

Thresholding

Nibali, et al.

[118]

2017 Deep convolutional

features

Deep residual learning, curriculum

learning, and transfer learning

Deep

Learning

Dou, et al.

[119]


features

multi-level 3D CNN Deep

Learning

da Silva, et

al. [120]


features

Deep learning and genetic

algorithms.

Deep

Learning

Sun, et al.

[121]

2017 Shape, texture and

deep features

CNN, DBN, Autoencoder Deep

Learning

Gupta, et al.

[122]


intensity features

Neural network classifier Supervised

Learning

22


Zhang, et al.

[123]

2018 Morphology based

3D skeletonization

feature, shape feature

and intensity features

ACM, thresholding, morphological

operation,, SVM

Morphology

Based

Jaffar, et al.

[124]

2018 Intensity and gradient Differential evolution based

thresholding

Hybrid

Method

Teramoto

and Fujita

[125]

2018 Shape and texture Thresholding, nodule enhancement

filter, rule based classifier, SVM

Hybrid

Method

Gong, et al.

[126]

2018 Shape, intensity and

texture features

Random forest classifier, level set

segmentation

Hybrid

Method

Ali, et al.

[127]

2018 Deep features Reinforcement learning based deep

neural network

Deep

Learning

Jiang, et al.

[128]

2018 Deep features Four channel CNN, Frangi

filter

Deep

Learning

Li, et al.

[129]

2018 Deep features CNN-ensemble Deep

Learning

Huidrom, et

al. [130]

2019 Shape, intensity and

gradient features

Neural Network optimized with

GA and PSO

Supervised

Learning

Shaukat, et

al. [131]


intensity features

ANN, thresholding and watershed

segmentation

Supervised

Learning

Xie, et al.

[132]

2019 Deep features with R-

CNN

R-CNN and Boosting CNN Deep

Learning

Kasinathan,

et al. [133]

2019 Deep features ACM, enhanced AlexNet Deep

Learning

2.2.1 Supervised Learning Methods

It is a machine learning process in which a classifier is trained on a set of data with

predefined class labels. After the training process, the classifier is tested against the

test data to classify the desired patterns in the correct classes [67]. In [103]

feedforward and backpropagation neural networks were used. To train the neural

network, they used statistical features extracted from the segmented lung portion of

CT images. Thirteen predefined training functions for backpropagation network were

used in addition to two newly proposed training functions. Best accuracy was

23

achieved with backpropagation algorithm with the training function.

Another ANN, thresholding and intensity feature based method was proposed in

[109]. A 3D matrix model by taking SVM as a base classifier was applied to detect

the nodules from 1D and 2D models [99].

Chen, et al. [95] presented a diagnosis system based on neural network and logistic

regression methods. Limitation of this approach is that it was only tested on a small

dataset, and no external evaluation was done. Discrete Time Cellular Neural Network

(DTCNN) was applied to identify non-pleural nodules from helical CT images [83].

To train the classifier, local shape features were extracted. For juxta-pleural nodules

detection, morphological, thresholding and labeling were performed. After that, the

classification was applied by training Linear Discriminant Analysis (LDA) classifier

on geometric and gray level features. Murphy, et al. [87] used local shape features and

kNN classifier. The system was tested over 813 scans and 80% sensitivity with 4.2

FPs/scan.

SVM-based supervised learning was developed for lung cancer diagnosis [93]. Before

applying SVM, nodule segmentation was carried out as a preprocessing step. After

that, shape and texture features were extracted for training the classifier. Another

supervised learning genetic programming classifier [96] was developed. In the first

step, lung area was segmented from CT image using thresholding. In the second step,

multiple thresholds combined with 3D pruning were applied to identify nodule

candidates. Finally, features were extracted, and by using these features classification

is performed. The performance is evaluated over LIDC.

A three stage hierarchical block classification based diagnosis system was developed

in [98]. The 3-D CT images were divided into hierarchical blocks and by entropy

analysis method required blocks were identified. After that, the thresholding method

was used for segmentation. Three types of features were extracted from segmented

regions, including 2-D geometric, 2-D texture, and 3-D geometric features. Finally,

SVM was applied to classify pulmonary nodules.

Choi and Choi [54] proposed another lung cancer detection method based on three-

dimensional shape features. To refine the feature vector, a wall elimination method

was used. Multi-scale dot enhancement filter [98] was used to obtain lungs portion,

and SVM was used for classification.

24

An optimum feature subset selection method for improvement of the diagnosis

process was presented in Boroczky, et al. [84]. For this purpose, they used a genetic

algorithm for optimal feature selection from 2D and 3D geometric features of lung CT

images. From each structure, 23 features were selected and then ten best features with

high relevancy were extracted by using GA optimization. SVM was trained on the

selected feature subset, which achieved 100% sensitivity and 56% specificity. Chan–

Vese (CV) model [110] consisting of three steps, including segmentation, candidate

nodule extraction, and FP reduction, improved the diagnosis process.

A feature based lung nodule detection technique [89] was developed, comprising of a

feature set of shape, intensity and gradient features, extracted from segmented

images. Different subsets of these features were used in classification. A lung nodule

detection method was proposed in [101], where two segmentation techniques adaptive

fuzzy thresholding and Active Contour Models (ACM) were collectively used. SVM

was applied for nodule classification with 2D stochastic and 3D anatomical features.

An automated diagnosis system was developed to identify the nodules which might be

missed on initial CT scan visual analysis [76]. For this purpose, they used multiple

techniques, including gray-level thresholding, rule-based scheme, and a cascaded

automated classification. Thresholding was used to detect 3D structures within the CT

scan. After that, an automated analysis was done on the detected structures to identify

nodule candidates. In the third step, cascade and rule-based classifiers were used to

identify actual nodules. Stable 3D Mass-Spring Model (MSM), was proposed to

detect small sized pulmonary nodules and juxta-pleural nodules [94]. A diagnosis

system based on SVM and rule base classifiers was proposed in [117]. Region

growing was applied to extract nodule candidates, and classifiers were trained with

HOG features.

Gupta, et al. [122] presented a three phased method. In the first phase, they performed

the lung region segmentation based on thresholding approach. In the second phase,

nodule candidates were extracted using morphology. In the final step, nodule

classification was performed based on a three-layer neural network. A neural network

classifier optimized with GA and PSO was developed for nodule detection and

classification [130]. Intensity, geometric and gradient features were extracted to train

the network. A two layer ANN was trained using shape, intensity and texture features

25

to detected nodules [131]. Before classification, lung region and nodule candidates

were extracted using threshold and watershed algorithms.

Ensemble classification is another supervised learning technique that performs

classification by combining the decision of individually trained classifiers [134]. A

nodule detection method was proposed in [111] based on shape, size, texture features

and ensemble classification. An ensemble classifier based on the random forest

supported by the clustering applied for diagnosis [90]. Firstly, nodules and non-

nodules instances were clustered, and after that, by training labels, random forest

classification was applied. In this technique, clustering was combined with the

classification to achieve high accuracy. Another ensemble classification technique

based on the combination of a genetic algorithm and random subspace method to

improve the performance was proposed [91]. For the classifier training, the features

were selected in two steps. Initially, 126 geometric and texture features were selected

and divided into different subsets to find a more appropriate set of features to achieve

better performance.

A binary ensemble classifier [104] was applied to reduce the false positive rate.

Weighted voting was applied for the ensemble of classifiers. Another ensemble

learning technique was proposed using the neural network ensemble to detect nodules

from CT images [92]. To train the networks, texture and geometric features were

used. Afterwards, they combined the individual networks, and to improve the results;

they used Bayesian rule for ensemble learning. Resultant ensemble neural network

produced better results than individual networks.

2.2.2 Unsupervised Learning Methods

Segmentation is a crucial factor for lung nodule detection in a diagnosis system.

Various segmentation techniques were proposed by different researchers in the last

two decades. Unsupervised learning techniques, including k-means [135], Fuzzy c-

means (FCM) is a variation of fuzzy logic are useful clustering algorithm [136]. In

[137] FCM was used for lung image segmentation from CT of nodule from the lung

region. In this section, different unsupervised learning methods are discussed in detail.

Tolouee, et al. [138] proposed a three-stage model for high-resolution CT images,

which mainly comprised of extraction of a region of interest, feature extraction and

finally, classification to normal and abnormal lung tissue patterns. For this purpose, a

26

set of thresholding, filtering, and morphological operators were used. For feature

extraction, wavelet features were extracted and finally, fuzzy k-nearest neighborhood

classifier was used for nodule classification.

Hessian matrix and Laplacian of Bilateral (LoB) filter was developed, which extends

Laplacian of Gaussian [112], but it needs high computational time. Han, et al. [114]

proposed different levels of vector quantization for lung and nodule segmentation.

The geometric, intensity, gradient and hessian features were used for false positive

reduction.

Gurcan, et al. [77] proposed a diagnosis system in which they used k-means clustering

for nodule candidate segmentation. After that, a rule-based classifier was combined

with linear discriminant analysis (LDA) and produced good accuracy. A graph cut

algorithm with Gaussian mixture models (GMMs) for lung segmentation was

proposed to improve the lung nodule detection [113].

A neural network based unsupervised learning algorithm was proposed in [86], which

used dot enhancement filter to select nodule candidates by enhancing spherical-

shaped objects. The neural classifier was developed, which used a training elimination

approach similar to the Self-Organization Map (SOM) for false positive reduction.

2.2.3 Diagnosis Based on Template Matching

Template matching is an emerging area in computer vision. It is an effective way to

recognize the objects from an image by using a template image [139]. In medical

imaging use of template matching has gained a profound interest in recognizing

abnormal growth within the human organs [140]. However, there is one limitation that

it cannot produce good results with small sized objects. Lee, et al. [75] developed two

template matching techniques. The first technique, named Genetic Algorithm

Template Matching (GATM) was used for extracting nodules. The second was a Lung

Wall Template Matching (LWTM) technique supported by conventional semicircular

models to detect those nodules which were attached to the lung walls. After initial

identification, the next task was to improve the results by reducing the false positive

rate for which they used thirteen features containing inverse difference moment,

entropy, area, and contrast. GATM was extended using the shape properties of

nodules [49]. 3D shape based features combined with global intensity distribution

were used to define fitness function.

27

Osman, et al. [85] presented a 3D template matching technique for lung nodule

detection. At the first step, ROI was identified by an eight directional search in each

slice. Combining 2D images, a 3D image was generated, which was matched with a

predefined 3D template to identify nodules. The evaluation of this technique was

performed only over six cases of LIDC dataset. Ozekes, et al. [51] proposed Genetic

Cellular Neural Networks (G-CNN) for region lung segmentation. The use of Fuzzy

rule based thresholding produced 100% sensitivity but a high false positive rate.

2.2.4 Diagnosis Based on Fuzzy Logic

Fuzzy logic is used in various kinds of decision problems [141]. It is based on

approximate reasoning ranging from a decision value from 0 to 1. It has some

applications generally in image processing and particularly in medical imaging. Lin

and Yan [78] proposed a fuzzy logic based pulmonary nodule detection method. In

the first step, lung area was extracted using some image processing techniques

comprising thresholding, morphology closing, and labeling. From the segmented lung

region, three features circularity, area, and mean brightness were extracted. A fuzzy

neural model was created and trained by these features to detect pulmonary nodules

with good accuracy. Lin, et al. [82] proposed another fuzzy-based method, which was

an extension of the previous approach with improvement in false positive rate. A

fuzzy rule based system using 2D and 3D features was given in [102]. The system

was optimized using a genetic algorithm (GA).

2.2.5 Diagnosis Based on Thresholding

Thresholding is a simplest but an effective way to identify objects within the image

[142]. It is also a very useful method for image segmentation by using pixel

intensities to partition the image [143]. In medical imaging, thresholding is a very

useful method to identify objects like nodules, vessels, and lesions depending upon

different gray levels. A thresholding based diagnosis system for cancer detection from

chest CT images was presented in [60]. Multiple thresholding operations were

performed during segmentation, edge enhancement, background elimination, and

nodule boundary detection. Finally, a probabilistic neural network was applied by

shape and texture features to distinguish the infected and normal sections of the

image. Thresholding also produced good results in detection [74]. To extract lung

28

volume from image gray-level thresholding was used as a segmentation technique.

After that, three-dimensional objects were extracted as nodule candidates, and finally,

they used linear discriminate analysis for final nodule identification with better

performance. Mukhopadhyay [58] proposed a segmentation framework for lung

nodules. The author used different thresholding techniques to segment region of

interest (ROI) and nodules of different types. The method in [59], performed 2D

preprocessing, 3D analysis and finally feature extraction. In 2D preprocessing, lung

volume was extracted using gray level thresholding. In the next step, thirty-six

thresholds (ranging from 50 to 225 gray levels) were applied to identify the nodule

candidates. Finally, to reduce false positive rate, a feature vector with nine features

was created, including three gray levels, three morphological and three 2D features.

On the basis, of this feature vector classification of nodule and non-nodule was

performed by using linear discriminative analysis.

A local density maximum algorithm was proposed by Zhao [79]. The diagnosis

process mainly consisted of three steps in which different types of thresholds were

used to segment lung portion, nodule candidate selection, and false positive reduction.

Shape and texture feature based juxta-pleural nodule detection method was proposed

by Taşcı and Uğur [108], in which they introduced seven new features. Otsu

thresholding and morphological operations were used for segmentation, and finally,

for classification, they tested ten different classifiers and found the best results for the

generalized linear model regression (GLMR) classifier. Teramoto and Fujita [97]

proposed a nodule enhancement method to improve the detection process. The

thresholding was applied for ROI extraction, and a cylindrical filter was proposed to

improve the detection process.

2.2.6 Diagnosis Based on Mathematical Morphology

Mathematical morphology is an effective image processing technique used in

segmentation [144]. Various researchers have developed pulmonary nodule detection

methods based on the morphological operations fully or with partial support. Kostis,

et al. [80] performed an automated diagnosis based on mathematical morphology.

This technique used a different combination of morphological filters to extract lung

nodules from high-resolution CT images. A morphological matching algorithm

improved the detection performace [81]. The system comprised of a combination of

29

2D and 3D of morphology operations. At first, 2D operations were performed based

on gray-level thresholding combined with morphological operations to extract lungs

portion from CT image. After that, 3D region growing and labeling with the support

of morphological closing were used, which further leads to extract the nodule

portions. Finally, by shape and nodule geometry classification was performed to

differentiate nodule and non-nodules.

Zhang, et al. [123] applied various morphological operations combined with other

image processing and machine learning techniques. At first, lung region extraction

was performed by the active counter model. After lung region segmentation, nodule

candidates were detected by using thresholding, morphological operations, and

connected component labeling. After that, feature extraction was performed

comprising of 3D skeletonization features, texture and shape features.

2.2.7 Multiple Techniques/Hybrid Approaches

Recently, researchers are interested in developing hybrid techniques by combining

multiple algorithms. Ye, et al. [57] presented a multistep method which comprises

various algorithms to detect lung nodules automatically. In the first step, segmentation

was performed on input CT images to separate the lung portion from the remaining

parts of the image. Fuzzy thresholding algorithm was used for this purpose. In the

next step, preprocessing was performed on the segmented regions; for this purpose

anti-geometric diffusion was used. In this method, the edges were diffused from the

image which supports the precise geometric feature extraction. A weighted SVM

minimized the chances of FPs resulting in more than 90% accuracy. Another lung

nodule detection algorithm comprising of six stages was presented in [88]. At the

first stage, they eliminated the additional information related to the patient‘s body and

labeled the stage as thorax extraction. In the next stage, they extracted the correct

portion of lung parenchyma [145] to support the accurate identification of dense

structures inside it. Afterwards, the tubular structure was eliminated. All these

eliminations improved the identification of correct portion of nodules. These stages

used thresholding, region growing, genetic algorithm and SVM [105]. The technique

was tested against LIDC-IDRI dataset with 80%, 20% data for training and testing,

respectively.

30

A multilevel segmentation technique for lung nodule detection presented in [106].

The fuzzy connectedness method comprised of mask generation, automatic seed point

selection, and adaptive thresholding. The process was used to remove the blood

vessels. In the next phase, automatic seed point selection was performed, which was

based on thresholding and 2/3D connectivity analysis. Finally, genetic algorithm and

multiple morphological operations were performed for final detection. For

experimentation images from LIDC and LIDC-IDRI databases were used and 100%

detection rate with 5% false positive was achieved.

Brown, et al. [107] presented a segmentation and detection method of nodules based

on watershed, intensity thresholding and Euclidean Distance Transformation (EDT).

Han, et al. [61] proposed vector quantization (VQ) method for lung ROI segmentation

and nodule candidate detection. A hybrid method was developed that included various

techniques applied to CT images one after other [56]. The system diagnosed cancer in

three major steps including lung tissue segmentation based preprocessing step, a

filtering step, and false positive reduction step. The histogram thresholding, seeded

region growing, and morphological filters were applied to the input CT images for

preprocessing and extraction of lung volume from the raw images. The nodule

candidates were identified, and geometric features were extracted using a 3D radial

based filter and scale space analysis, respectively. The final step was the false positive

reduction, in which two approaches were used, including heuristic and supervised. In

the heuristic approach, simple thresholding based on geometric features was used to

reduce the number of ROIs. On the other hand, in a supervised approach SVM was

used based on heuristics and maximum intensity projection Zernike moments [146].

A hybrid method was proposed by Shen, et al. [115] to identify juxta-pleural nodules

from CT images by combining chain code and SVM classifier. Bidirectional chain

code was used to detect nodule boundary. Three features, including relative boundary

distance, boundary segment concave degree, and relative position information were

used to train the classifier. Finally, SVM was used for classification. Jaffar, et al.

[124] used differential evolution based thresholding for lung segmentation, and then

ensemble learning was applied on the basis of gradients and intensity features for

classification of nodules.

Teramoto and Fujita [125] proposed a hybrid method for PET/CT images.

Thresholding and cylindrical nodules enhancement filters were applied. Rule based

31

classifier and three SVM classifiers were trained over shape features for final

classification of nodules. Gong, et al. [126] presented a hybrid diagnosis method. Otsu

thresholding, region growing and morphological filtering were applied for lung region

extraction. Afterwards, 3D tensor filtering and level set segmentation along with local

image information was used for nodule candidate detection. Subsequently, shape,

intensity and texture features were extracted. On these features, random forest

classifier was trained and tested to distinguish nodules and non-nodules.

2.2.8 Deep Learning Based Methods

Artificial intelligence technology is changing the lives of people and becoming a hot

spot in the development and research of science and technology [147]. Deep learning,

such as artificial intelligence technology, builds a neural network imitating the

neurons of a person to form the learning machine is a unique idea in the field of

artificial intelligence. In particular, in recent years, deep learning has become an

important and dynamic area of research that provides a platform for different

classification problems [148]. The key feature of deep learning is to create an

extensive network of neurons that improves the accuracy of classification and

prediction problems [149]. A major focus of deep learning research in computer

vision is on natural images. However, in recent years, the researchers have also

applied deep learning algorithms in medical imaging, but their results are not

promising [118].

Hua, et al. [116] used a deep belief network for lung nodule classification using

texture and morphological features and compared the performance with the

conventional neural network. Another deep learning method was presented by Sun, et

al. [121]. They used three deep learning algorithms, including autoencoder, deep

belief network and CNN. The method was evaluated over LIDC-IDRI dataset. Dou, et

al. [119] also proposed deep learning based method. A multi-level 3D CNN

framework was developed for this purpose. Another deep learning based method was

proposed by Ali, et al. [127]. Reinforcement learning based deep neural network was

trained and tested for automated detection of nodules. da Silva, et al. [120] proposed

nodule classification method by combining deep learning and genetic algorithms.

Jiang, et al. [128] presented a patch based deep learning approach for lung nodule

detection from 2D images. The lung segmentation was performed based on connected

32

component analysis and morphological operation. The multiscale Frangi filter was

used for nodule candidate detection by removing the vessels. Finally, classification

was performed based on CNN. The method was tested over LIDC-IDRI dataset.

Li, et al. [129] designed three convolutional neural networks with three different input

sizes, including 12x12, 32x32 and 60x60, for nodule candidate detection. For

classification of nodules, they developed an ensemble CNN and achieved 95%

sensitivity with 5.0 FPs per image. Xie, et al. [132] proposed 2D CNN for nodule

candidate detection and boosting technique. Kasinathan, et al. [133] enhanced a CNN

based on AlexNet architecture for nodule classification. The candidate nodules were

detected based on ACM and local image bias field formulation. The method achieved

97% sensitivity with 3.8 FPs/image.

2.3 Discussion and Analysis

The current review of lung nodule detection methods mainly focuses on the methods

given during recent times. These methods are critically viewed under some analysis

parameters, especially on performance, nodule size, datasets, and nodule types. The

primary focus of a diagnosis system is the detection and classification of nodules with

high sensitivity and less number of false positives. It is a complicated task to compare

the soundness of different methodologies due to the different parameters used during

their analysis, e.g. if two diagnosis systems have produced the results on different

datasets, focused different kinds of nodules as well as sizes then it is quite

unjustifiable to compare those. In recent years, some authors have claimed sensitivity

higher than 95% [51, 54, 84, 85, 90, 109]. However, these methods were unable to

overcome the false positive rate. Osman, et al. [85] and Ozekes, et al. [51] achieved

100% sensitivity, but their results are on a dataset containing only large nodules of

size greater than 5mm. Lung nodule detection methods lack in achieving high

accuracy and have high false positive rate due to limitations in the handling of

different types of nodules like juxta-pleural and juxta-vascular nodules, especially of

smaller sizes due to the following reason. These issues are due to the complex

structure of the human lung.The limitations of the existing methods are illustrated in

Table 2.2.

33

Table 2.2: Limitations in recently developed methods

Methodology Year Limitations

Ozekes, et al. [51] 2008 Used larger nodules of >5.6 mm. High FPs/scan of 13, template

matching was used and failed to deal with smaller nodules.

da Silva Sousa, et al.

[88]

2010 Low sensitivity of <85%. The method was not evaluated over some

standard datasets. Large size nodules were included in the evaluation.

Messay, et al. [89] 2010 Accuracy is just 80%; juxta-pleural nodules were missed during the

smoothing process.

Magalhães Barros

Netto, et al. [93]

2012 Sensitivity is low at 85.9%; evaluation was performed with a small

dataset containing 48 nodules.

Keshani, et al. [101] 2013 Low sensitivity 89% and high FPs/scan 7.3, multiple diseases were

handled with features of the same type resulting in poor performance.

Choi and Choi [98] 2013 Used a subset of dataset for performance evaluation.

Choi and Choi [54] 2014 High false positive rate of 6.67/scan; evaluation was performed over

subset of dataset containing 148 nodules.

Kuruvilla and

Gunavathi [103]

2014 Only statistical features were used. A very high false positive rate of

30/scan.

Cao, et al. [104] 2014 Accuracy < 85%. False positives were not identified.

de Carvalho Filho, et

al. [105]

2014 Sensitivity is very low 86%. FPs/scan not identified.

Badura and Pietka

[106]

2014 Accuracy is given, but no values of sensitivity and specificity are

mentioned. It used only those nodules which were annotated by all

radiologists.

Wang, et al. [110] 2015 Performance is not properly measured; only false positive rate is

mentioned. Performance of system not evaluated by a standard dataset.

Shen, et al. [115] 2015 Method is evaluated only in terms of accuracy. It handled juxta-pleural

nodules only.

Kaya and Can [111] 2015 Low sensitivity of 82%. Number of FPs not reported.

Hua, et al. [116] 2015 Poor results of 73% sensitivity, 82% specificity. No. FPs mentioned.

Features are not properly elaborated. Method is not tested over some

standard dataset.

Taşcı and Uğur

[108]

2015 Method was evaluated in terms of accuracy. Only juxta-pleural

nodules were addressed. Only one type of feature was used. FP not

reported.

Mukhopadhyay [58] 2016 Very low sensitivity. Only segmentation is performed.

34

Methodology Year Limitations

Firmino, et al. [117] 2016 Only one type of feature was used for classification. HOG features

Nibali, et al. [118] 2017 A large number of deep features are fused which increases the

computation time. Moreover, FPs are not reported.

Zhang, et al. [123] 2018 Used a smaller subset of LIDC-IDRI dataset of 71 scans containing

168 nodules.

Huidrom, et al.

[130]

2019 The performance was evaluated over 300 scans, but nodule

information is not provided. Moreover, the number of FP was not

reported.

Shaukat, et al. [131] 2019 Simple thresholding was applied to extract the lung ROI. It leads to

missing the juxta-pleural nodules. The method was evaluated on 84

scans.

Xie, et al. [132] 2019 The nodule detection is based on 2D properties, which leads to the

wrong diagnosis. As a result, sensitivity is low as 83% and FPs are

very high as 8/scan.

Following are the major findings of this review:

The diagnosis methods miss juxta-vascular nodules during the process of

vessel removal, which leads to the wrong diagnosis [35, 48, 54-56, 89, 94, 97,

103, 124].

There is a challenge of high sensitivity with low FPs/scan [48, 54-56, 61, 89,

94, 97, 103, 107, 118, 119, 123, 124, 127].

One type of feature vector cannot represent all kinds of nodules [83].

In morphological operation, window selection cannot work well for all slices

and can miss juxta-vascular nodules [84].

The segmentation process can completely or partially eliminate juxta-pleural

nodules [55]. During the diagnosis process, a fundamental step is the

extraction of the lung portion. Different segmentation methods are used for

this purpose. During this process, juxta-pleural nodules are completely or

partially removed from ROI due to their presence on the boundary of the lung.

The detection of these nodules is another challenge.

Template matching leads to wrong candidate selection due to similar shaped

objects, especially for small nodules [49, 51, 75, 85].

35

The nodule candidate detection is a key step in intelligent lung nodule

detection and classification.

The shape features, along with intensity and texture features can be better

representative of the nodules.

Considering these findings, we have designed our proposed methods to improve the

detection process and reduce the number of false positives. In the first phase of lung

region extraction, optimal thresholding is applied which varies from slice to slice. It

results in precise segmentation. Further analysis is also performed for boundary

correction to ensure the detection of juxta-pleural nodules. In the initial candidate

detection process, 3D properties are taken into account. In addition, 2D as well as 3D

connectivity and shape properties of nodules and vessels are incorporated to separate

the nodules from vessels. In the third phase of feature extraction, instead of using one

type of feature, shape, intensity, gradient and texture features are extracted.

Furthermore, both 2D and 3D features are extracted for better identification of

nodules. In Method II and IV, feature reductions are also applied which reduces the

training time of the classifier. In the final phase of classification, we investigate the

effects of single classifier (Method I), multiple classifiers (Method II), ensemble of

classifiers (Method III) and deep learning (Method IV). This impacts performance and

also helps in comparative analysis. Furthermore, in Method III and IV we have

applied 3D segmentation of juxta-vascular nodules instead of conventional 2D

segmentation, which requires restructuring of nodules. This improves the overall

performance in general and the detection of juxta-vascular nodules in particular.

Another key feature of the proposed methods is that these have the ability to deal with

different types of nodules, including isolated, juxta-pleural or juxta-vascular.

2.4 Summary

In this chapter, a systematic review has been presented of diagnostic systems for

pulmonary nodules from CT images. Our analysis found that LIDC-IDRI is the

largest and the most authentic benchmark for performance evaluation of diagnostic

methods, compared to other databases. Additionally, feature extraction, classification,

and segmentation techniques were discussed in detail related to different diagnosis

systems. During the analysis of performance evaluation of diagnosis systems, an issue

of high false positive rate produced during the nodule detection process was primarily

36

due to juxta-pleural, juxta-vascular and smaller nodules. Consequently, there is a need

to overcome these issues to improve the performance of diagnosis systems. This

review will help researchers to find accurate information about lung cancer and its

diagnosis.

37

Chapter 3

Proposed Methods

38

3.1 Introduction

This chapter presents four methods for lung nodule detection and classification, which

are proved to be more effective than the existing ones. In the first method, optimal

thresholding and polygon approximation is performed for ROI extraction and nodule

candidate detection respectively. Nodule candidates are enhanced, and a hybrid

feature vector is proposed by encapsulating shape, intensity, and HOG features. SVM

is applied for the classification on the basis of the proposed hybrid feature vector. In

the second method, nodule candidate detection is improved by applying ACM and

shape properties of nodules. Moreover, feature reduction is performed to create a

better feature descriptor, which improves the classification performance. Four

classifiers are trained and tested to classify the nodules. In the third method, a novel

3D multistage segmentation method is proposed to improve the lung volume

segmentation and nodule candidate detection process. Furthermore, SVM-ensemble

classifier is proposed for better classification of nodules. In the fourth and final

method, a diagnosis method is proposed in which every step of the diagnosis process

is improved. Threshold selection is improved by FODPSO optimization. A novel

method for nodule candidate detection is proposed based on geometric fit in

parametric form. In feature extraction, 2D, as well as 3D shape and texture features

are extracted. In texture features, spatial information is also incorporated to extract

true representative features. Finally, deep learning approaches, including autoencoder

and softmax are trained and tested for feature reduction and classification. A block

diagram in Fig 3.1 illustrates all the proposed methods and their major steps. The

proposed methods are listed below, and details are discussed in the succeeding section

of this chapter.

i. Lung nodule detection using polygon approximation and hybrid features from

lung CT images

ii. 3D nodule candidate detection supported by hybrid features to reduce false

positives

iii. Multistage segmentation model and SVM-ensemble for precise lung nodule

detection

iv. Lung nodule detection and classification based on geometric fit in parametric

form and deep learning

39

Fig 3.1: Block diagram of proposed methods

40


and Hybrid Features from Lung CT Images

A lung nodule detection method is presented in this section which is based on CT

images.

Fig 3.2: Major steps of the proposed method

The pictorial representation of basic units of the proposed system is given in Fig 3.2.

It shows three significant steps, which are further divided into substeps. At first, lung

region is separated by optimal gray level thresholding and morphological operations.

On the next step, polygon approximation is used to extract nodule candidates from the

region of interest (ROI) by eliminating vessels. After that, nodule candidates are

enhanced, which leads to improved classification performance. Finally, the false

positive reduction step is performed, comprising of two substeps, feature extraction

and classification. A hybrid feature vector is provided as input to the classifier which

41

is composed of geometric, intensity and Histograms of Oriented Gradients (HOG)

features. And for classification, SVM is used to distinguish nodules and non-nodules

by features, extracted in the previous step.

The main contributions of this study are:

1. Lung region extraction is performed on the basis of optimal gray level

thresholding.

2. A novel nodule candidate detection method based on polygon approximation

is proposed.

3. A hybrid feature vector is created for better training of classifiers.

4. SVM classifier is trained and tested with different kernels for better

classification of nodules.

3.2.1 Lung Region Extraction using Optimal Threshold

Thresholding is a simple and effective method of segmentation in image processing

[142]. Segmenting a grayscale image creates a binary image with reference to the gray

values of related regions. If the intensity of a pixel is less than the threshold, then the

corresponding pixel in the mask image is set to white, and otherwise black. However,

in some cases, a single threshold may not be effective. Similar situation occurs in the

lung region extraction process, since in a 3D scan, middle slices may require different

threshold as compared to the starting and ending slices, because the middle slices

contain more details of the lung. So, there is a need to update the threshold or use of

multiple thresholds. The pixel intensities of the lung region in a CT image are mostly

in the range of -950 to -500 Hounsfield Unit (HU). The extent of dense area,

including bone, blood, and the chest wall, is higher than -500 HU [150]. The air

attenuation is at -1000 HU. Taking into account these points, -500HU is selected as an

initial threshold . This threshold is updated using an iterative process based on the

Eq. 3.1.

(3.1)

where is the threshold of the iteration, is the mean intensity value of the

foreground which is the lung region, and is mean intensity value of the

background, segmented in iteration on the basis of the threshold . This

42

process is repeated until an optimal threshold is computed as shown in Fig 3.3. If a

pixel at position is greater than threshold then the corresponding pixel in

binary image at the same location will be 1 otherwise 0.

Fig 3.3: Lung region extraction process

The ROI is extracted using the lung mask, which includes both vessels and nodules

that are further processed to remove vessels. Figure 3.4 shows input images, their

corresponding histogram and binary mask generated after applying the threshold.

(a) (b) (c)

Fig 3.4: Gray level distribution (a) input CT images (b) histogram of images (c)

threshold images

43

3.2.2 Nodule Candidate Selection using Polygon Approximation

One of the critical steps in lung nodule detection is candidate nodule extraction and

elimination of unwanted shapes in ROI. A more precise selection leads to better

performance of a diagnosis system. A polygon approximation method is proposed to

distinguish nodules from vessels. In this technique, a ratio between the length and

height of polygon surrounding the shape within ROI is calculated. If this ratio is more

than the threshold specified, then the shape is eliminated. Otherwise, it is selected as a

nodule candidate. Fig 3.5 illustrates the candidate selection.

(a) (b)

Fig 3.5: Nodule candidate selection (a) objects in ROI (b) marked by the polygon

In polygon approximation, the objects are distinguished by the ratio of polygon sides

as given in Eq. 3.2.

(3.2)

where represents the width and indicates the length of the polygon is enclosing

the candidate object. Algorithm 3.1 is proposed to detect the nodule candidates using

this ratio formula. The structure of vessels is rectangular whereas the nodules are

round shape objects. Therefore, the objects with a ratio closer to 1 have a high

probability of being an actual nodule, whereas objects with smaller ratio are more

likely to be a vessel. In Algorithm 3.1, the value of ratio threshold is set to 0.5, which

depicts that the objects, with a width less than half of their length, are considered as

vessels and vice versa.

44

Algorithm 3.1: Nodule Candidate Refinement

Input: Set of segmented candidate object:

Output: Refined nodule candidate set:

Begin

1: ► Set of nodule candidates

2: ► Set of vessels

3: for each ► Process each segmented object in

4: if ( ) then ► Ratio computed by Eq. 3.2

5: ► Add the object to set of vessels

6: else

7: ► Add the object to set of nodule candidates

8: end if

9: end for

End

3.2.3 Nodule Candidate Enhancement

Contrast enhancement plays a critical role in image segmentation and classification to

achieve better performance by improving the image quality. Herein, the images

containing nodule candidates are enhanced using Contrast Limited Adaptive

Histogram Equalization (CLAHE) [151]. In this approach, rather than using global

contrast enhancement, contextual contrasts are enhanced. This technique distributes

gray values using a full range of gray to improve the hidden features of the image. To

reduce the impact of noise, contrast limiting is applied to neighborhood pixels.

CLAHE is an enhanced version of Adaptive Histogram Equalization (AHE) [152].

The process of candidate nodule enhancement improves the quality of objects within

the image, thus helping in distinguishing nodules and non-nodules, and improving the

overall performance of the diagnosis method.

3.2.4 False Positive Reduction using Nodule Properties

In this step, the initial nodule candidate set is refined on the basis of nodule shape

properties. In LIDC, the size of nodule has a range of 3mm to 30mm. These objects

are removed using the above criteria to improve the set of nodule candidates, which

45

helps in reducing false positive rate. Figure 3.6 is the pictographic representation of

the process.

(a) (b)

(c) (d)

Fig 3.6: ROI and nodule segmentation (a) input CT image (b) lung mask (c) ROI with

vessels and nodules (d) nodule candidate region

3.2.5 Hybrid Intensity-Geometric and Gradient Feature Extraction

After candidate nodules detection, the next step is to detect actual nodules and non-

nodules. For this purpose, there is a need to classify nodule candidates by their

features. In the proposed method, a hybrid feature vector is created, which comprises

of geometric, intensity and HOG features, described in succeeding sections.

3.2.5.1 First Order Histogram Based Intensity Features

The intensity characteristics play a vital role in object detection. First order histogram

is used to extract intensity features that quantitatively describe statistical

characteristics of an image [153]. The features obtained in this process present a

useful characteristic of the image. One of the intensity features is a mean value, which

measures an average value of image intensity. Similarly, variance, skewness, and

kurtosis are also used to evaluate intensity distribution within the image. Skewness is

a measure that tells the symmetries of the image. Kurtosis is the flatness of the

histogram. Energy is a measurement of local homogeneity in the image that represents

46

uniformity of texture. Histogram entropy measures the randomness of the image. The

mathematical representations of these features are given from Eq. 3.3 to Eq. 3.8.

∑

(3.3)

∑

(3.4)

∑

(3.5)

∑

(3.6)

∑

(3.7)

∑

(3.8)

where is maximum gray level in the image and is the probability of a gray level

computed by the first order histogram.

3.2.5.2 Geometric Features

The geometric features are useful to describe the shape of an object within the image

[154]. In addition to intensity features, the proposed technique makes use of

geometric features, which are further subdivided into 2D and 3D features. 2D features

are based on the information extracted from a slice. Whereas, 3D features are based

on 3D scans by examining voxals. It describes the shape of objects in the nodule

candidate set. The area as a geometric feature is given by Eq. 3.9 below.

Area ∑ (3.9)

where is a median slice of the 3D scan; the median slices have more details of

internal structure as compared to other slices. Eq. 3.10 gives the mathematical

representation of the radius feature.

47

(3.10)

where D is diameter which represents the maximum length of the bounding box. Eq.

3.11 gives the 3D volume feature.

∑

(3.11)

where V represents the total number of voxels in the segmented object, circularity and

compactness are presented by Eq.3.12 and Eq.3.13 respectively.

(3.12)

(3.13)

where A and r are the area and radius respectively as defined in Eq. 3.9 and Eq. 3.10.

These are useful features for image analysis to recognize the geometric shape of

objects. In the current scenario, these features are intended to play a vital role to

distinguish nodules and non-nodules based on their geometrical structure.

3.2.5.3 Histogram of Oriented Gradients

In this method, gradient orientation in a localized portion of an image is counted

which is similar to the process of the histogram of orientation contrast of edges. It is

useful to recognize the shape of an object by computing gradient orientation in a local

region, containing the object [155]. Figure 3.7 shows nodule candidates and HOG

descriptor.

(a) (b)

Fig 3.7: Gradient orientation (a) nodule candidates (b) HOG descriptor

48

3.2.6 Support Vector Machine for Classification

After the formulation of feature vectors, SVM is used to classify the nodules. SVM is

the leading classifier in the domain of statistical learning [68, 156]. The reason behind

the use of SVM is its strength of handling high-dimensional data by using kernel

tricks. Secondly, SVM avoids overfitting by the use of regularization parameter, and

finally, it avoids local minima and is more robust in convergence to global minima.

Let { } are the examples on whom SVM has to perform

classification. The optimization function is given in Eq. 3.14 and Eq.3.15.

∑

(3.14)

(3.15)

where is weight vector and represents bias. The kernels are represented using Eq.

3.16 to Eq. 3.18.

Linear Kernel (3.16)

Polynomial Kernel

(3.17)

Radial Based Function ( ‖ ‖

) (3.18)

where and are vectors in input space, represents the degree of polynomial,

‖ ‖ is the Euclidian distance and is standard deviation, defined as Gaussian

distribution.

The proposed method is evaluated over a publicly available dataset LIDC. The results

are discussed in the next chapter.

3.3 Method II: 3D Nodule Candidate Detection Supported by

Hybrid Features to Reduce False Positives

The nodule candidate detection is a key step in intelligent lung nodule detection and

classification. In the previous method, the nodule candidate detection is performed

49

from 2D scans. Moreover, candidates are refined on the basis of the ratio of major and

minor axis lengths. In this study, the nodule candidate detection process is improved

by considering the 3D properties of nodules, in addition to 2D properties.

Furthermore, some additional parameters along with ratio are also used for better

refinement of nodule candidate set.

The proposed method performs four major tasks, as illustrated in Fig 3.8. It is

necessary to remove the background before the start of the nodule candidate detection

process. If nodule candidate detection is performed without removing the background,

then there is a chance that many objects residing outside the lung region may be

included within the nodule candidate set, which may lead to the wrong diagnosis. To

avoid this issue, optimal gray level thresholding, connected component labeling and

hole filling for the lung region extraction is applied. In the second step, a novel

method for nodule candidate detection is proposed based on Active Contour Model

(ACM) [157], 3D neighborhood connectivity and geometric properties based rules. In

this phase, some of the false positives are eliminated by checking the 3D

neighborhood connectivity and by validating the geometric properties of the candidate

objects. In the third phase, seven feature vectors are created with individual and

combinations of geometric, texture and HOG-PCA features. Finally, to reduce false

positives, classification is performed. Four different classifiers including SVM, k-NN,

Naïve Bayes and AdaBoost are used with different variations.

The major focus of this study is the reduction of false positive nodules, along with a

good score of sensitivity. The main contributions of this study are as under:

1. Lung region extraction is performed on the basis of optimal gray level

threshold based on HU values.

2. A novel hybrid technique for nodule candidate detection is proposed by

encapsulating ACM, 3D connectivity and geometric properties of nodules.

The method selects optimum candidate set, which reduces the computation

time of proceeding phases of the diagnosis process.

3. A hybrid feature vector is created for better representation of nodule

candidates. The size of each individual feature vector is also considered as a

major factor. A smaller vector may lose its significance if it is encapsulated

with a very large feature vector. To overcome this issue HOG features are

50

reduced first to make compatible in size with the other two base feature

vectors, including texture and geometric feature vectors.

4. Finally, four different classifiers including SVM, k-NN, NB, and AdaBoost

are trained and tested to distinguish actual nodules from the candidate set.

These contributions have impacted the performance of the diagnosis method.

Fig 3.8: Major steps of the proposed method

3.3.1 Lung Volume Extraction using Optimal Threshold

The initial lung mask is extracted using the optimal threshold. After that 3D

connected component labeling, morphological dilation and opening operations are

used for hole filling and boundary smoothing. To include the nodules attached to the

lung‘s boundary, also known as juxta-pleural nodules, the bidirectional chain code

[115] is applied, which results in boundary correction of the segmented mask. The

lung region is separated by applying the lung mask over the corresponding input slice,

which results in the segmented image with extracted ROI. The resultant ROI contains

the lung portion of the CT image.

51

(a) (b) (c)

Fig 3.9: Lung region extraction (a) input CT images (b) histogram of images (c) final

lung masks of the images

Fig 3.9 shows that different slices have different intensity distributions, due to which,

a single threshold cannot work for every slice. The images in Fig 3.9(a) represent the

input CT images. These images are chosen to show the different views of a scan. The

slices shown in Fig 3.9(a) are slice-number 14, 96, 122 and 142 of LIDC scan

13614193285030092. The thresholds are calculated based on the HU intensity values

52

in the images. The histograms are shown in Fig 3.9(b) illustrate the HU distribution

within each image.

3.3.2 Nodule Candidate Detection

In a diagnosis system, the candidate nodule extraction and elimination of unwanted

shapes is an important step. A more precise selection leads to better performance of a

diagnosis system. The inclusion of unwanted objects in the candidate set affects the

performance of the detection method. However, strict criteria may lead to missing the

actual nodules, resulting in the wrong diagnosis. Considering these issues, a 3D

nodule candidate detection method based on the Active Counter Model (ACM), 3D

neighborhood detection and geometric properties of nodules is proposed. Figure 3.10

illustrates the working model of the proposed candidate detection method.

Fig 3.10: Nodule candidate detection process

The ACM [157], also known as a snake, is an effective method for image

segmentation. The major advantage of ACM is that it extracts the objects from an

image with closed and smooth boundaries. The fundamental idea of ACM is to extract

the desired object from an image by energy minimization. It comprises of three types

of energies including internal energy, external energy, and constraint energy. The

relationship of a snake with these energies is shown in Eq. 3.19.

(3.19)

53

Moreover, the internal energy is comprised of elastic energy and bending energy

given from Eq. 3.20 to Eq. 3.23. This elastic energy is given by the first-order

derivative weighted by (s). Scalar controls the contribution of elastic energy using

point spacing. The bending energy contribution is controlled by the product of second

order derivate and scalar . The scalars and control the snake to identify the

object boundary, resulting in precise segmentation.

∫

(3.20)

⁄ (3.21)

where represents the snake position and is elastic energy controlling weight

along the contour.

∫

(3.22)

⁄ (3.23)

External energy takes smaller values as the boundaries of ROI and is defined as a

function of Eimage(x, y) shown in Eq. 3.24 and Eq. 3.25.

∫ ( )

(3.24)

( ) (3.25)

where is Gaussian of standard deviation , and represents the snake position.

represents the gradient operator, and represents the linear convolution. These

energies generate a flexible cover that surrounds the object and useful for its

segmentation. This technique is applied to segmented ROI containing nodules and

vessels. ACM results in detection of boundaries of the objects within ROI, which

includes vessels and nodules. It creates a large number of objects within ROI, which

can become the candidates for nodule. In addition, ACM also eliminates very small

54

objects which help in reduction of noise objects. These candidate objects are refined

on the basis of 3D connectivity and geometric properties. Neighborhood detection is

important as nodule is a mass that must have its existence in multiple adjacent slices

within a CT scan. If an object has existence in one slice, then it must have similar

intensity values in the predecessor or successor slices. Using this property, unwanted

objects are removed from the ROI. In the next step, the nodule candidate object‘s

filtration criteria are defined, by using geometric properties of the nodule. Fig 3.11(a)

shows ROI extracted from input CT images. The individual objects are extracted from

ROI using active contour shown in Fig 3.11 (b). These objects are filtered through 3D

neighborhood detection. After that, nodule candidates are extracted using Algorithm

3.2.

Algorithm 3.2: Refining Nodule Candidates

Input: Set of segmented candidate object:

Output: Refined nodule candidate set:

Begin

1: ► Set of nodule candidates

2: ► Set of vessels

3: for each ► Process each segmented object in

4: if ( ) then ► Ratio computed by Eq. 3.2

5: ► Add the object to set of vessels

6: else if ( and and

7: ► Add the object to set of nodule candidates

8: else

9: discard ► Other than vessel or nodule

10: end if

11: end if

12: end for

End

55

(a) (b) (c)

Fig 3.11: Nodule candidate detection (a) segmented ROIs, (b) corresponding contours of

the objects within the ROI’s, (c) the final selected candidates.

In Algorithm 3.2, nodule geometric properties are used, including the ratio of length

and breadth, area and circularity, to refine the nodule candidate set. The mismatched

objects are removed from the candidate set, which helps in reducing false positive

rate. The input to Algorithm 3.2 is segmented images containing lung ROI. First, the

automated contour is applied, to extract all the objects in the ROI. To extract nodule

candidates, the vessels and noise objects are eliminated from the set of segmented

objects. The ratio between the length and width of each object is computed by using

the bounding box over each object. A set is created that includes segmented objects

from ROI. Two additional sets named and are also created. Initially, and

sets are empty. Then between the length ( ) and width ( ) of each object is

computed and compared with , where stands for ratio threshold. If is

smaller than the threshold , it is considered as a vessel and not the round shaped

object. If the selected object is not a vessel, then it is checked with respect to the other

nodule properties i.e., area and circularity. Two thresholds for area and

are

defined, where stands for the lower threshold and

stands for the higher

56

threshold of the area. Furthermore, is defined as circularity threshold. If the

condition at line number six in Algorithm 3.2 is satisfied, then the object is considered

as nodule candidate; otherwise, it is considered as noise and rejected. Figure 3.11 (c)

shows the extracted nodule candidates from the corresponding images given in Fig

3.11 (a) respectively.


In the proposed technique, seven different feature vectors are created by different

combinations of geometric, texture and HOG-PCA features shown in Fig 3.12. The

texture, geometric and HOG features are used as base features. However, the HOG

features result in a large feature vector that is not compatible with the other two

features to become a part of the single vector. To overcome this issue, the PCA is

applied over the HOG features to reduce their dimensionality and make them

compatible with other features.

3.3.3.1 Texture Features

Texture features [158] are very effective in object identification and classification.

These features quantify the different properties of objects like smoothness, roughness,

bumpiness by using the intensity values of the object pixels. These features represent

the true characteristics of nodule texture and intensity in 2D, as well as 3D scan.

Taking into account these properties, these features are categorized into 2D texture

and 3D texture features. The 2D features include kurtosis, skewness, variance, mean,

entropy and energy. The 3D features include absolute value, correlation, inertia,

angular second moment, entropy, inverse difference, and maximum probability.

Angular second moment measures the uniformity of an object. Correlation measures

how much a voxel is associated with its neighboring pixels. Inertia measures the

contrast of intensity of a pixel with its neighborhood values within the candidate.

Absolute value computes the absolute intensity level in the candidate nodule. Inverse

difference measures the contrast of intensities within the object. Entropy measures the

randomness of the voxels in nodule candidates. If one combination of frequencies

dominates then maximum probability gives high value. The mathematical

representations of 3D texture features are given from Eq. 3.26 to Eq. 3.32.

57

∑ ∑[ ]

(3.26)

∑ ∑

(3.27)

∑ ∑

(3.28)

∑ ∑

(3.29)

∑ ∑

(3.30)

∑ ∑

(3.31)

(3.32)

where represents a maximum gray level intensity, represents the occurrence of

gray levels in the image. It gives the relationship between the intensities of pixels.

The are the standard deviations along horizontal and vertical axis respectively.


The most important features, to distinguish a nodule from other objects, are the

geometric features because a nodule is a round shaped object with size 3-30 mm

[159]. In addition to this property, a nodule is a solid object which has its presence in

multiple slices in a CT image. Considering these properties of nodules, 2D as well as

3D geometric features are used. The 2D features describe the shape of an object

within a 2D slice of a 3D CT scan, whereas, 3D features represent the 3D shape

properties of an object within a 3D scan [57]. The 2D geometric features include area,

radius, perimeter, and circularity. The 3D geometric features include volume,

compactness, elongation, bounding box dimensions and principal axis lengths. The

bounding box dimensions give three features, corresponding to each dimension, i.e.,

58

x, y, and z, respectively. Principal axis length consists of two features including major

and minor axis lengths. As a result, eight 3D geometric features are extracted for each

candidate object. As a whole, twelve geometric features are included in the feature

vector by combining four 2D and eight 3D features.

3.3.3.3 Histogram Oriented of Gradient and Principal Component Analysis

(HOG-PCA)

Histogram Oriented of Gradient [155] features describe the directional change in the

intensities of the image. This property helps in object detection. Gradient distribution

provides useful information in distinguishing the nodules and non-nodules. HOG is an

important gradient descriptor. During the HOG feature extraction process, histogram

of the oriented gradient is calculated based on the pixel intensity, i.e., for each pixel

gradient magnitudes and , and orientation angle are calculated.

Where, are gradient magnitudes along the x and y-axis. These histograms are

combined to form a feature vector. To differentiate the nodules and non-nodules, edge

orientation intensities are used. Combining HOG features with shape and intensity

features provides valuable information of the nodule structures. However, a HOG

descriptor is a high dimensional feature vector which increases the computation time

of the classifier. To overcome this issue, PCA is applied to the HOG. PCA is a widely

used method for feature reduction [160]. It transforms the high dimensional data into

a vector by preserving the actual information. As a result, the top seven HOG-PCA

features are selected.

3.3.3.4 Hybrid Feature Vector

One type of feature vector cannot describe all the properties of an object. Considering

this issue, three types of features are encapsulated to create a hybrid feature. Let F =

(FG,FT,FH) is a hybrid feature vector, where FG, FT, FH are geometric, texture and

HOG-PCA features respectively, and are represented by Eq. 3.33 to Eq. 3.35.

59

Fig 3.12: Feature extraction and selection

⟨ ⟩ (3.33)

⟨ ⟩ (3.34)

⟨ ⟩ (3.35)

where , and are the lengths of individual feature vectors. In this way, feature

vectors of all the images are created, which results in very high dimensional data.

There is a need to select good representative features to overcome the high

dimensionality and improve performance. The researchers used PCA, LDA and

various kinds of optimization techniques for feature reduction. The common practice

is first to create a hybrid high dimensional feature vector and then select optimum

feature set. But there is a big issue that the ratio between a number of features for each

category mismatches. In this study, 12 geometric and 13 texture features are

extracted. However, the number of HOG features is very high as compared to

geometric and texture features. If a large number of HOG features are combined with

other features for feature selection, the resultant feature vector may not contain some

60

or all useful texture and shape features. Therefore, PCA is applied to the HOG feature

vector and seven top features are extracted. These features along with other two

feature vectors are created using geometric and texture features. Seven feature vectors

are created by different combinations of geometric, texture and HOG-PCA features.

Fig 3.12 shows the working model of feature extraction and selection process.

3.3.4 Classification

The final phase of the proposed method is classification, in which four classifiers are

used, including SVM [68], k-NN [161], Naïve Bayes [162] and AdaBoost [163].

These classifiers are trained and tested over the features extracted from nodule

candidates. SVM is a popular classifier used for two-dimensional as well as

multidimensional data by the use of kernel tricks [164]. To get the advantage of this

property, SVM is trained and tested with three kernels including linear, polynomial

and RBF.

The k-NN is used as the second classifier. It is a kind of lazy learner, i.e. it does not

create a model for given data. Instead, it simply stores data. This classifier works less

in training time but extensively during testing. One of the advantages of k-NN is that

it creates a model by grouping the nearest neighbors with similar properties [165].

Here, k represents the number of neighbors involved in decision making.

Experimentation is performed with 3-NN and 5-NN. Naïve Bayes is used as the third

classifier. It is a probabilistic learning algorithm. It is driven from Bayes‘ theorem as

given in Eq. 3.36.

(3.36)

where X represents the attribute in the feature vector, is the target class. The major

advantage of this classifier is efficiency. It takes less time for training when applied to

a large dataset [166]. The fourth classifier in this study is AdaBoost. It is a kind of

ensemble classifier based on the concept of boosting. It works in a cascading fashion

i.e., make sure that it should mainly focus on those training examples which are

misclassified by the previous weak classifiers [167]. The proposed method is

evaluated over publicly available benchmark dataset LIDC. The results and

61

experimentation details are discussed in chapter 4, which shows that the proposed

method has achieved remarkable improvements in the results.

3.4 Method III: Multistage Segmentation Model and SVM-

Ensemble for Precise Lung Nodule Detection

Human lung has a complex structure, due to which it becomes a challenging task for

the radiologists to identify the diseased region. This complex structure also remains a

challenge for automated lung cancer diagnosis methods. Therefore, a simple

segmentation method cannot produce good results. Considering these points, a

multistage segmentation method is presented in this section to improve the diagnosis

procedure. Fig 3.13 shows a workflow diagram of the proposed method.

Segmentation of lung nodules is a challenging phase in the detection process due to

the complex structure of the lung. Furthermore, handling of small, boundary attached

(juxta-pleural) and vessel attached (juxta-vascular) nodules is a big challenge. The

proposed technique addresses these issues in order to give a better detection method

for lung cancer. The lung region is extracted by applying corner seeded region

growing, combined with optimal thresholding. In addition to this, morphological

operations are applied in boundary smoothing, hole filling, and juxta-vascular nodule

extraction. In addition to segmentation, the selection of relevant features and a robust

classifier helps to improve accuracy and reduce false positives. Considering the

aforementioned geometric properties, along with 3D edge information are applied to

extract nodule candidates. Geometric Texture Features Descriptor (GTFD), followed

by Support Vector Machine based ensemble (SVM-ensemble) classification is

employed to identify actual nodules from the candidate set. SVM-ensemble classifier

is applied, which is based on three base classifiers, namely, SVM-linear, SVM-

polynomial and SVM-RBF (Radial Based Function). A publicly available dataset

LIDC-IDRI is used to evaluate the performance of the proposed method.

The main contributions of the proposed method are:

1. A multistage segmentation model is proposed for precise lung region extraction

and nodule candidate detection.

2. Differential evolution is used for threshold selection.

3. 3D morphological operation is used to separate juxta-vascular nodules from the

vessels.

62

4. Geometric Texture Feature Descriptor (GTFD) is used for the better

representation of nodule candidates.

5. SVM-ensemble classifier is proposed for classification of nodules and non-

nodules.

Fig 3.13: Workflow of the proposed method

3.4.1 Nodule Segmentation

Pulmonary nodule segmentation is an essential phase in the performance of an

automatic diagnostic system. During the diagnosis procedure, precise and accurate

segmentation of nodules impacts the performance. Segmentation of smaller nodules,

ranging from 3-30mm, juxta-pleural nodules, and vascular attached nodules is a

significant test for the segmentation method.

3.4.1.1 Lung Region Extraction

In the first step, the image background is removed by using corner seeded region

growing combined with thresholding approach. The thresholding is a straightforward

63

and efficient method for image segmentation. It performs segmentation on the basis of

pixel intensities. It is observed that simple thresholding is not useful for lung region

extraction due to overlapping between the intensities of background and some

sections of ROI. To overcome this issue, the background of input images is removed.

Furthermore, the threshold generated for one slice is not useful to other slices because

of its variation in the gray level for each slice as shown in Fig 3.14. Therefore, the

suggested segmentation technique used a combination of differential evolution based

optimal thresholding [168] and corner seeded region growing.

After the removal of image background, differential evolution based optimal

thresholding is used to identify the lung boundary and extraction of the lung region.

An initial threshold of -500HU is applied because most of the lung region lies in the

range of -950HU to -500HU. It is an iterative step in which the threshold is

recalculated in each iteration.

In an image, a histogram gives the probability distribution of gray levels. This

probability distribution is calculated by Eq. 3.37.

∑ ∑

√

[

]

(3.37)

The total number of categories within the image is represented by . and are

the probability and probability distribution functions within the category ,

respectively. and are the mean and standard deviation. To compute optimal

threshold, the overall probability error in two different categories is minimized by Eq.

3.38.

∫ ∫

(3.38)

This error is with respect to the threshold . The overall error is calculated as given

by Eq. 3.39.

64

∑

(3.39)

Henceforth, a threshold image is generated that contains the lung mask.

Morphological operations are applied to refine the lung boundary and extraction of

juxta-pleural nodules. Morphological closing is performed on the binary mask, where

a disk of size 25 pixels is taken as the structuring element. The process of closing also

includes some extra regions which can affect the accuracy of segmentation. To solve

this problem, morphological opening with the same structuring element, a disk of size

25 pixels, is applied. Fig 3.15 highlights the process of juxta-pleural nodule

extraction. Finally, the lung region is extracted by using the resultant mask and input

CT image. The whole procedure is repeated for all the slices of each scan. Fig 3.16

shows the step by step results of segmentation.

(a) (b)

Fig 3.14: Gray level distribution (a) input CT images (b) their respective histograms

65

(a) (b) © (d)

Fig 3.15: Juxta-pleural nodule extraction (a) input images with juxta-pleural nodules,

(b) corresponding lung masks missing pleural nodules, (c) boundary correction using

morphology, (d) extracted lungs with juxta-pleural nodules

66

(a) (b)

(c) (d)

Fig 3.16: Lung region extraction (a) input image, (b) background removed, (c) lung

mask, (d) extracted lung

3.4.1.2 Nodule Candidate Detection

After ROI extraction, the next step is to extract nodule candidates by eliminating

vessels and other unwanted objects. For this purpose, a combination of morphological

operations, edge detection, bounding box and shape information are applied to the

lung ROI for eliminating vessels and noise objects. The major issue in nodule

candidate detection is juxta-vascular nodules due to their attachment to the vessels

[169]. In most of the previously known techniques, these nodules are eliminated with

the removal of vessels. To avoid nodule elimination, 3D connectivity information is

used. As discussed earlier that a nodule is a 3D object, therefore, it must show its

existence in consecutive slices of a scan. Considering this property of nodule, the

presence of each object with its predecessor and successor slices is investigated. Let

is a candidate for a nodule pixel, where is the slice number and ,

represent the location of the pixel within that slice. Then 3D neighborhood is checked

by comparing the intensity of that pixel. Each pixel of slice has nine

neighborhoods in the slice and nine in the slice as shown in Fig 3.17. The

67

similar intensity pixels must be there, within the eighteen neighborhood pixels, either

in predecessor or in successor slices or in both.

Fig 3.17: 3D connectivity of candidate pixels

The morphological opening is used to separate juxta-vascular nodules from vessels.

The structuring element used for this purpose is a ‗sphere‘ of size 2 pixel radius,

which is effective in 3D lung structure. The nodule is a round shaped object, and the

use of a ‗sphere‘ as structuring element preserves the nodules and breaks the

connection of nodules and vessels. Afterwards, 3D connected component analysis is

performed, and regions with centers within 5mm are merged. Canny edge detection

method is used to identify the boundaries of each object [170]. It results in a large

number of candidate objects. To refine this set of objects, the nodule size and shape

information is taken into account. The nodule is a round shaped object that ranges

from 3mm to 30mm in size within the ROI. Considering these properties, the nodule

candidate set is defined on the basis of diameter, elongation, and circularity. The

objects with a diameter less than 3mm and greater than 30mm are removed and

considered to be unwanted objects as their sizes do not match with the size

specifications of nodules. In addition to diameter, those objects are also excluded,

which are too elongated, i.e., their major axis length is 2.25 times more than the minor

axis length. The third parameter is circularity, which defines the roundness of the

object. The roundness of an object ranges from 0 to 1, where 1 represents perfect

round. All those objects which have circularity less than 0.65 are also excluded from

the candidate set. The elongation and circularity threshold values are empirically

selected based on the observation that if circularity value is increased more than 0.65

and elongation value is decreased less than 2.25, this causes the removal of some of

the actual nodules. On these parameters, the nodule candidate set is refined, and

68

nodules are distinguished from other objects like vessels and noise objects as

presented in Fig 3.18. Some of the extracted nodule candidates are illustrated in Fig

3.19.

(a) (b)

Fig 3.18: Nodule candidate refinement (a) objects in ROI (b) nodule candidate regions

Fig 3.19: Extracted nodule candidates

3.4.2 Geometric Texture Feature Descriptor (GTFD)

In object detection and classification, precise and relevant features play a significant

role. In recent studies, a variety of features including geometric, texture, Local Binary

Patterns (LBP), wavelet, intensity, and gradient are used for the detection and

classification of nodules because a single type of feature vector cannot give an

accurate representation of objects, resulting in unsatisfactory performance. To

overcome these issues, researchers also used hybrid feature vectors by encapsulating

more types of features in a single feature vector. In this study, a GTFD by combining

texture and shape features in 2D and 3D is devised that successfully delineates the

nodules and non-nodules.

69


Geometric features like circularity, area, and volume are the shape descriptors of an

object [171]. Therefore, these features are helpful to distinguish nodules from other

objects within the lung region. The median slice of a scan contains maximum

information of objects present in the scan. Therefore, 2D geometric features are

extracted from the median slices. However, 3D geometric features are extracted from

3D segmented objects. The features include area, radius, circularity, volume,

compactness, and elongation.


Texture features characterize the regions of images on the basis of their pixel

intensities [158]. Ten texture features, four 2D including mean, variance, skewness,

energy, and six 3D features including angular second movement (energy), contrast,

correlation, homogeneity, and entropy, are used in this study.

3.4.3 Ensemble Classification

SVM-ensemble classification is used to identify the actual nodules from a set of

candidate nodules. An ensemble classifier makes a final decision on the test dataset

derived from the output of individually trained base classifiers [172]. Let there be

base classifiers as: { } over a dataset and target . Each classifier is

trained and evaluated over the same test examples. The decisions of individual

classifiers are combined to make a final decision. Majority voting is a useful

combiner, in which decision about the class of a test example is made through the

number of votes, determined by Eq. 3.40.

∑

(3.40)

where assigns class labels to the test feature vector on the basis of maximum

votes given to a class and represents the total number of base classifiers. Let if there

are three base classifiers; two predict that the test example is of class A and one

predicts it in class B, then majority voting predicts it as class A. In this study, SVM is

used as a base classifier with three different kernels including linear, polynomial and

Gaussian RBF.

70

Fig 3.20: Workflow of SVM-ensemble for nodule classification

SVM is known to be a good classifier that learns easily and produces good

classification results. However, a single classifier may not learn all the parameters to

produce good global optimum result. To meet this limitation, an SVM based ensemble

classifier is developed for nodule classification. Fig 3.20 shows the workflow of

SVM-ensemble classifier for the proposed method to distinguish nodules and non-

nodules. During the training phase, each individual SVM based classifier is trained

over the labeled data provided for training. SVM-linear, SVM-polynomial, and SVM-

RBF are trained as base classifiers. The individual classifiers are aggregated on the

basis of voting. In the testing phase, an unlabeled test example is supplied as input to

all the individual classifiers simultaneously, and a final decision is made on the basis

of Eq. 3.40. The proposed method is evaluated over the largest publicly available

dataset LIDC-IDRI. The experimentation and results are discussed in the next chapter.


Geometric Fit in Parametric Form and Deep Learning

Although, the most recent studies have presented various lung nodule detection and

classification methods to improve the performance of the diagnosis process, but to

overcome the number of false positives remains a challenge. In addition,

71

improvement in one stage of diagnosis method may not assert a strong impact on the

performance. In this study, these points are addressed, and the main focus is to reduce

the number of false positives by contributing at every stage of the diagnosis process.

The proposed lung nodule detection method to reduce false positives is illustrated in

Fig 3.21 consisting of four major steps. First, the optimal threshold is computed on

the basis of Fractional Order Darwinian Particle Swarm Optimization (FODPSO)

[173], which produces an initial binary lung mask. To improve the structure of this

mask for accurate segmentation, hole-filling and boundary smoothing are performed

by applying morphological operations. The lung region segmentation is performed on

the basis of this mask. After segmentation, a novel method based on the geometric fit

in parametric form [174], is proposed to detect circular shapes that can be a part of the

nodule candidate set. These candidates are refined on the basis of geometry and shape

properties of the nodules. In the third step, feature extraction and fusion are

performed. The geometric and texture features are extracted and fused to a hybrid

feature vector. A CT scan contains 3D information of lung along with the 2D

information. Taking into account this property, 3D texture and geometric features are

also extracted, along with 2D features. As a result, a hybrid feature vector is created.

In the fourth and final step, feature reduction and classification are performed. A two-

level stacked autoencoder (SA) is designed for feature reduction, and the softmax

activation function is trained to distinguish nodules and non-nodules.

The performance of the proposed method is evaluated over LIDC-IDRI dataset. The

performance factors, including sensitivity, specificity, accuracy, and number of

FPs/scan remain the primary focus of this study.

The main contribution of the proposed algorithm is given under:

1. FODPSO based optimal threshold to segment the lungs volume.

2. Proposed a novel method for nodule candidate detection using geometric fit in

parametric form by using geometric properties of nodules. Nodules pruning is

performed by using geometric information based rules.

3. Proposed a Hybrid Geometric Texture Feature (HGTF) descriptor for better

representation of nodule candidates, which improves the classification.

4. Deep learning for feature reduction and classification using stacked

autoencoder and softmax.

72

Fig 3.21: Block diagram of the proposed method

3.5.1 Lung Volume Segmentation

The extraction of lung region from CT images is a much needed step before nodule

candidate detection. The segmentation separates the lung cavity and the surroundings,

including background and lung anatomy [175]. This removes the objects residing

outside the lung area that impedes the objects, residing outside the lung region, to

become part of the candidate set. The lung volume segmentation method, as shown in

Fig 3.22, consists of the following steps:

1. Initial lung mask extraction by applying FODPSO based optimal threshold.

2. Lung region extraction based on 3D connected component labeling.

3. Final lung mask generation by refining the lung boundary.

A lung CT scan can be categorized into two major regions; low intensity region and

high intensity region. These regions represent different anatomical structures. The

surrounding of the lung cavity lies in the high-density region, whereas the lung cavity

and air around the body belong to the low-density region. Before nodule candidate

detection, it is required to separate the lung volume from low-density regions because

the Region of Interest (ROI) is the lung volume. The major problem with CT data is

that all slices of a single patient contain a different portion of the lung; for example

initial slices contain only a small part of the lung, middle slices contain major lung

parts, and final slices also contain some part of the lung. Therefore, it is not possible

to have a single threshold to segment lung from all slices of a patient.

The lung tissues are usually in the range of -950 to -500 Hounsfield Units (HU) in CT

image, while the extent of dense area, including bone, blood, and the chest wall, is

• LIDC/IDRI Dataset

Input CT Images

• FODPSO Thresholding

• Hole Filling

• Boundary Smoothing

Lung Volume Segmentation

• Geometric Fit in Parametric Form

• Geometric Rule Base Refinement

Nodule Candidate Detection

• Geometric 2D/3D

• Texture

• Hybrid

Feature Extraction

• Autoencoder + Softmax based Deep Learning

Classification

73

higher than -500 HU. On the other hand, air attenuation is at -1000 HU. Taking into

account these points, -500 HU is selected as an initial threshold. However, there is an

issue that middle slices of a 3D scan cannot be handled with this threshold as

compared to initial and last slices. A fixed threshold may not work well for all slices

due to this reason. Therefore, it is required to find out an optimal threshold to handle

the different types of slices. The most effective method to obtain the optimal threshold

is the one that maximizes the variance between the classes [176, 177] and can be

calculated using Eq. 3.41.

(3.41)

where and are the probabilities of foreground and background respectively. ,

and represents the mean gray level intensities of foreground, background and

complete image. An optimization method can be a useful tool in optimal threshold

selection. Considering this point, FODPSO is applied to find an optimal threshold that

can handle all types of slices.

Fig 3.22: Proposed lungs segmentation method

3.5.1.1 Fractional-Order Darwinian PSO based Optimal Threshold

Several variations of Particle Swarm Optimization (PSO) are introduced in recent

years, including Darwinian Particle Swarm Optimization (DPSO) [178], Fractional

Order Particle Swarm Optimization (FOPSO) [179] and FODPSO [173, 180]. The

FODPSO is based on Darwin‘s evolutionary theory. It is an extension of PSO but

with improved balance between exploration and exploitation with the help of

fractional order extension of FODPSO. The particles in FODPSO are known as

intelligent swarms interacting with each other and with the environment to achieve a

common goal. Using these principles, FODPSO overcomes the weakness of PSO,

74

which get stuck in local optimization. The FODPSO overcomes the PSO weakness of

convergence. It controls convergence of swarm towards the solution. The FODPSO

executes several PSO algorithms in parallel on the same test problem. When a swarm

is stuck in local optima then search in that area is stopped, and the swarm is moved to

another area for the search. The swarms, which show better performance, are

rewarded, and the swarms with poor performance are punished. The algorithm is

initialized by a population of randomly created particles. The optimization process is

performed iteratively on the basis of pBest and gBest. Each swarm moves in search

space to a position , within the image intensity levels from to and

velocity . The practical best

and global best

information is used to update

the position and velocity of a particle and computed by Eq. 3.42 and Eq. 3.43.

( )( )

(3.42)

(3.43)

The fractional order controls the position and velocity by Grünwald–Letnikov

fractional derivative. The and are the weights that control local and global

performance. The and are random multipliers ranging from 0 to 1 and is the

time. On the basis of these equations, an optimum solution is computed. In the

threshold selection process, FODPSO particles are initialized with zero velocity and

positions are assigned within the gray levels. Each pixel is assigned a gray level

ranging from to , where is the lowest and is the maximum possible value

of a gray level. The total number of pixels for each gray level can be computed using

Eq. 3.44.

∑

∑

(3.44)

where is pixel count of the image and represents the number of gray levels. The

number of pixels in each category is computed by histogram .

75

The probability distribution is calculated by Eq. 3.45.

∑ ∑

√

[

]

(3.45)

where an image contains number of categories. The is probability,

probability distribution, is mean and is variance of category . To compute the

optimal threshold, the overall probability error is minimized by the fitness function.

The threshold image is a binary image created using Eq. 3.46, on the basis of

final optimum threshold.

{

(3.46)

where, is the source image, represents maximum gray level in the image

and is optimum threshold. The problem space is explored by each particle for the

candidate solution. Eq. 3.47 gives the mathematical representation of particles, which

is used to construct the mixture of Gaussian probability function.

(3.47)

where and represent the mean and variance of particle within the search

space. The possible variance for each particle is computed and compared with other

particles within a swarm. The particle which achieves a high class variance is

considered as best performer and other particles move towards its direction. The

fractional extension improves exploitation behavior, which leads to an improved

convergence.

3.5.1.2 Fitness Function

The optimization is an iterative process that continues its execution until the fitness

criteria are achieved. In image segmentation, the goal of optimization is to minimize

the error between the original histogram and Gaussian functions shown in Eq. 3.48.

76

∑ ( ) ( )

|(∑

) |

(3.48)

The total number of gray levels is represented by . The ( ) and ( ) represents

the probability and histogram of gray level respectively. The represents the

number of categories within the image. is the probability within the category and

is penalty associated with the probability of a category within the image. The

optimal threshold is computed through an iterative process, which continues till the

stopping criteria are fulfilled.

The initial segmented lung mask is generated using the optimal threshold computed

by FODPSO. This segmented mask contains the body and non-body voxels. To refine

this mask and extract the body voxels surrounded by non-body voxels, 3D connected

component labeling is applied. The 18 neighborhood pixels are used from predecessor

and successor slices of the 3D scan. The two largest regions are selected as the lung

body regions. Then hole filling is performed using morphological operations.

Fig 3.23 shows the segmentation results, which illustrates that FODPSO based

thresholding method worked well for lung volume extraction. The extracted lung

masks may not include the juxta-pleural nodules as shown in Fig 3.24(b), which affect

the diagnosis performance. The eight directional chain code analysis [181] is

performed for boundary correction, given in Algorithm 3.3, which results in juxta-

pleural nodule inclusion, as shown in Fig 3.24(c). The distance between the critical

points is calculated and compared with a nodule size and merged with those sections

which have distance less than nodule size. Fig 3.24(d) shows the final segmented

regions with juxta-pleural nodules.

77

(a) (b) (c)

Fig 3.23: Lung region extraction (a) input CT images (b) threshold images (c) extracted

lung regions.

78

(a) (b) (c) (d)

Fig 3.24: Juxta-pleural nodule extraction (a) Images with pleura nodule, (b) initial lung

mask, (c) after boundary correction, (d) final ROI

Algorithm 3.3: Boundary Correction Algorithm

Input: Initial lung mask:

Output: Final lung mask after boundary correction:

Begin

1: = ► Compute boundary mask

2: ► Compute the gradient

3: ►Identify critical points on the basis of gradient

4: for each ►Process each pair of critical points

5: ► Segment length between x,y

6:

7:

8: if ►Check boundary segment is concave

9: if ► Compare with nodule size

10: = ► Update path from x to y

11: end if

12: end if

13: end for

14: ► Update lung mask

End

79

3.5.2 Nodule Candidate Detection using Geometric Fit in Parametric

Form

The nodule candidate detection has prime importance in the automated diagnosis

process. This step requires the consideration of key properties of nodule candidates to

detect them accurately. The main target of this stage is to perform initial candidate

detection with 100% sensitivity, which reflects the inclusion of all nodules in the

candidate set. However, a major issue of this step is the inclusion of a large number of

non-nodules in the candidate set, which leads to high false positives. The higher

sensitivity with low false positives depicts the strength of a detection method. A novel

nodule candidate detection method, based on geometric fit in parametric form, is

proposed to address the aforementioned issues.

As discussed earlier, the nodules are round shape circular objects of different sizes

ranging from 3mm to 30mm [64]. Furthermore, in a CT scan, nodules appear to be the

brighter objects with maximum intensity at the center and mostly within dark

surroundings. Considering these points, the nodule detection strategy is developed, on

the basis of circular region detection using geometric fit in parametric form, by

considering the maximum intensity point as the centroid of the circle. The circle in

parametric form is represented by Eq. 3.49 and Eq. 3.50.

(3.49)

(3.50)

The geometric fit refers to the circles for which the sum of square distance from a

given points is minimal. Consider two points and , the distance between these

points is calculated as given in Eq. 3.51.

(3.51)

The minimum distance problem is solved using the Newton method, to minimize the

algebraic distance, which is expedient to get good initial values of , and . Then

minimum quadratic function is computed by minimizing , , and . The

Jacobian describes how this kind of change is transformed. To find area of the

80

circle, the Cartesian coordinates are changed to polar coordinates, which results in the

following transformations given in Eq. 5.52 and Eq. 3.53.

∬

(3.52)

The resultant Jacobian is computed as given by Eq. 3.53.

|

| |

| (3.53)

The geometric fit is used to detect the nodule candidates from the segmented ROI of

each slice. The radius value from 1.5mm to 15mm is applied, to detect all sizes of

nodules. The contrast between each object, with center at and radius and its

surroundings is computed. The is computed using geometric fit in parametric

form. The object for which the sum of the squared distance to a reference point is

minimal is considered as the best fit. An object is considered best fit or geometric fit,

if its distance from the reference point, given in annotation, is less than 1.5 times its

radius [159]. In addition, a nodule has its presence in multiple slices of a scan;

therefore, in addition to the area, there is a need to calculate the volume as well. The

Jacobian for the 3D object is applied and is given in Eq. 3.54.

∭

∭ |

|

(3.54)

The , and are the coordinates of the actual object in scan, and , and are the

coordinates of the reference object within the annotation. The Gauss-Newton method

is an iterative process which produces the optimum values of , and . In this

process, all the circular objects are processed and checked over the criteria to become

a nodule candidate. For further refinement of the nodule candidate set, geometric rule

base refinement is performed. These properties are area, circularity, radius, perimeter

and elongation.

81

Table 3.1: Threshold values for nodule candidate set refinement

Parameter Reference value

Radius From 1.5mm to 15 mm

Area 7.0 to 706.9

Perimeter 9.42mm to 94.25mm

Volume 14.13 to 141337

Circularity Greater than 0.65mm

Elongation Less than 2.25mm

Table 3.1 gives the reference values for the parameters used for the refinement of the

nodule candidate set. The objects with a radius less than 1.5mm and greater than

15mm are eliminated as their sizes do not match with the actual nodule properties.

The objects having area from 7.0 to 706.9 and perimeter from 9.42mm to

94.25mm are selected as nodule candidates. The circularity measures the roundness of

the objects, ranging from 0 to 1. The objects with circularity 1 are considered as

perfectly round. The objects in the candidate set with circularity less 0.65 are

excluded due to low roundness. The elongation is the ratio between major and minor

axis lengths. The objects having elongation greater than 2.25 are excluded from

candidate set because these are too elongated and considered as vessels. The threshold

values of circularity and elongation are selected experimentally. If the circularity

value is increased from 0.65 and the elongation value is decreased from 2.25, then

some of the actual nodules are also excluded resulting in low sensitivity. The nodule

candidates‘ regions are shown in Fig 3.25, and some of the extracted candidates are

shown in Fig 3.26.

82

(a) (b)

Fig 3.25: Nodule candidate refinement (a) segmented ROIs, (b) the final candidate

regions

(a) (b)

Fig 3.26: Nodule candidates (a) nodules (b) non-nodules


Image features play a key role in the classification process. A precise feature set can

describe an image more accurately [84]. A feature vector is a numerical representation

of an image or object within the image. Similarly, a single type of feature may not be

a true representative of an object. To overcome this issue, there is a need to create a

feature vector that incorporates more than one type of features [104]. This type of

feature vector is known as a hybrid feature vector. In the proposed method, each

83

feature vector incorporates 44 features, including 32 texture and 12 geometric

features. Besides the advantage of combining multiple features, there is a limitation

that it increases the dimensionality of a feature vector [182]. A high dimensional

feature vector increases the classifier training time. Feature reduction helps to

overcome this issue. A precise reduction curtails the feature vector without losing

important information. In the proposed method, a stacked autoencoder based deep

learning network is developed for the feature dimensionality reduction discussed in

section 3.5.4


The texture features characterize the image regions by their pixel intensities [183].

For texture analysis, Gray Tone Spatial Dependence Matrices (GTSDM) [184] is

computed, based on different angular ratios and distances among the neighborhood

pixels in 2D as well as 3D. These features also incorporate the spatial information

along with the texture, which includes the distance and angular relationships between

a nodule candidate and its surroundings [185]. At the first step, GTSDM matrix is

computed, and features are extracted by using this matrix. For 2D features, eight

neighborhoods are considered with angular information on 0, 45, 90 and 135 degrees

[186]. In a 3D space, a voxel has 26 neighborhoods, nine in each predecessor and

successor slices. Considering the spatial information along with their intensity levels,

the 3D texture features are extracted. Fig 3.27 (b) shows the 3D neighborhood of the

central voxel.

(a) (b)

Fig 3.27: Texture special information (a) pixels with neighborhood directions

(b) 3D neighborhoods

84

The 2D texture features, including variance, skewness, Kurtosis, and energy are

computed in eight directions, then mean and range for each feature value is computed.

In addition to 2D features, six 3D features, including angular second movement,

correlation, inertia, absolute value, inverse difference and entropy are computed.

Suppose is a feature matrix, is the cell of that matrix containing the feature

value in direction. For 2D features, four feature values in eight directions are

computed. Then mean and range of each feature value is computed, which results in a

1x8 sized 2D feature vector. The association of the voxels in 3D space is different as

compared to the 2D plane which requires that rather than computing features in all 26

neighborhoods, try different combinations for sampling the data. Considering these

points, six 3D features in 18 and 26 neighbors are computed. After that, mean and

range for each feature are computed same, as 2D features, which results in a 1x24

feature vector. Finally, 32 texture features are extracted including eight 2D and twenty

four 3D features.


The second group of features is the geometric features that describe the shape

information of the nodules. These are computed from the segmented candidate

objects. It is a very useful descriptor to distinguish true positives from false positives.

To distinguish nodules from other objects, it is desirable to use shape information.

The geometric shape descriptor includes 2D as well as 3D geometric features [187].

The 2D geometric features include area, radius, perimeter, and circularity. In addition

to 2D features, 3D features are also very important for nodule classification because a

nodule is a 3D object which has its existence in multiple adjacent slices. The 3D

geometric features include volume, compactness, elongation, bounding box

dimensions and principal axis lengths. The principal axis length includes two features,

i.e. major and minor axis length. The bounding box dimensions include three features

in each dimension x, y, z-axis, which results in eight 3D geometric features. As a

result, 12 geometric features are extracted.

3.5.3.3 Feature Fusion to Create Hybrid Feature Vector

Feature fusion is an encapsulation of different types of feature vectors into a single

feature vector [188]. The feature fusion of multiple features improves the

classification accuracy because different objects within a dataset have different

85

properties with respect to the feature types [189]. In the proposed method, two level

serial fusion is applied, as shown in Fig 3.28. At the first level, 2D and 3D geometric

and texture features are fused, which results in geometric and texture feature vectors.

On the second level of fusion, these texture and geometric feature vectors are fused to

form a hybrid geometric texture feature (HGTF). The fusion of these individual

vectors is carried out on serial basis, to form a hybrid vector based on Eq. 3.55 to Eq.

3.57, by fusion of these vectors.

(3.55)

(3.56)

(3.57)

where, and are fused geometric and texture feature vectors respectively, and

is the final hybrid feature vector. represents the serial fusion.

Fig 3.28: Feature fusion for the hybrid feature vector

3.5.4 Stacked Sparse Autoencoder and Softmax Based Feature

Optimization and Classification

Classification is the final step of the diagnosis process which distinguishes nodules

from other objects. The classifier is trained on labeled examples and in the testing

phase, unlabeled data is provided to a trained classifier, which classifies it to a

relevant class. Feature reduction is a process that improves the performance of a

classifier by selecting more appropriate feature subset from a large feature set. The

commonly used methods for feature reduction are principal component analysis

(PCA), genetic optimization and PSO based optimization [105]. Deep learning is an

important area of research that provides a platform for different classification

86

problems. The key feature of deep learning is to create a large neural network that

improves the accuracy of classification and prediction problems. An autoencoder is a

kind of deep neural network which has an important role in the feature reduction and

optimization problems. It is an unsupervised three layer network with same sized

input and output layers. Its transformation of input to the hidden layer is known as

encoding, and from hidden to the output layer is called decoding [190]. It maps the

high dimensional feature vector to the lower dimensional feature vector.

However, a single autoencoder may not be very accurate. To overcome this

shortcoming, a deep neural network is proposed including two staged stacked

autoencoder for feature reduction and softmax for classification. A deep autoencoder

takes a high dimensional input feature space and then reduce it by a series of smaller

hidden layers. Reduction of data to lower dimension improves the performance of a

classifier. Autoencoder maps input feature vector to hidden layer

where k<n. That hidden layer is a new feature representation. When the size of the

hidden layer is less than the input layer then autoencoder maps high dimensional input

feature vector to low dimensional feature vector. These features are then used as input

to the second hidden layer of autoencoder. In this way a two-level stacked

autoencoder performs feature optimization. The network is trained with scaled

conjugate gradient [191]. The Eq. 3.58 and Eq. 3.59 give the update rules and

activation functions for the training of network.

(3.58)

where,

is the bias of the layer, is the learning rate and is the additional

learning rate.

(∑

) (3.59)

where

is the neuron of the layer,

are the synaptic weights, is the

input to the neuron and is the bias. Fig 3.29 shows the network structure of the

proposed deep net. During the training of network, each input feature is passed to

all the hidden layers in a feedforward manner and then to the output layer to compute

87

. The error is computed, which gives the deviation from input to output. This error

is backpropagated to perform weight updation. The first layer receives HGTF feature

vector as an input and passes it to the first autoencoder, which encodes these features

and produces a new feature vector. This new feature vector is passed to the second

autoencoder as an input following the same procedure. In this way, a stacked

autoencoder is constructed. Finally, the low dimensional feature vector is given to the

softmax classifier to distinguish nodules and non-nodules. In contrast to autoencoder,

softmax is based on supervised learning. The softmax is a probability distribution

function that maps a K-dimensional vector of arbitrary values to the K-dimensional

real values computed by Eq. 3.60.

∑

(3.60)

where ∑ works as a regularization parameter.

Fig 3.29: Stacked autoencoder for feature optimization

88

3.6 Summary

This chapter presents four proposed methods to detect and classify lung nodules,

which provide the second opinion to the radiologists during the diagnosis process.

The primary focus of these methods is to reduce the number of false positives along

with high sensitivity. The precise segmentation of the lung region and nodule

candidates improves the detection process. Similarly, precise features and robust

classifier identify the actual nodules from the candidate set. In the proposed methods,

these points are considered. Lung region is extracted with the target that all the juxta-

pleural nodules must become part of ROI. In the nodule segmentation process, precise

segmentation of juxta-vascular nodules remains the primary focus. To achieve these

targets 3D image information is also considered along with the 2D information. For

classification of nodules and non-nodules, different types of 2D, as well as 3D

features including shape, texture, intensity and gradient features, are extracted.

Finally, instead of relying on state of the art classifiers, ensemble classifiers, and deep

learning based feature reduction and classification are techniques proposed. The

experimental setup, implementation details, and the results of these methods are

discussed in the next chapter.

89

Chapter 4

Results and Discussion

90

4.1 Introduction

In this chapter, the performance of the proposed methods is evaluated. The following

subsections contain a detailed discussion of the experimental setup, dataset

description, and the annotation details, quantification measures, experiments, and

comparison of proposed methods against the methods reported in the literature. The

experimentation statistics are illustrated in Fig 4.1, including the dataset details, data

distribution, and performance measures used to evaluate the significance of the

proposed methods.

Fig 4.1: Experimentation statistics of all proposed methods

91


and Hybrid Features from Lung CT Images

In this study, an authentic publicly available dataset LIDC is used to evaluate the

performance of the proposed method. It contains 3D scans, each containing more than

100 slices with slice thickness between 1.25mm to 3mm and dimension of 512×512

pixels. Each image has a 12-bit grayscale resolution with the pixel size ranging from

0.5mm to 0.8mm. Moreover, along with the image data, another repository of XML

files is also provided for the convenience of researchers to get details about images

available in the dataset. In the present case, all LIDC scans are used in

experimentation. Further, the classification of nodules is performed using linear, RBF

and polynomial kernels. The performance of classifiers is evaluated over different

distributions of data into training and testing ratio.

4.2.1 Experimentation and Results

MATLAB 2015a is used as an implementation tool. The results using SVM with

different kernels are shown in the tables from Table 4.1 to Table 4.3. The Linear

kernel is applied with the division of dataset in training and testing with the ratios of

20-80, 30-70, 50-50 and 30-70. In the case of linear kernel, it can be seen from Table

4.1 that the proposed technique produced 83.4% accuracy but with a very high false

positive rate.

The second kernel used is polynomial. The proposed technique is evaluated with

polynomials of degree 2, 3, 4, 5 and 6, respectively, with the same distribution of test

and training examples used in the linear kernel. It can be observed from Table 4.2 that

the best results produced by the polynomial kernel of degree 5 are on 70-30

distribution for training and testing, respectively.

Table 4.1: Results produced by SVM linear kernel

Training-Testing Accuracy(%) Sensitivity(%) Specificity(%) FPs/Scan AUC

80-20 83.4 85.6 74.3 25.7 0.82

70-30 82.1 85.2 78.7 22.3 0.81

50-50 78.2 75.4 71.4 28.6 0.78

92

Table 4.2: Results produced by SVM polynomial kernel


80-20 94.1 95.3 93.2 6.8 0.97

70-30 98.8 97.7 96.2 3.8 0.99

50-50 91.2 89.6 87.5 12.5 0.94

The Gaussian RBF kernel is applied to train and test the SVM classifier. The

accuracy, sensitivity, specificity, and FPs/scan are computed. The RBF kernel

produced the best results, as reported in Table 4.3, with 70-30 distribution of training

and test data, respectively.

Table 4.3: Results produced by SVM-RBF


80-20 96.3 96.0 95.8 4.9 0.98

70-30 98.8 97.8 97.2 3.7 0.99

50-50 95.2 89.3 95.3 6.1 0.98

Table 4.4 shows the best results obtained by all three kernels used in the proposed

technique. The linear kernel produced the worst performance with 83.4% accuracy

and 85.6% sensitivity with an FPs/scan of 25.7. It is due to the limitation of the linear

kernel in handling the high dimensional data. On the other hand, the RBF kernel has

given an excellent performance with an accuracy of 98.8% and 97.8% sensitivity with

very low FPs/scan of 3.7. Here it is clear that SVM-RBF kernel produces best results.

Figure 4.2 shows the comparison of classification results graphically.

Table 4.4: Best results by SVM classifier with different kernels

Kernel Training-Testing Accuracy(%) Sensitivity(%) Specificity(%) FPs/Scan AUC

Linear 70-30 83.4 85.6 74.3 25.7 0.82

Polynomial 70-30 98.8 97.7 96.2 3.8 0.99

RBF 70-30 98.8 97.8 97.2 3.7 0.99

93

Fig 4.2: Comparison charts of classification results

4.2.2 Comparison of Proposed Technique with Existing Techniques

The most important factor in performance evaluation is to compare the results with

existing methods in the same domain. It is a challenging task due to the heterogeneity

in evaluation parameters like dataset, nodule size, number of scans and result metrics.

However, some common parameters are identified for comparison. In this regard, six

parameters are used, including dataset, accuracy, sensitivity, specificity FPs/scan and

the size of the nodules considered for experimentation. The most important among

these parameters is the number of false positives because it severely affects the results

of a diagnosis system. Reducing false positives may lead to compromise the

sensitivity, specificity, and accuracy. A good diagnosis system generates a minimum

number of false positives without compromising over the score of other parameters.

Therefore, a good combination of these scores shows the strength of a diagnosis

system.

94

Ta

ble

4.5

: Com

pariso

n o

f the p

rop

ose

d m

ethod

with

the resu

lts of ex

isting

techn

iqu

es

Meth

od

olo

gy

Y

ear

Da

taset

Acc

ura

cy (%

) S

ensitiv

ity(%

) S

pec

ificity(%

) F

Ps/S

can

N

od

ule S

ize

Deh

mesh

ki, et al. [4

9]

20

07

Clin

ical 9

0.0

-

90

.0

14

.6

3- 2

0 m

m

Ye, et al. [5

7]

20

09

Clin

ical -

90

.2

- 8

.2

3-2

0 m

m

Messa

y, et al. [8

9]

20

10

LID

C

80

.4

82

.66

- 3

3

-30

mm

Cascio

, et al. [94

] 2

01

2

LID

C

97

.0

88

.0

97

.0

6.1

3

–3

0 m

m

Kuru

villa a

nd

Gu

navath

i [10

3]

20

14

LID

C

- 9

1.4

-

30

3–

30

mm

Han

, et al. [61

] 2

01

5

LID

C

- 8

9.2

-

4.0

3

–3

0 m

m

Taşcı an

d U

ğur [5

5]

20

15

LID

C

95

.88

-

- -

3–

30

mm

Do

u, et al. [1

19

] 2

01

7

LID

C

- 9

0.7

-

4.0

3

-30

mm

Jiang, et al. [1

28

] 2

01

8

LID

C

- 8

0.0

6

- 4

.7

3-3

0m

m

Ali, et al. [1

27

] 2

01

8

LID

C-ID

RI

64

.4

58

.9

55

.3

- 3

-30

mm

Shaukat, et al. [1

31

] 2

01

9

LID

C

93

.7

95

.5

94

.28

5

.72

3-3

0 m

m

Huid

rom

, et al. [13

0]

20

19

LID

C-ID

RI

93

.23

9

3.2

6

93

.2

- 3

-30

mm

Pro

po

sed M

etho

d

- L

IDC

9

8.8

9

7.8

9

6.4

3

.7

3- 3

0 m

m

94

95

Fig 4.3: Comparison chart of proposed and existing system results

96

Table 4.5 gives a valuable comparison of the proposed technique with other existing

methods in the same domain. Table 4.5 and Figure 4.3 show that the proposed method

outperformed the existing methods in most of the cases. On the other hand, the

proposed technique‘s results are comparable with a few existing methods like Messay,

et al. [89] produced less FPs/scan than the proposed method, but in case of sensitivity,

there is a huge difference. Cascio, et al. [94], also produced better specificity but

lacked in comparison of sensitivity and FPs/scan, which are key parameters to

evaluate a diagnosis method.

4.3 Method II: 3D Nodule Candidate Detection Supported by

Hybrid Features to Reduce False Positives

The details of the dataset, performance evaluation parameters, experimentation, and

comparison of results with existing techniques are discussed in this section.

4.3.1 Dataset and Evaluation Parameters

To evaluate the significance of a diagnosis system, it is very important to evaluate on

some standard benchmarks. There are two key factors involved in performance

evaluation. First is to test the proposed method on some standard dataset and

secondly, use some standard evaluation parameters. Henceforth, the images for

experimentation are taken from a standard dataset LIDC [31]. It is publicly available

at National Biomedical Imaging Archive (NBIA). It contains 84 scans of which 58

scans contain nodules. Each image of this dataset has 512x512 dimensions. Each scan

contains more than 100 images of a patient [34, 35]. All the scans are used in this

study for performance evaluation. The evaluation parameters have a key role in

comparing the performance of different methods. There is a need to set some standard

parameters. In this study, six parameters, including sensitivity, specificity, accuracy,

FPs/scan, nodule size and area under the ROC curve are considered as evaluation

parameters. Actual nodules which are correctly identified by the classifier represented

as TP, whereas TN represents actual nodules which wrongly classified as non-

nodules. FP is the number of non-nodules classified as nodules and FN represents

those candidates, which are incorrectly classified as nodules.

97

4.3.2 Experimentation

Several experiments are performed over a different combination of features and

classifiers. Four different classifiers are used, i.e. SVM, k-NN, Naïve Bayes and

AdaBoost, to distinguish nodules and non-nodules. During this process, seven feature

vectors are created as follows:

V1 = Texture features

V2 = Geometric features

V3 = HOG-PCA

V4 = Geometric +Texture

V5 = Geometric + HOG-PCA

V6 = Texture + HOG-PCA

V7 = Geometric + Texture +HOG-PCA

The holdout cross validation is used for the classification. The classifiers are trained

and tested on 80-20, 70-30, 60-40 and 50-50 distribution, respectively. To avoid the

biasness, experiments are repeated ten times, and the mean score is reported in tables

from Table 4.6 to Table 4.12.

Table 4.6: Results with the geometric feature vector

Classifier Accuracy(%) Sensitivity(%) Specificity(%) AUC

SVM-Linear 86.5 86.2 71.1 0.74

SVM-RBF 93.2 90.1 90.4 0.87

SVM-poly-2 90.1 89.5 86.2 0.86

SVM-poly-3 88.9 88.0 84.3 0.86

3-NN 81.0 83.0 74.7 0.84

5-NN 75.4 77.6 67.1 0.73

Naïve Bayes 94.2 95.7 90.5 0.93

AdaBoost-50 95.5 90.6 93.7 0.95

AdaBoost-60 95.7 93.0 94.6 0.95

AdaBoost-70 96.2 95.4 95.0 0.96

AdaBoost-80 97.0 96.1 95.7 0.96

AdaBoost-90 97.3 96.9 96.5 0.97

AdaBoost-100 97.3 96.8 96.4 0.97

98

Table 4.7: Results with the texture feature vector


SVM-Linear 82.1 85.2 69.7 0.74

SVM-RBF 93.1 90.2 87.4 0.88

SVM-poly-2 90.4 89.4 86.9 0.87

SVM-poly-3 91.2 89.6 87.5 0.87

3-NN 78.9 80.9 73.6 0.86

5-NN 74.2 74.5 65.8 0.72

Naïve Bayes 90.1 90.6 89.4 0.92

AdaBoost-50 92.2 88.6 91.5 0.95

AdaBoost-60 93.6 90.9 91.3 0.95

AdaBoost-70 94.1 93.3 93.1 0.96

AdaBoost-80 94.7 94.2 93.3 0.96

AdaBoost-90 95.2 94.8 94.2 0.97

AdaBoost-100 95.2 94.8 94.0 0.97

Table 4.8: Results with HOG-PCA feature vector


SVM-Linear 79.1 80.4 69.1 0.72

SVM-RBF 83.4 85.6 83.2 0.86

SVM-poly-2 85.0 82.2 84.1 0.87

SVM-poly-3 86.2 84.4 84.3 0.88

3-NN 77.9 79.9 72.6 0.82

5-NN 72.7 70.3 66.1 0.70

Naïve Bayes 91.1 92.6 87.4 0.92

AdaBoost-50 92.4 87.5 90.6 0.91

AdaBoost-60 92.6 89.9 91.5 0.92

AdaBoost-70 93.1 92.3 91.9 0.92

AdaBoost-80 93.9 93.0 92.6 0.93

AdaBoost-90 94.2 93.8 93.4 0.95

AdaBoost-100 94.2 93.7 93.3 0.95

99

Table 4.9: Results with texture + geometric feature vector


SVM-Linear 91.4 90.6 85.3 0.90

SVM-RBF 98.3 97.7 97.2 0.99

SVM-poly-2 94.2 93.6 90.0 0.96

SVM-poly-3 95.8 95.2 93.2 0.97

3-NN 82.1 84.1 73.8 0.86

5-NN 78.3 79.5 68.1 0.74

Naïve Bayesian 95.0 96.3 90.9 0.93

AdaBoost-50 96.6 91.7 94.8 0.96

AdaBoost-60 96.8 94.1 95.7 0.97

AdaBoost-70 97.3 96.5 96.1 0.97

AdaBoost-80 98.1 97.2 96.8 0.98

AdaBoost-90 98.4 98.0 97.6 0.99

AdaBoost-100 98.4 97.9 97.5 0.99

Table 4.10: Results with texture + HOG-PCA feature vector


SVM-Linear 87.2 82.4 83.3 0.85

SVM-RBF 96.2 93.3 93.9 0.97

SVM-poly-2 93.5 92.4 89.8 0.96

SVM-poly-3 93.9 94.0 92.4 0.97

3-NN 81.3 82.2 72.4 0.86

5-NN 77.2 78.1 68.2 0.74

Naïve Bayesian 94.8 95.4 90.1 0.93

AdaBoost-50 96.1 91.3 94.3 0.96

AdaBoost-60 96.3 93.6 95.1 0.97

AdaBoost-70 96.7 96.2 95.5 0.98

AdaBoost-80 97.5 96.6 96.4 0.98

AdaBoost-90 97.9 97.5 97.1 0.99

AdaBoost-100 97.8 97.5 96.9. 0.99

100

Table 4.11: Results with geometric + HOG-PCA feature vector


SVM-Linear 90.2 89.7 83.8 0.89

SVM-RBF 96.7 93.8 94.1 0.97

SVM-poly-2 93.7 92.7 89.9 0.96

SVM-poly-3 94.6 94.3 92.9 0.96

3-NN 82.0 83.7 73.4 0.86

5-NN 76.6 76.1 67.7 0.73

Naïve Bayesian 96.8 97.7 91.8 0.93

AdaBoost-50 96.4 91.6 94.6 0.97

AdaBoost-60 96.7 93.7 95.3 0.97

AdaBoost-70 97.2 96.3 95.7 0.97

AdaBoost-80 97.7 96.7 96.6 0.98

AdaBoost-90 98.2 97.8 97.3 0.99

AdaBoost-100 98.0 97.6 97.2 0.99

Table 4.12: Results with geometric + texture + HOG-PCA feature vector


SVM-Linear 91.5 90.5 88.1 0.90

SVM-RBF 98.6 97.9 97.4 0.99

SVM-poly-2 95.2 96.0 94.3 0.98

SVM-poly-3 97.6 97.2 96.2 0.98

3-NN 83.4 85.6 74.3 0.87

5-NN 79.0 80.9 70.1 0.78

Naïve Bayesian 93.3 94.8 87.6 0.89

AdaBoost-50 96.9 92.0 95.1 0.97

AdaBoost-60 97.1 94.3 96.0 0.98

AdaBoost-70 97.6 96.8 96.3 0.98

AdaBoost-80 98.3 97.4 97.1 0.99

AdaBoost-90 99.2 98.3 98.0 0.99

AdaBoost-100 99.2 98.3 97.8 0.99

101

SVM is applied with different three kernels, including linear, RBF and polynomial of

degree 2 and 3. K-NN is applied with k=3and k=5 neighborhoods. Similarly,

AdaBoost is also used with different variations in the number of iterations. The

iterations of AdaBoost represent the number of weak classifiers used in

experimentation. The increase in the number of weak classifiers is directly

proportional to the time complexity, which leads to slow down the execution process.

On the other hand, a lower number of weak classifiers results in reduced accuracy.

Therefore, it is much needed to select an optimum value of weak classifiers resulting

in good accuracy with less computation time. The highest accuracy is achieved over

90 iterations. Figure 4.4 shows the ROC curve for a different number of iterations.

For implementation, MATLAB 2016a is used.

The tables from Table 4.6 to Table 4.8 illustrate the best results produced by each

classifier with a single type of feature vector. In addition to that, hybrid feature

vectors V4, V5, V6, and V7 are also used for classification. The results with these

feature vectors are shown in Table 4.9 and Table 4.12. SVM produced the best results

with 70%-30% data distribution for training and testing, respectively. The best results

are with RBF kernel, i.e., 93.1% accuracy, 90.2% sensitivity and 87.4% specificity

over texture features, as shown in Table 4.6; 93.2% accuracy, 90.1% sensitivity and

90.4% specificity with geometric features reported in Table 4.7. Similarly, 83.4%

accuracy, 85.6% sensitivity and 83.2% specificity with HOG-PCA reported in Table

4.8. On the other hand, Table 4.9 to Table 12 show results with hybrid features and

best results are with feature vector V7, which is a combination of all three types of

features including geometric, texture and HOG-PCA, reported in Table 4.10 and

Table 4.11. These results are 99.2% of accuracy, 98.3% of sensitivity and 98.0% of

specificity with 3.3 FPs/scan. The results show that a single type of feature vector

cannot be a true representative of nodules. It is clear that results with hybrid feature

vectors are better than the single type of features. The comparison of results, reported

from Table 4.6 to Table 4.8, illustrates that in a single type of feature vector, the

geometric features are better representatives of nodules. However, one type of feature

vector is not the best descriptor of the nodules.

102

(a)

(b)

Fig 4.4: ROC analysis (a) AdaBoost results with different number of iterations over V7

(b) SVM results over hybrid feature vector V7

103

(a)

(b)

Fig 4.5: ROC analysis (a) results using feature vector V7 (b) results using feature vector

V5

104

Further, Fig 4.4 and Fig 4.5 give a valuable performance comparison of all the

classifiers with different feature vectors. It shows results over feature vector consists

of texture and geometric and HOG-PCA features. These results show that AdaBoost

has better performance over the other classifiers. Fig 4.5 (a) shows the results of

optimum feature vector, including intensity, shape and HOG features.

It is clear that the results are improved using optimum feature vectors for AdaBoost

and SVM. However, the performance is desecrated in the case of Naïve Bayes. Fig

4.5 shows that AdaBoost outperforms the other classifiers. The ROC values of

AdaBoost, SVM and Naïve Bayes and k-NN are 0.99, 0.99, 0.87 and 0.87,

respectively.

4.3.3 Comparison of Proposed Method with Existing Techniques

A brief comparison is given in Table 4.13, and it shows a clear picture that the

proposed method has produced better performance than the existing techniques in

most of the parameters. Messay, et al. [89] produced less number of false positives

than the proposed method but lacked in accuracy and sensitivity. Magalhães Barros

Netto, et al. [93] produced a remarkable score in FPs/scan, but that score is over a

smaller subset of the dataset. They have used only 29 scans of LIDC dataset. Besides

false positives, they have a very low score in other parameters.

Dehmeshki, et al. [49] evaluated their technique over a private dataset instead of the

standard dataset and nodule size of 3-20mm with high FPs/scan. The recently

developed deep learning methods [118, 119, 127] for lung nodule detection has not

shown promising results. Nibali, et al. [118] also pointed out that deep learning is

facing challenges in automated diagnosis. The comparison of results depicts that the

proposed method has reduced the number of FPs/scan with good accuracy, sensitivity,

and specificity.

105

Ta

ble

4.1

3: R

esults c

om

pariso

n o

f the p

rop

osed

meth

od

with

existin

g tech

niq

ues

Meth

od

olo

gy

Y

ear

Da

tasets

Acc

ura

cy(%

) S

ensitiv

ity(%

) S

pec

ificity(%

) F

Ps/S

can

N

od

ule S

ize(mm

)

Deh

mesh

ki, et al. [4

9]

20

07

Clin

ical 9

0.0

-

- 1

4.6

3

- 20

Messa

y, et al. [8

9]

20

10

LID

C

80

.4

82

.66

- 3

.0

3-3

0

Mag

alhães B

arros N

etto, et al. [9

3]

20

12

LID

C

90

.65

8

5.9

3

90

.79

0

.13

8

3–

30

Cho

i and

Cho

i [54

] 2

01

4

LID

C

99

.0

97

.5

97

.5

6.7

6

3-3

0

Kuru

villa a

nd

Gu

navath

i [10

3]

20

14

LID

C

- 9

1.4

-

30

3–

30

Han

, et al. [61

] 2

01

5

LID

C

- 8

9.2

-

4.0

3

–3

0

Taşcı an

d U

ğur [5

5]

20

15

LID

C

95

.88

-

- -

3-3

0

Shen, et al. [1

15

] 2

01

5

LID

C

97

.7

- -

- 3

-30

Nib

ali, et al. [11

8]

20

17

LID

C

89

.9

91

.1

88

.6

- 3

-30

Do

u, et al. [1

19

] 2

01

7

LID

C

- 9

0.7

-

4.0

3

-30

Jaffar, et al. [12

4]

20

18

LID

C

98

.7

97

.5

98

.18

- 3

-30

Ali, et al. [1

27

] 2

01

8

LID

C-ID

RI

64

.4

58

.9

55

.3

- 3

-30

Huid

rom

, et al. [13

0]

20

19

LID

C-ID

RI

93

.23

9

3.2

6

93

.2

- 3

-30

Shaukat, et al. [1

31

] 2

01

9

LID

C

93

.7

95

.5

94

.28

5

.72

3-3

0

Pro

po

sed M

etho

d

L

IDC

9

9.2

9

8.3

9

8.0

3

.3

3-3

0

105

106

4.4 Method III: Multistage Segmentation Model and SVM-

Ensemble for Precise Lung Nodule Detection

In this method, an authentic publicly available dataset LIDC-IDRI [34] is used to

evaluate the performance of the proposed method. The tube current range, from 40

milliampere-seconds (mA.s) to 627mA.s, is used for image acquisition. The tube

potential energies have a range from 120kV to 140kV. The low-dose scans are more

effective to identify small size lung nodules [192]. The dataset contains 1018 CT

scans of 1010 patients collected from 7 academic centers. Each scan contains

approximately 100 to 400 slices with slice thickness between 0.6mm to 5mm and

dimension of 512×512 pixels. Each image has a 12-bit grayscale resolution with the

pixel size ranging from 0.5mm to 0.8mm. Out of 1018 cases, 121 cases have slice

thickness ≥ 3mm, which is not suitable for the computer based diagnosis [193]. In

addition to this, there are nine cases that have inconsistent slice spacing. By excluding

these cases, a total of 888 scans are left for the automated diagnosis. Every scan of the

dataset is reviewed by four experienced radiologists for nodule annotation. The

annotation information for each scan is provided in XML format. All the radiologists

have a consensus on 777 nodules. In this study, those nodules are considered for

experimentations, which are marked by all the four radiologists. Around 45% scans of

dataset contain no nodules, which are annotated by all the four radiologists, and some

scans have multiple nodules. Considering these facts, those scans are excluded, which

have zero nodules, and those scans are selected, which contain the number of nodules

in the range of 1-8. As a result, 250 scans are finally selected, with 567 nodules.

Candidate nodules are classified according to annotations provided by the expert

radiologists. If a nodule candidate meets the criteria, it is considered as hit and

classified as a nodule. On the other hand, if it does not meet the criteria, it is taken as

a miss and considered as a non-nodule. A candidate is considered hit if its radius is

0.8 to 1.5 times of the actual nodule or its distance with reference to any

corresponding annotated nodule is less than 1.5 times the radius of that nodule [54].

The performance of classifiers is checked over different distributions of data into

training and testing ratio.

107

4.4.1 Experimentation and Results

MATLAB 2016a is used as an implementation tool. Three different distributions of

data are used as 80-20, 70-30 and 50-50 percent for training and testing purposes,

respectively. The hold-out validation is used for training and testing the classifiers.

The process is repeated 10 times to avoid the biasness and average score is computed.

Various experiments are executed with different combinations of feature sets. The

results are shown in Table 4.14 to Table 4.16 with geometric, texture and GTFD

features, respectively. SVM is trained with three different kernels, namely SVM-

linear, SVM-polynomial, and SVM-RBF. The output of individual classifiers over the

test examples is evaluated. Majority voting is applied to make a final decision by

SVM-ensemble. From Table 4.14 to Table 4.16, it is clear that SVM-ensemble has

produced significantly better results as compared to individual classifiers.

Table 4.14: Results with geometric features

Classifier Training-Testing Sensitivity(%) Specificity(%) Accuracy(%) AUC

SVM-linear 80-20 85.2 72.1 82.3 0.90

70-30 85.7 73.2 82.5 0.90

50-50 83.3 70.3 79.8 0.90

SVM-polynomial 80-20 88.6 87.3 90.5 0.93

70-30 90.3 88.4 91.9 0.93

50-50 87.2 84.7 88.8 0.93

SVM-RBF 80-20 92.7 90.9 92.5 0.96

70-30 94.6 92.2 93.4 0.96

50-50 90.6 88.6 89.2 0.96

SVM-ensemble 80-20 94.2 91.4 93.2 0.98

70-30 95.1 93.7 94.4 0.99

50-50 92.0 89.8 92.2 0.98

Performance of the proposed method is evaluated with the individual as well as

combined feature sets. The results show that combined feature vectors produced a

better performance as compared to a single feature type.

108

Table 4.15: Results with texture features


SVM-linear 80-20 84.5 72.1 81.8 0.88

70-30 85.2 72.7 82.1 0.88

50-50 83.1 69.9 79.3 0.87

SVM-polynomial 80-20 88.1 86.4 89.8 0.91

70-30 89.6 87.5 91.2 0.91

50-50 86.8 84.2 88.3 0.90

SVM-RBF 80-20 92.3 90.2 91.9 0.94

70-30 93.9 91.8 92.7 0.94

50-50 90.0 87.9 88.4 0.94

SVM-ensemble 80-20 93.4 90.9 92.2 0.97

70-30 94.8 93.2 93.7 0.97

50-50 91.3 89.1 90.8 0.97

Table 4.16: Results with GTFD


SVM-linear 80-20 88.2 75.6 85.3 0.93

70-30 89.1 76.7 85.8 0.93

50-50 86.4 73.2 83.1 0.92

SVM-polynomial 80-20 91.7 90.6 93.9 0.98

70-30 93.6 91.2 95.3 0.98

50-50 91.0 88.9 92.8 0.97

SVM-RBF 80-20 96.9 96.7 97.4 0.99

70-30 97.7 97.4 98.3 0.99

50-50 95.8 95.8 96.8 0.98

SVM-ensemble 80-20 98.0 97.6 98.4 0.99

70-30 98.6 98.2 99.0 0.99

50-50 97.2 96.0 97.3 0.99

109

The performance with texture features is lower than the geometric features. However,

when GTFD is used by encapsulating both geometric and texture features, better

results are obtained as compared to a single type of feature vector. Furthermore, best

results are shown when experimentation is done over the combination of 70-30, i.e.,

70% training and 30% testing.

Table 4.16 depicts that GTFD presented better results than individual features. In

comparison to Table 4.14 and Table 4.15, it gives a clear picture that geometric

features are better descriptors of nodules as compared to texture features. However,

on an individual basis, they cannot produce the best results. Fig 4.6 illustrates a

receiver operating characteristic (ROC) curve for the results of classification. It is

evident that ensemble based SVM shows better results as compared to RBF,

polynomial and linear kernels.

Fig 4.6: ROC curve with GTFD feature vector

The nodule candidate detection process comprises of two stages. In the first stage,

66021 nodule candidates are detected with 261.8 FPs/scan. At this step, all 567 actual

nodules are also recognized with 100% sensitivity. The second stage is related to

candidate refinement on the basis of nodule size and shape information. In this stage,

the number of candidates is reduced to 46176, and FPs/Scan is decreased to 182.5

with 45614 total false positives. However, this refinement phase has missed 5 actual

nodules and sensitivity is reduced to 99.1%.

The classification phase, based on SVM-ensemble, reduces the number of false

positives to 838, which is 3.4/scan. The free-response receiver operating characteristic

110

(FROC) curve in Fig 4.7 (a) illustrates the number for FPs/scan produced by four

classifiers. The SVM with linear kernel delivers 9.6 FPs/scan with 89.1% sensitivity,

SVM-polynomial produces 6.4 FPs/scan with 93.6% sensitivity, and SVM-RBF

shows comparatively good results with 4.2 FPs/scan and sensitivity of 97.7%.

However, SVM-ensemble reduces the FPs/scan to 3.4% with the highest sensitivity of

98.6%.

(a)

(b)

Fig 4.7: FROC analysis (a) FROC curve shows false positive reduction by each classifier,

(b) FROC curve illustrates the sensitivity comparison of each classifier at 3.4 FPs/scan

111

In Fig 4.7 (b), another FROC curve is given, which illustrates the comparison of

classifiers‘ sensitivity at 3.4 FPs/scan. It depicts that SVM-ensemble has the highest

sensitivity, i.e., 98.6% and SVM-linear has the lowest. SVM-RBF has better

sensitivity than SVM-polynomial at 3.4 FPs/scan. It is evident from FROC curves that

minimum FPs/scan with SVM-ensemble is reduced to 1 but at the cost of sensitivity

degradation with 92.1%.

4.4.2 Comparative Analysis

For estimating the significance of the proposed method in relation to reported

techniques in the literature, a comparison of datasets and their results is presented in

Table 4.17 and Table 4.18, respectively.

Table 4.17: Comparison of datasets used in existing and proposed systems

Methodology Year Dataset Scans Nodules Nodule Size

Dehmeshki, et al. [49] 2007 Clinical 70 179 3-20 mm

Ye, et al. [57] 2009 Clinical 108 220 3-20 mm

Tan, et al. [48] 2011 LIDC 125 80 3-30 mm

Riccardi, et al. [56] 2011 LIDC 154 117 3-30 mm

Cascio, et al. [94] 2012 LIDC 84 148 3-30 mm

Teramoto and Fujita [97] 2013 LIDC 84 103 5-20 mm

Choi and Choi [54] 2014 LIDC 84 148 3-30 mm

Kuruvilla and Gunavathi [103] 2014 Clinical 155 110 3-30 mm

Han, et al. [61] 2015 LIDC-IDRI 205 490 3-30 mm

Nibali, et al. [118] 2017 LIDC-IDRI - - 3-30 mm

Jaffar, et al. [124] 2018 LIDC 84 148 3-30mm

Teramoto and Fujita [125] 2018 Clinical 100 186 -

Zhang, et al. [123] 2018 LIDC-IDRI 71 168 3-30mm

Shaukat, et al. [131] 2019 LIDC 84 148 3-30mm

Huidrom, et al. [130] 2019 LIDC-IDRI 300 - 3-30mm

Proposed Method LIDC-IDRI 250 567 3-30 mm

112

Ye, et al. [57] used clinical data of 108 scans containing 220 nodules of size 3-20mm

with a slice thickness of 0.5-2mm. Table 4.17 illustrates that some methods used 84

CT scans of LIDC dataset consisting of 148 nodules marked by any of the radiologists

[54, 94, 124, 131]. Dehmeshki et al. [49] made use of 70 CT scans of a clinical dataset

containing 179 nodules of size 3-20mm with a slice thickness of 0.5-1.25mm.

Kuruvilla and Gunavathi [103] utilized a clinical dataset of 155 scans with a slice

thickness of 0.75-1.25mm, including 110 nodules. In addition, they have also used

some scans of LIDC dataset without LIDC annotations. Han, et al. [61] used 205 CT

scans of LIDC-IDRI dataset containing 490 nodules marked by any of the

radiologists. The nodules included in their study were of 3-30mm size. Teramoto and

Fujita [97] made use of 84 CT scans of LIDC dataset with 103 nodules of size 5-

20mm marked by any of the radiologists. Nibali, et al. [118] utilized LIDC-IDRI

dataset for experimentations, but details are not provided. Tan, et al. [48] made use of

125 CT scans of LIDC dataset containing 80 nodules of size 3-30mm marked by all

the four radiologists. Riccardi, et al. [56] used 154 CT scans of LIDC dataset

consisting of 117 nodules with size 3-30mm marked by all the four radiologists.

Huidrom, et al. [130] used 300 scans, but nodule information is not provided. Zhang,

et al. [123] used a smaller subset of LIDC-IDRI dataset of 71 scans containing 168

nodules. The proposed method is evaluated over 250 scans of LIDC-IDRI dataset

containing 567 nodules with size 3-30mm annotated by all four radiologists.

The statistics in Table 4.17 highlight the fact that the proposed method is tested over a

more significant number of scans and nodules as compared to other research works

reported in the literature. Table 4.18 compares the results of the proposed and existing

methods. Kuruvilla and Gunavathi [103] produced 91.4% sensitivity but with a very

high FPs/scan of 30. Moreover, they did not provide specificity. Choi and Choi [54]

claimed 97.5% sensitivity as well as specificity with 6.76 FPs/scan. Ye, et al. [57]

produced 8.2 FPs/scan with 90.2% sensitivity. Cascio et.al [94] claimed 88%

sensitivity and 97% accuracy with 6.1 FPs/scan. Teramoto and Fujita [97] achieved a

low FPs/scan of 4.2 but at the cost of lower accuracy of 80%. Han, et al. [61] claimed

a low FPs/scan of 4.14, but their sensitivity was low as 89.2%. Nibali, et al. [118]

showed 89.9%, 91.07% and 88.64% accuracy, sensitivity, and specificity,

respectively.

113

Table 4.18: Comparison of results with existing methods

Methodology Year Accuracy(%) Sensitivity(%) Specificity(%) FPs/Scan

Ye, et al. [57] 2009 - 90.2 - 8.2

Tan, et al. [48] 2011 - 87.5 - 4.0

Riccardi, et al. [56] 2011 - 71.0 - 6.5

Cascio, et al. [94] 2012 97.0 88.0 - 6.1

Teramoto and Fujita [97] 2013 80.0 - - 4.2

Choi and Choi [54] 2014 99.0 97.5 97.5 6.76

Kuruvilla and Gunavathi [103] 2014 - 91.4 - 30

Han, et al. [61] 2015 95.88 89.2 - 4.14

Nibali, et al. [118] 2017 89.9 91.07 88.64 -

Jaffar, et al. [124] 2018 98.7 97.5 98.18 -

Teramoto and Fujita [125] 2018 - 83.0 - 5.0

Zhang, et al. [123] 2018 - 89.3 - 2.1

Shaukat, et al. [131] 2019 93.7 95.5 94.28 5.72

Huidrom, et al. [130] 2019 93.23 93.26 93.2 -

Proposed Method 99.0 98.6 98.2 3.4

Tan, et al. [48] produced 87.5% sensitivity with 4.0 FPs/scan and Riccardi, et al. [56]

claimed 6.5 FPs/scan with a low sensitivity of 71.0%. Huidrom, et al. [130] produced

93.26% sensitivity, but number of FPs is not reported. Zhang, et al. [123] claimed

better FPs/scan than the proposed method. However, their sensitivity is very low in

comparison to the proposed method. Besides, they have used a smaller size dataset for

experimentation. By concluding the whole discussion, it is evident that the proposed

method has shown better results, i.e. 99.0% accuracy, 98.6% sensitivity and 98.2%

specificity with a 3.4 FPs/scan which is a clear proof of the effectiveness.


Geometric Fit in Parametric Form and Deep Learning

The details of the dataset, data distribution, results of segmentation, initial candidate

detection and classification are discussed in this section. MATLAB 2016b is used as

114

an implementation tool, on Intel Xeon 3.5 GHz with 16 GB RAM, Windows Server

2012 R2 Standard.

4.5.1 Dataset Selection

The performance evaluation is conducted over a publicly available dataset LIDC-

IDRI [34]. It is the largest dataset of lung CT images, which has 1018 scans of 1010

patients. Each scan contains 100 to 400 slices with slice thickness 0.6mm to 5mm and

dimensions 512x512. The nodules in the dataset are of size 3mm to 30mm in diameter

[35]. There are 121 scans which have slice thickness ≥3mm. These scans are not

suitable to use in computer-based diagnosis systems [193]. In addition to that nine

scans have inconsistent slice spacing and not suitable for automated diagnosis

systems. As a result, 888 scans are left. Four experienced radiologists examine all the

scans for nodule annotation. These annotations are given in XML format, along with

each scan. Total 777 nodules are used for experimentation, which are marked by all

the radiologists.

4.5.2 Segmentation Results

Accurate results of segmentation affect the performance of successor steps. The initial

values of FODPSO parameters are given in Table 4.19.

Table 4.19: FODPSO initialization parameters

Parameter Initial Value

Total number of iterations ( ) 100

Number of swarms ( ) 4

Minimum number of swarms ( ) 2

Number of particles in each swarm ( ) 20

Minimum number of particles in each swarm ( ) 10

Maximum number of particles in each swarm ( ) 30

Maximum number of swarms ( ) 6

Fractional coefficient ( ) 0.6

Global weight controller ( ) 1.2

Local weight controller ( ) 0.8

115

The position of each particle is initialized within the gray intensity levels from to

. The results of the segmentation using FODPSO are compared with the existing

methods. Three quantitative measures are used, including Jaccard measure [194],

probabilistic random index (PRI) [195] and variation of information (VoI) [196], as

performance measures. The Jaccard similarity coefficient compares the similarity

between the segmented image ( ) and the reference ground truth image ( ). The

mathematical form of Jaccard is given in Eq. 4.1.

(4.1)

The second parameter, probabilistic random index (PRI) is used to check the

consistency of the segmented image with reference to the actual image. It checks the

degree of closeness of intensities of a pair of pixels in the segmented and original

image. The PRI value ranges from 0 to 1, where 1 yields an ideal segmentation. The

PRI value can be calculated by Eq. 4.2.

( )∑ ∑ ( ( )

( ))

(4.2)

where is a discriminant function that determines whether the pixel in has the

same label with the corresponding pixel in and is the number of overlapping

regions. In information theory, the Variation of Information (VoI) also known as

shared information, quantifies the degree of mutual information in the segmented

image and the ground truth image. The VoI can be computed by Eq. 4.3

(4.3)

where H(Is) and H(Ir) are the entropies calculated by the Eq. 4.4 and Eq. 4.5 given

below.

∑

(4.4)

116

∑

(4.5)

M(Is, Ir) represents the mutual information and is computed by Eq. 4.6.

∑∑

(4.6)

The value of VoI ranges from 0 to ∞, where 0 represents perfect segmentation. The

segmentation results of FODPSO are compared with the well-known methods like

fuzzy entropy based optimal threshold, Otsu method and use of superpixel for

segmentation. Table 4.20 gives a brief comparison of the segmentation results, which

depicts that proposed FODPSO based segmentation has significantly improved the

results.

Table 4.20: Comparison of segmentation results

Method Year Measure Without Nodules Pleural Nodules

Jaffar, et al. [197] 2010 VoI 1.5428 3.893

PRI 0.9527 0.8729

Jaccard 0.9254 0.8723

Choi and Choi [96] 2012 VoI 1.779 3.825

PRI 0.9581 0.8556

Jaccard 0.9375 0.8535

Liao, et al. [198] 2016 VoI 1.3815 2.2052

PRI 0.9624 0.8926

Jaccard 0.9413 0.8821

Jaffar, et al. [124] 2018 VoI 1.1152 1.9748

PRI 0.9695 0.9142

Jaccard 0.9587 0.9032

Proposed Method

VoI 1.013 1.9712

PRI 0.9703 0.9201

Jaccard 0.9621 0.9093

117

(a)

(b)

Fig 4.8: 3D lung volume (a) left view and (b) right view

118

(a)

(b)

Fig 4.9: Extracted 3D ROI with nodules and vessels (a) normal view of the

complete lung (b) a portion containing juxta-vascular nodule

119

The 3D volume is shown in Fig 4.8, which also indicates a juxta-pleural nodule. It

illustrates the accuracy of the segmentation. Furthermore, Fig 4.9 shows a 3D ROI of

the segmented lung region, including vessels and nodules. The 3D volumes are shown

in Fig 4.8 and Fig 4.9 consists of 250 slices with 512x512 dimensions on z, x, and y-

axis, respectively. Fig 4.9 (b) shows a portion of ROI shown in Fig 4.9 (a), which

highlights some juxta-vascular nodules.

4.5.3 Initial Candidate Detection Results

After the lung volume extraction, the nodule candidate detection is performed. The

initial candidates are extracted on the basis of geometric fit in parametric form.

Afterwards, a geometric rule-based refinement is performed on the nodule candidate

set. In the initial stage, large numbers of nodule candidates are generated. In this

phase, all actual nodules are successfully included in the candidate set with 100%

sensitivity with 205.4 FPs/scan. In the refinement phase of candidate detection, the

number of false positives is reduced to 82584 with 93 FPs/scan. However, during the

refinement phase, 13 actual nodules are also missed, and sensitivity degrades to

98.33% with 1.67% false negative rate. The number of false positives is further

reduced in the classification phase.

4.5.4 Classification Results

After segmentation, nodule candidates are extracted, including 764 actual nodules and

the remaining are non-nodules. It includes juxta-vascular, juxta-pleural and isolated

nodules of size ranging from 3mm to 30mm. The classification is performed with

deep learning. The classifiers are trained and tested over holdout cross validation with

three different data distributions 80-20, 70-30 and 50-50. The experiments with each

data distribution are repeated 10 times and mean values are computed. The best

results are achieved with 70-30 data distribution, where 70-30 represents 70% training

and 30% test data, respectively. These results are presented in Table 4.21 and Table

4.22. The first phase of this learning network is a stacked autoencoder followed by

softmax classifier. The stacked autoencoder is used for feature optimization, which

improves the performance of the classifier. In the feature optimization step, two level

autoencoder is used. The weight regularization is initialized to 0.001, sparsity

regularization to 4, and sparsity proportion to 0.05. Fig 4.10 shows the working

structure of a simple autoencoder. It takes a 44 size feature vector and encodes it to a

120

reduced set of 10 feature vector. In the second step, it decodes back to a 44 size

feature vector. However, in the case of stacked autoencoder, instead of decoding the

encoding layer pass the output to the next encoding layer. As a result, a hierarchy of

encoding is formed, which results in the lower dimensional feature vector. The size 10

is selected empirically. Experimentation is performed with different sizes of hidden

layers ranging from 5 to 14, and it is found that the best classification is at value 10.

This feature vector is passed to the softmax to distinguish nodules from other objects.

Fig 4.10: The first layer of autoencoder

In Fig 4.11 two autoencoders are stacked, which reduces the 44-dimensional feature

vector to a 10-dimensional feature vector. This feature reduction has improved the

performance of the classifier in terms of computation time and accuracy.

Fig 4.11: Stacked autoencoder and softmax for feature reduction and classification

Table 4.21 shows the classification results with and without feature reduction. It is

clear that when the classification is performed by softmax without feature

optimization, the results are poor. The results with one level autoencoder are

significantly improved, but the overall performance is very poor. However, when a

two level autoencoder is used, it significantly improves the classification

performance.

121

Table 4.21: Results with different variations of deep net

Classifier Accuracy(%) Sensitivity(%) Specificity(%) FPs/Scan

Two Level SA+softmax 96.9 95.6 97.0 2.8

One encoder +softmax 90.3 89.8 84.6 11.7

Without autoencoder 75.9 68.6 49.8 45.2

The results of the deep net classification are compared with the state of the art

classifiers, including SVM [199], k-NN [200], Naïve Bayes [201] and Decision Tree

[202] as illustrated in Table 4.22. The proposed deep net has outperformed the state of

the art classifiers in all the metrics. Fig 4.12 shows free-response receiver operating

characteristic (FROC) curves, which gives a brief comparison of sensitivity and

FPs/scan produced by each classifier. Fig 4.12(a) illustrates the FPs produced by each

classifier.

SVM produced 3.4 FPs/scan with 92.6% sensitivity, Naïve Bayes produced 90.8%

sensitivity with 6.4 FPs/scan, k-NN produced 85.3% sensitivity with 8.8 FPs/scan and

decision tree produced 89.2% sensitivity with 6.9 FPs/scan, whereas the proposed

method achieved 95.6% sensitivity with very low FPs of 2.8 per scan. Fig 4.12 (b)

shows another FROC curve providing a comparison of sensitivity produced by each

classifier at 2.8 FPs/scan. It shows that there is a wide difference of sensitivity

produced by the proposed method and other state of the art classifiers like SVM has

86.9%, Naïve Bayes has 82.0%, k-NN has 75.3%, and decision tree has 79.0%. These

results are clear evidence of the significance of the proposed method.

Table 4.22: Comparison of classification results

Classifier Accuracy(%) Sensitivity(%) Specificity(%) FPs/Scan

SVM 96.3 92.6 96.4 3.4

Naïve Bayes 91.2 90.8 91.6 6.4

k-NN 84.2 85.3 83.8 8.8

Decision Tree 89.7 89.2 89.6 6.9

SA +softmax 96.9 95.6 97.0 2.8

122

(a)

(b)

Fig 4.12: FROC analysis (a) FROC curve showing the performance of each classifier, (b)

FROC curve comparing the performance of classifiers at 2.8 FPs/scan

123

4.5.5 Comparative Analysis

Table 4.23 provides a brief comparison of the proposed and reported methods in the

literature based on common parameters. It is important to mention that some methods

use clinical data and others have used LIDC and its extended version LIDC-IDRI.

However, researchers have used different subsets of these datasets with variations in

nodule sizes and annotations. Kuruvilla and Gunavathi [103] produced 91.4%

sensitivity and 30 FPs/scan on a dataset of 155 scans containing 110 nodules.

Riccardi, et al. [56] worked on 154 LIDC scans containing 117 nodules, but their

results are not promising, i.e. 6.5 FP/scans with 71.0%.sensitivity. Brown, et al. [107]

achieved 2.0 FPs/scan, but these results are on a small dataset of 108 CT scans

containing 68 nodules. In addition, they have only considered large nodules of size

4mm ≥, and sensitivity is also very low. Nibali, et al. [118] produced 89.9% accuracy,

91.1% sensitivity, and 88.6% specificity respectively, but they have not provided the

details about the size of dataset and number of FPs. Cascio et.al [94] produced 88.0%

sensitivity, 97.0% accuracy and 6.1 FPs/scan on 84 CT scans containing 148 nodules,

from LIDC, of size 3-30mm with consensus of all four radiologists. Taşcı and Uğur

[55] targeted only juxta-pleural nodules and evaluated performance on 24 scans

containing 124 nodules produced 95.88% accuracy. The evaluation was performed

over LIDC dataset. Teramoto and Fujita [97] used 84 scans of LIDC dataset and

claimed 4.2 FPs/scan and 80% accuracy. They have not considered the smaller

nodules <5mm. In another method, Teramoto and Fujita [125] used clinical data and

produced 83% sensitivity and 5.0 FPs/scan. Tan, et al. [48] produced 4.0 FPs/scan

and 87.5% sensitivity over a subset of LIDC dataset of 125 CT scans containing 80

nodules of 3-30mm size, annotated by all 4 radiologists. Messay, et al. [89] reported

3.0 FPs/scan, 82.7% sensitivity and 80.4% accuracy. Total 143 nodules from 84 CT

scans of LIDC dataset were used for the evaluation. Han, et al. [61] evaluated their

method over a subset of LIDC-IDRI dataset of 205 CT scans containing 490 nodules

with agreement level one and claimed overall 82.7% sensitivity with 4.0 FPs/scan.

Dou, et al. [119] used 888 CT scans containing 1186 nodules. They achieved 90.7%

sensitivity and 4.0 FPs/scan. Ali, et al. [127] used 888 scans containing 1148 nodules

to evaluate their proposed CNN, but their results are not promising. Xie, et al. [132]

achieved 86.42% sensitivity during the candidate detection process. In classification,

it is reduced to 83.4%, 83.2%, and 73.4% with 8, 4, and 0.125 FPs/scan respectively.

124

Tab

le 4

.23

: Com

para

tive a

naly

sis in term

s of d

ata

set an

d resu

lts

Meth

od

olo

gy

Y

ear

Sca

ns

No

du

les N

od

ule S

ize

Da

taset

Acc

ura

cy(%

) S

ensitiv

ity(%

) S

pec

ificity(%

) F

Ps/S

can

Messa

y, et al. [8

9]

20

10

84

14

3

3-3

0m

m

LID

C

80

.4

82

.7

- 3

.0

Tan

, et al. [48

] 2

01

1

12

5

80

3-3

0m

m

LID

C

- 8

7.5

-

4.0

Riccard

i, et al. [56

] 2

01

1

15

4

11

7

3-3

0m

m

LID

C

- 7

1.0

-

6.5

Cascio

, et al. [94

] 2

01

2

84

14

8

3-3

0m

m

LID

C

97

.0

88

.0

- 6

.1

Teram

oto

and

Fujita [9

7]

20

13

84

10

3

5-2

0m

m

LID

C

80

.0

- -

4.2

Kuru

villa a

nd

Gu

navath

i [10

3]

20

14

15

5

11

0

3-3

0m

m

Clin

ical -

91

.4

- 3

0.0

Bro

wn, et al. [1

07

] 2

01

4

10

8

68

4-3

0m

m

LID

C-ID

RI

- 7

5.0

-

2.0

Han

, et al. [61

] 2

01

5

20

5

49

0

3-3

0m

m

LID

C-ID

RI

95

.88

82

.7

- 4

.0

Taşcı an

d U

ğur [5

5]

20

15

24

12

4

- L

IDC

9

5.8

8

- -

-

Nib

ali, et al. [11

8]

20

17

- -

3-3

0m

m

LID

C-ID

RI

89

.9

91

.1

88

.6

-

Do

u, et al. [1

19

] 2

01

7

88

8

11

86

3-3

0m

m

LID

C-ID

RI

- 9

0.7

-

4.0

Teram

oto

and

Fujita [1

25

] 2

01

8

10

0

18

6

3-3

0m

m

Clin

ical -

83

.0

- 5

.0

Ali, et al. [1

27

] 2

01

8

88

8

11

48

3-3

0m

m

LID

C-ID

RI

64

.4

58

.9

55

.3

-

Zhan

g, et al. [1

23

] 2

01

8

71

16

8

3-3

0m

m

LID

C-ID

RI

93

.6

89

.3

- 2

.1

Shaukat, et al. [1

31

] 2

01

9

84

14

8

3-3

0m

m

LID

C

93

.7

95

.5

94

.28

5.7

2

Huid

rom

, et al. [13

0]

20

19

30

0

- 3

-30

mm

L

IDC

-IDR

I 9

3.2

3

93

.26

93

.2

-

Xie, et al. [1

32]

20

19

88

8

- 3

-30

mm

L

IDC

-IDR

I -

83

.4

- 8

.0

Pro

po

sed M

etho

d

- 8

88

77

7

3-3

0m

m

LID

C-ID

RI

96

.9

95

.6

97

.0

2.8

124

125

Huidrom, et al. [130] used 300 CT scans of LIDC-IDRI dataset and produced 93.26%

sensitivity, but details about the number of nodules and FPs are not reported. Shaukat,

et al. [131] achieved 95.5% sensitivity with 5.72 FPs/scan. Zhang, et al. [123] claimed

a low FPs/scan of 2.1, but their sensitivity is significantly low, i.e., 89.3%.

Furthermore, they have achieved these results on a smaller subset of LIDC-IDRI

dataset contacting only 71 scans. Brown, et al. [107] claimed better FPs/scan than the

proposed method. However, they have used a smaller dataset and nodules greater than

4mm whereas, in this study, nodules greater than 3mm are considered for evaluation.

These facts depict the strength of the proposed method.

The results have been verified by clinical experts who have shown confidence that the

proposed methods work on clinical data and will facilitate the radiologists during the

diagnosis process by providing a second opinion, which will save the time of

radiologists.

4.6 Summary

This chapter presents the experiments and results of the proposed methods and their

comparison to the methods reported in the literature. The experimentation is

performed with the most popular and authentic publicly available datasets, including

LIDC and LIDC-IDRI, the latter being the largest. The first and second methods are

evaluated on the LIDC dataset and reduced the number of false positives to 3.7 and

3.3 per scan, respectively. The third and fourth methods are evaluated on the LIDC-

IDRI and reduced the number of false positives to 3.4 and 2.8 per scan, respectively.

The experimental results, therefore, show that there is a significant improvement in

results. Each method has reduced the number of false positives with high sensitivity.

126

Chapter 5

Conclusion and Future Work

127

5.1 Conclusion

Precise and intelligent detection and classification of pulmonary nodules is an arduous

task in the field of medical image analysis. The complex lung structure makes this

problem even more difficult. A radiologist has to examine a large number of CT

images during the diagnosis process. Automated diagnosis systems provide support to

the radiologists to identify the disease in less amount of time by providing the second

opinion. However, there is a need to handle different types of nodules more carefully.

Precise segmentation of nodules helps to overcome these issues. Moreover, relevant

features and robust classifier(s) also assert an enormous impact on the performance of

a diagnosis method.

This thesis has highlighted the issues in the existing approaches on different phases of

the diagnosis process, including segmentation, feature extraction and classification.

We proposed novel methods to address these issues. Our research provides precise

detection and classification of different nodules, including isolated, juxta-pleural and

juxta-vascular nodules. Four methods are proposed in this thesis on the basis of

problems identified from the literature.

In the first method, a novel diagnosis method is proposed based on polygon

approximation and hybrid features. During this study, dynamic optimal thresholding

is computed to extract the lung region. Polygon approximation is proposed to

eliminate the vessels and extraction of nodule candidates from ROI. CLAHE is used

to enhance the nodule candidates. A hybrid feature vector based on geometric,

intensity and HOG features is constructed, and SVM is applied using linear, RBF and

polynomial kernels to distinguish nodules and non-nodules. Experiments show that

the RBF kernel has outperformed the linear and polynomial kernels regarding the

accuracy, sensitivity, and specificity. Experimentation is performed on a publicly

available dataset LIDC, and achieved an accuracy of 98.8%, 97.8% sensitivity with

3.7 FPs per scan.

In method II, a novel nodule candidate detection method is presented. Furthermore,

for better classification of these nodule candidates, seven different feature vectors are

created using a different combination of individual features. Four different classifiers

are trained and tested to identify actual nodules. A number of experiments are

performed with different distributions of data selected from LIDC dataset. It is found

128

that the performance of k-NN and Naïve Bayes is not good. However, SVM has

produced better results than these two. The best results are achieved with AdaBoost

classifier showing an accuracy of 99.2% with sensitivity 98.3%, specificity 98.0%,

and 3.3 FPs/scan. These results are better than the existing techniques in the literature.

In the third method, a more effective 3D segmentation method for lung region and

nodule candidates is proposed. A critical issue, with the existing nodule diagnosis

methods, is the detection of small, boundary attached and vessel attached nodules. If a

detection method misses any nodule, it leads to an inaccurate diagnosis, which might

be life-threatening. To handle such problems, a nodule segmentation method from 3D

CT scans is presented. The first step of the proposed method is based on a hybrid

multistep segmentation technique to extract lung ROI and nodule candidates. SVM-

ensemble is used to distinguish nodules from the set of selected candidates. The

classification is based on a GTFD feature vector encapsulating texture and geometric

features. The proposed method is tested LIDC-IDRI, which is an extended version of

LIDC dataset. The presented method achieved 99.0% accuracy with 98.6% sensitivity

and 98.2% specificity. The number of FPs/scan is also low as 3.4/scan.

The primary focus of the fourth method is the reduction of false positives by

preserving sensitivity. Initial segmentation of lung ROI is performed on the basis of

FODPSO. In the second step, a novel nodule candidate detection method based on

geometric fit in parametric form is proposed, which distinguishes the nodules and

vessels precisely. It impacts the overall performance of the detection method.

Furthermore, a hybrid GTFD is created for better representation of the candidate

nodules. Finally, a stacked autoencoder and softmax algorithm are applied for feature

optimization and classification of nodules. The dataset used in this study is collected

from LIDC-IDRI dataset with nodules of size ≥3mm. The proposed method

significantly improves the results and achieves a low false positive rate as 2.8/scan

with 96.9% accuracy, 95.6% sensitivity, and 97.0% specificity. The comparative

analysis depicts that the proposed methods have significantly reduced the FPs/scan

with high sensitivity. The results have also been verified by the clinical experts.

The comparison of these methods can help the users to select the more appropriate

method to use. The first method performs 2D segmentation of slices. This

shortcoming is eliminated by the second method by performing 3D image analysis

during segmentation and in feature extraction phases. Juxta-vascular nodules are

129

difficult to segment due to their attachment with vessels. In the third method, a robust

multistage segmentation model is proposed to detect different types of nodules

precisely. Instead of slice level segmentation, a novel scan level 3D segmentation is

performed to extract juxta-vascular nodules. It is an important aspect of this method,

which improves the detection process. It outperforms the existing methods. The

SVM-ensemble classifier is implemented based on three heterogeneous kernels. These

improvements impacted the performance of diagnosis methods and outperformed the

existing methods.

In the fourth method, instead of improving one phase of the diagnosis process, we

have improved every phase of the diagnosis process. Furthermore, deep learning is

applied for feature reduction and classification, which strengthens the diagnosis

method to process such a large number of images. It results in the evaluation of the

diagnosis method on 888 CT scans with more than 10 million images. Method II is

better than Method I as it improves the nodule candidate detection process by refining

the initial candidate set. It reduces the number of FPs. Method III and IV are better

than method II due to the better segmentation of juxta-vascular nodules. Furthermore,

in method IV deep learning base nodule classification is performed, which has the

strength to process big data. It is a hot topic in the current era of research. Concluding

the discussion, we can say that all four methods have their strengths. However,

Methods II, III and IV are significantly stronger than the first one. If users have big

data of patients and high processing machines then Method IV will be the best choice

as it also has the strength of deep learning. Combining multiple methods may be a

good approach to achieve good performance. However, it requires high processing

machines. Furthermore, each diagnosis method produces false positives. A

combination of multiple methods may increase false positives. This can be another

area of research to investigate in future.

The investigations conducted in this research lead to the following conclusions:

A robust optimization algorithm improves the computation of threshold

selection, which leads to precise lung volume extraction and better detection

of juxta-pleural nodules from a CT scan.

A multistage segmentation method can extract different types of nodules more

precisely, which impacts the overall performance of a diagnosis method.

130

Scan level segmentation of juxta-vascular nodules improves the sensitivity and

also reduces the false positives as compared to slice level segmentation.

The geometric features, along with texture features, improve the performance

of the diagnosis system. However, the geometric features are more significant

to detect and classify nodules by investigating the shape of the objects within

the CT scan.

Ensemble classifiers and deep neural networks have improved classification

performance. Deep learning also provides the strength to a diagnosis method

to deal with large number of scans.

5.2 Future Work

The proposed methods are tested and evaluated for isolated, juxta-pleural and

juxta-vascular nodules. However, in the future, these methods can be extended

to other types of lung diseases by identifying disorders in the tissue patterns

such as endobronchial nodules, carcinomas presenting as irregular wall

thickening of lung bullae, hilar lesions. The main issue with these diseases is

the availability of a small data, which results in poor training of the classifiers.

Deep learning is a new hot area of research. In this thesis, deep learning is

applied, including autoencoder for feature reduction and softmax for nodule

classification. In the future, Convolutional Neural Network (CNN) can be used

for direct feature extraction and classification. However, there is a big issue

with CNN that it requires high computation during the training process, which

requires high-performance machines and Graphics Processing Unit (GPU).

Current repositories of lung CT scans contain one scan for each patient. The

diagnosis process can be improved by examining multiple scans of a patient

taken on different time spans. The change in the lung can be observed

precisely by comparing two different scans of a patient.

131

References

[1] P. Jha, M. MacLennan, F. J. Chaloupka, A. Yurekli, C. Ramasundarahettige,

K. Palipudi, et al., "Global hazards of tobacco and the benefits of smoking

cessation and tobacco taxes," DISEASE CONTROL PRIORITIES• THIRD

EDITION, p. 175, 2015.

[2] G. Veronesi, M. Bellomi, J. L. Mulshine, G. Pelosi, P. Scanagatta, G.

Paganelli, et al., "Lung cancer screening with low-dose computed

tomography: a non-invasive diagnostic protocol for baseline lung nodules,"

Lung Cancer, vol. 61, pp. 340-349, 2008.

[3] R. L. Siegel, K. D. Miller, and A. Jemal, "Cancer statistics, 2017," CA: A

Cancer Journal for Clinicians, vol. 67, pp. 7-30, 2017.

[4] W. H. Organization. (2019, 14/01/2019). Cancer fact sheet Available:

http://www.who.int/mediacentre/factsheets/fs297/en/

[5] T. Sobue, N. Moriyama, M. Kaneko, M. Kusumoto, T. Kobayashi, R.

Tsuchiya, et al., "Screening for lung cancer with low-dose helical computed

tomography: anti-lung cancer association project," Journal of Clinical

Oncology, vol. 20, pp. 911-920, 2002.

[6] Y. Kurihara, S. Matsuoka, T. Yamashiro, A. Fujikawa, S. Matsushita, K.

Yagihashi, et al., "MRI of Pulmonary Nodules," American Journal of

Roentgenology, vol. 202, pp. W210-W216, 2014.

[7] R. E. Coleman, "PET in lung cancer," Journal of Nuclear Medicine: Official

Publication, Society of Nuclear Medicine, vol. 40, pp. 814-820, 1999.

[8] Z. Wang, Y. Hu, Y. Wang, W. Han, L. Wang, F. Xue, et al., "Can CT

Screening Give Rise to a Beneficial Stage Shift in Lung Cancer Patients?

Systematic Review and Meta-Analysis," PloS One, vol. 11, p. e0164416,

2016.

http://www.who.int/mediacentre/factsheets/fs297/en/

132

[9] J. Zhao, S. H. Wang, X. Liu, Y. Liu, and Y. Q. Chen, "Early diagnosis of

cirrhosis via automatic location and geometric description of liver capsule,"

The Visual Computer, vol. 34, pp. 1677-1689, 2018.

[10] D. Sathish, S. Kamath, K. Prasad, and R. Kadavigere, "Role of normalization

of breast thermogram images and automatic classification of breast cancer,"

The Visual Computer, vol. 35, pp. 57-70, 2019.

[11] W. H. Organization. (2019, 14/01/2019). Lung cancer fact sheet Available:

http://gco.iarc.fr/today/data/factsheets/cancers/15-Lung-fact-sheet.pdf

[12] W. H. Organization. (2019, 14/01/2019). Cancer facts of Pakistan Available:

http://gco.iarc.fr/today/online-analysis-pie

[13] M. K. Gould, J. Fletcher, M. D. Iannettoni, W. R. Lynch, D. E. Midthun, D. P.

Naidich, et al., "Evaluation of patients with pulmonary nodules: when is it

lung cancer? ACCP evidence-based clinical practice guidelines," Chest

Journal, vol. 132, pp. 108S-130S, 2007.

[14] I. Sluimer, A. Schilham, M. Prokop, and B. van Ginneken, "Computer analysis

of computed tomography scans of the lung: a survey," IEEE Transactions on

Medical Imaging, vol. 25, pp. 385-405, 2006.

[15] F. Girvin and J. P. Ko, "Pulmonary nodules: detection, assessment, and CAD,"

American Journal of Roentgenology, vol. 191, pp. 1057-1069, 2008.

[16] J. L. Causey, J. Zhang, S. Ma, B. Jiang, J. A. Qualls, D. G. Politte, et al.,

"Highly accurate model for prediction of lung nodule malignancy with CT

scans," Scientific Reports, vol. 8, p. 9286, 2018.

[17] C. Tao, D. S. Gierada, F. Zhu, T. K. Pilgram, J. H. Wang, and K. T. Bae,

"Automated matching of pulmonary nodules: evaluation in serial screening

chest CT," American Journal of Roentgenology, vol. 192, pp. 624-628, 2009.

http://gco.iarc.fr/today/data/factsheets/cancers/15-Lung-fact-sheet.pdf

http://gco.iarc.fr/today/online-analysis-pie

133

[18] J. Pu, J. Roos, A. Y. Chin, S. Napel, G. D. Rubin, and D. S. Paik, "Adaptive

border marching algorithm: automatic lung segmentation on chest CT

images," Computerized Medical Imaging and Graphics, vol. 32, pp. 452-462,

2008.

[19] H. L. Neville, A. R. Hogan, Y. Zhuge, E. A. Perez, M. C. Cheung, L. G.

Koniaris, et al., "Incidence and outcomes of malignant pediatric lung

neoplasms," Journal of Surgical Research, vol. 156, pp. 224-230, 2009.

[20] I. Petkovska, M. S. Brown, J. G. Goldin, H. J. Kim, M. F. McNitt-Gray, F. G.

Abtin, et al., "The effect of lung volume on nodule size on CT," Academic

Radiology, vol. 14, pp. 476-485, 2007.

[21] C. B. Weldon and R. C. Shamberger, "Pediatric pulmonary tumors: primary

and metastatic," in Seminars in Pediatric Surgery, 2008, pp. 17-29.

[22] I. A. Cunningham, "Computed tomography," The Biomedical Engineering

Handbook, 1998.

[23] M. C. Ellis, C. J. Hessman, R. Weerasinghe, P. H. Schipper, and J. T. Vetto,

"Comparison of pulmonary nodule detection rates between preoperative CT

imaging and intraoperative lung palpation," The American Journal of Surgery,

vol. 201, pp. 619-622, 2011.

[24] D. LLC. (22/12/2014). "CT Scan vs PET Scan." Available:

http://www.diffen.com/difference/CT_Scan_vs_PET_Scan

[25] K. Awai, K. Murao, A. Ozawa, M. Komi, H. Hayakawa, S. Hori, et al.,

"Pulmonary Nodules at Chest CT: Effect of Computer-aided Diagnosis on

Radiologists‘ Detection Performance 1," Radiology, vol. 230, pp. 347-352,

2004.

[26] S. Copley, T. Bryant, A. Chambers, C. Harvey, J. Hodson, A. Graham, et al.,

"Observer accuracy in the detection of pulmonary nodules on CT: effect of

cine frame rate," Clinical Radiology, vol. 65, pp. 133-136, 2010.

http://www.diffen.com/difference/CT_Scan_vs_PET_Scan

134

[27] H. M. Marshall, R. V. Bowman, I. A. Yang, K. M. Fong, and C. D. Berg,

"Screening for lung cancer with low-dose computed tomography: a review of

current status," Journal of Thoracic Disease, vol. 5, p. S524, 2013.

[28] A. El-Baz, A. Farag, G. Gimel‘farb, R. Falk, and M. A. El-Ghar, "A novel

level set-based computer-aided detection system for automatic detection of

lung nodules in low dose chest computed tomography scans," in Lung Imaging

and Computer Aided Diagnosis, ed, 2011, pp. 221-238.

[29] A. Reeves, A. Biancardi, D. Yankelevitz, S. Fotin, B. Keller, A. Jirapatnakul,

et al., "A public image database to support research in computer aided

diagnosis," in Engineering in Medicine and Biology Society, 2009. EMBC

2009. Annual International Conference of the IEEE, 2009, pp. 3715-3718.

[30] X. Yi, E. Walia, and P. Babyn, "Generative adversarial network in medical

imaging: A review," Medical Image Analysis, vol. 58, p. 101552, 2019.

[31] S. G. Armato III, G. McLennan, M. F. McNitt-Gray, C. R. Meyer, D.

Yankelevitz, D. R. Aberle, et al., "Lung image database consortium:

Developing a resource for the medical imaging research community 1,"

Radiology, vol. 232, pp. 739-748, 2004.

[32] N. C. Institute. (2018, 24/01/2018). LIDC archive. Available:

https://www.cancerimagingarchive.net

[33] M. F. McNitt-Gray, S. G. Armato III, C. R. Meyer, A. P. Reeves, G.

McLennan, R. C. Pais, et al., "The Lung Image Database Consortium (LIDC)

data collection process for nodule detection and annotation," Academic

Radiology, vol. 14, pp. 1464-1474, 2007.

[34] S. G. Armato III, G. McLennan, L. Bidaut, M. F. McNitt-Gray, C. R. Meyer,

A. P. Reeves, et al., "The lung image database consortium (LIDC) and image

database resource initiative (IDRI): a completed reference database of lung

nodules on CT scans," Medical Physics, vol. 38, pp. 915-931, 2011.

http://www.cancerimagingarchive.net/

135

[35] A. P. Reeves, A. M. Biancardi, T. V. Apanasovich, C. R. Meyer, H.

MacMahon, E. J. van Beek, et al., "The Lung Image Database Consortium

(LIDC): a comparison of different size metrics for pulmonary nodule

measurements," Academic Radiology, vol. 14, pp. 1475-1485, 2007.

[36] J. Zhang, Y. Xia, H. Cui, and Y. Zhang, "Pulmonary nodule detection in

medical images: a survey," Biomedical Signal Processing and Control, vol.

43, pp. 138-147, 2018.

[37] T. Fawcett, "An introduction to ROC analysis," Pattern Recognition Letters,

vol. 27, pp. 861-874, 2006.

[38] A. P. Bradley, "The use of the area under the ROC curve in the evaluation of

machine learning algorithms," Pattern Recognition, vol. 30, pp. 1145-1159,

1997.

[39] C. Buissink, J. Thompson, M. Voet, A. Sanderud, L. Kamping, L. Savary, et

al., "The influence of experience and training in a group of novice observers:

A jackknife alternative free-response receiver operating characteristic

analysis," Radiography, vol. 20, pp. 300-305, 2014.

[40] M. Sprindzuk, V. Kovalev, E. Snezhko, and S. Kharuzhyk, "Lung cancer

differential diagnosis based on the computer assisted radiology: The state of

the art," Polish Journal of Radiology, vol. 75, p. 67, 2010.

[41] K. Doi, "Computer-aided diagnosis in medical imaging: historical review,

current status and future potential," Computerized Medical Imaging and

Graphics, vol. 31, pp. 198-211, 2007.

[42] G. S. Lodwick, T. E. Keats, and J. P. Dorst, "The coding of roentgen images

for computer analysis as applied to lung cancer," Radiology, vol. 81, pp. 185-

200, 1963.

136

[43] A. Hosny, C. Parmar, J. Quackenbush, L. H. Schwartz, and H. J. Aerts,

"Artificial intelligence in radiology," Nature Reviews Cancer, vol. 18, p. 500,

2018.

[44] V. Goh, S. Halligan, A. Gharpuray, D. Wellsted, J. Sundin, and C. I. Bartram,

"Quantitative Assessment of Colorectal Cancer Tumor Vascular Parameters

by Using Perfusion CT: Influence of Tumor Region of Interest 1," Radiology,

vol. 247, pp. 726-732, 2008.

[45] S. M Naqi and M. Sharif, "Recent Developments in Computer Aided

Diagnosis for Lung Nodule Detection from CT images: A Review," Current

Medical Imaging Reviews, vol. 13, pp. 3-19, 2017.

[46] S. Sun, C. Bauer, and R. Beichel, "Automated 3-D segmentation of lungs with

lung cancer in CT data using a novel robust active shape model approach,"

IEEE Transactions on Medical Imaging, vol. 31, pp. 449-460, 2011.

[47] Y. Yim and H. Hong, "Correction of segmented lung boundary for inclusion

of pleural nodules and pulmonary vessels in chest CT images," Computers in

Biology and Medicine, vol. 38, pp. 845-857, 2008.

[48] M. Tan, R. Deklerck, B. Jansen, M. Bister, and J. Cornelis, "A novel

computer‐aided lung nodule detection system for CT images," Medical

Physics, vol. 38, pp. 5630-5645, 2011.

[49] J. Dehmeshki, X. Ye, X. Lin, M. Valdivieso, and H. Amin, "Automated

detection of lung nodules in CT images using shape-based genetic algorithm,"

Computerized Medical Imaging and Graphics, vol. 31, pp. 408-417, 2007.

[50] S. Mattoccia, F. Tombari, and L. Di Stefano, "Efficient template matching for

multi-channel images," Pattern Recognition Letters, vol. 32, pp. 694-700,

2011.

[51] S. Ozekes, O. Osman, and O. N. Ucan, "Nodule detection in a lung region

that's segmented with using genetic cellular neural networks and 3D template

137

matching with fuzzy rule based thresholding," Korean Journal of Radiology,

vol. 9, pp. 1-9, 2008.

[52] H. Li, Y. Wang, K. R. Liu, S.-C. Lo, and M. T. Freedman, "Computerized

radiographic mass detection. I. Lesion site selection by morphological

enhancement and contextual segmentation," IEEE Transactions on Medical

Imaging, vol. 20, pp. 289-301, 2001.

[53] S. Ukil and J. M. Reinhardt, "Anatomy-guided lung lobe segmentation in X-

ray CT images," IEEE Transactions on Medical Imaging, vol. 28, pp. 202-

214, 2009.

[54] W.-J. Choi and T.-S. Choi, "Automated pulmonary nodule detection based on

three-dimensional shape-based feature descriptor," Computer Methods and

Programs in Biomedicine, vol. 113, pp. 37-54, 2014.

[55] E. Taşcı and A. Uğur, "Shape and texture based novel features for automated

juxtapleural nodule detection in lung cts," Journal of Medical Systems, vol. 39,

p. 46, 2015.

[56] A. Riccardi, T. S. Petkov, G. Ferri, M. Masotti, and R. Campanini,

"Computer-aided detection of lung nodules via 3D fast radial transform, scale

space representation, and Zernike MIP classification," Medical Physics, vol.

38, pp. 1962-1971, 2011.

[57] X. Ye, X. Lin, J. Dehmeshki, G. Slabaugh, and G. Beddoe, "Shape-based

computer-aided detection of lung nodules in thoracic CT images," IEEE

Transactions on Biomedical Engineering, vol. 56, pp. 1810-1820, 2009.

[58] S. Mukhopadhyay, "A Segmentation Framework of Pulmonary Nodules in

Lung CT Images," Journal of Digital Imaging, vol. 29, pp. 86-103, 2016.

[59] S. G. Armato III, M. L. Giger, and H. MacMahon, "Automated detection of

lung nodules in CT scans: preliminary results," Medical Physics, vol. 28, pp.

1552-1561, 2001.

138

[60] D. S. Elizabeth, H. K. Nehemiah, C. Raj, and A. Kannan, "A novel

segmentation approach for improving diagnostic accuracy of CAD systems for

detecting lung cancer from chest computed tomography images," Journal of

Data and Information Quality (JDIQ), vol. 3, p. 4, 2012.

[61] H. Han, L. Li, F. Han, B. Song, W. Moore, and Z. Liang, "Fast and adaptive

detection of pulmonary nodules in thoracic CT images using a hierarchical

vector quantization scheme," IEEE Journal of Biomedical and Health

Informatics, vol. 19, pp. 648-659, 2015.

[62] S. Armato and H. MacMahon, "Automated lung segmentation and computer-

aided diagnosis for thoracic CT scans," in International Congress Series,

2003, pp. 977-982.

[63] A. P. Reeves and W. J. Kostis, "Computer-aided diagnosis of small pulmonary

nodules," in Seminars in Ultrasound, CT and MRI, 2000, pp. 116-128.

[64] N. Duggan, E. Bae, S. Shen, W. Hsu, A. Bui, E. Jones, et al., "A technique for

lung nodule candidate detection in CT using global minimization methods," in

International Workshop on Energy Minimization Methods in Computer Vision

and Pattern Recognition, 2015, pp. 478-491.

[65] H. Liu and L. Yu, "Toward integrating feature selection algorithms for

classification and clustering," IEEE Transactions on Knowledge and Data

Engineering, vol. 17, pp. 491-502, 2005.

[66] Y. Saeys, I. Inza, and P. Larrañaga, "A review of feature selection techniques

in bioinformatics," Bioinformatics, vol. 23, pp. 2507-2517, 2007.

[67] S. B. Kotsiantis, I. Zaharakis, and P. Pintelas, "Supervised machine learning:

A review of classification techniques," ed, 2007.

[68] B. Schölkopf and A. J. Smola, Learning with kernels: support vector

machines, regularization, optimization, and beyond: MIT press, 2002.

139

[69] C. M. Bishop, Pattern recognition and machine learning vol. 1: springer New

York, 2006.

[70] J. Jiang, P. Trundle, and J. Ren, "Medical image analysis with artificial neural

networks," Computerized Medical Imaging and Graphics, vol. 34, pp. 617-

631, 2010.

[71] A. Khosravi, S. Nahavandi, D. Creighton, and A. F. Atiya, "Comprehensive

review of neural network-based prediction intervals and new advances," IEEE

Transactions on Neural Networks, vol. 22, pp. 1341-1356, 2011.

[72] T. G. Dietterich, "Ensemble methods in machine learning," in Multiple

Cassifier Systems, ed: Springer, 2000, pp. 1-15.

[73] B. van Ginneken, S. G. Armato III, B. de Hoop, S. van Amelsvoort-van de

Vorst, T. Duindam, M. Niemeijer, et al., "Comparing and combining

algorithms for computer-aided detection of pulmonary nodules in computed

tomography scans: the ANODE09 study," Medical Image Analysis, vol. 14,

pp. 707-722, 2010.

[74] S. G. Armato, M. L. Giger, C. J. Moran, J. T. Blackburn, K. Doi, and H.

MacMahon, "Computerized Detection of Pulmonary Nodules on CT Scans 1,"

Radiographics, vol. 19, pp. 1303-1311, 1999.

[75] Y. Lee, T. Hara, H. Fujita, S. Itoh, and T. Ishigaki, "Automated detection of

pulmonary nodules in helical CT images based on an improved template-

matching technique," IEEE Transactions on Medical Imaging, vol. 20, pp.

595-604, 2001.

[76] S. G. Armato, F. Li, M. L. Giger, H. MacMahon, S. Sone, and K. Doi, "Lung

Cancer: Performance of Automated Lung Nodule Detection Applied to

Cancers Missed in a CT Screening Program 1," Radiology, vol. 225, pp. 685-

692, 2002.

140

[77] M. N. Gurcan, B. Sahiner, N. Petrick, H.-P. Chan, E. A. Kazerooni, P. N.

Cascade, et al., "Lung nodule detection on thoracic computed tomography

images: preliminary evaluation of a computer-aided diagnosis system,"

Medical Physics, vol. 29, pp. 2552-2558, 2002.

[78] D.-T. Lin and C.-R. Yan, "Lung nodules identification rules extraction with

neural fuzzy network," in Neural Information Processing, 2002. ICONIP'02.

Proceedings of the 9th International Conference on, 2002, pp. 2049-2053.

[79] B. Zhao, "Automatic detection of small lung nodules on CT utilizing a local

density maximum algorithm," Journal of Applied Clinical Medical Physics,

vol. 4, pp. 248-260, 2003.

[80] W. J. Kostis, A. P. Reeves, D. F. Yankelevitz, and C. I. Henschke, "Three-

dimensional segmentation and growth-rate estimation of small pulmonary

nodules in helical CT images," IEEE Transactions on Medical Imaging, vol.

22, pp. 1259-1274, 2003.

[81] K. T. Bae, J.-S. Kim, Y.-H. Na, K. G. Kim, and J.-H. Kim, "Pulmonary

Nodules: Automated Detection on CT Images with Morphologic Matching

Algorithm—Preliminary Results 1," Radiology, vol. 236, pp. 286-293, 2005.

[82] D.-T. Lin, C.-R. Yan, and W.-T. Chen, "Autonomous detection of pulmonary

nodules on CT images with a neural network-based fuzzy system,"


[83] X. Zhang, G. McLennan, E. A. Hoffman, and M. Sonka, "Automated

detection of small-size pulmonary nodules based on helical CT images," in

Information Processing in Medical Imaging, 2005, pp. 664-676.

[84] L. Boroczky, L. Zhao, and K. P. Lee, "Feature subset selection for improving

the performance of false positive reduction in lung nodule CAD," IEEE

Transactions on Information Technology in Biomedicine, vol. 10, pp. 504-511,

2006.

141

[85] O. Osman, S. Ozekes, and O. N. Ucan, "Lung nodule diagnosis using 3D

template matching," Computers in Biology and Medicine, vol. 37, pp. 1167-

1172, 2007.

[86] A. Retico, P. Delogu, M. E. Fantacci, I. Gori, and A. P. Martinez, "Lung

nodule detection in low-dose and thin-slice computed tomography,"

Computers in Biology and Medicine, vol. 38, pp. 525-534, 2008.

[87] K. Murphy, B. van Ginneken, A. M. Schilham, B. De Hoop, H. Gietema, and

M. Prokop, "A large-scale evaluation of automatic pulmonary nodule

detection in chest CT using local image features and k-nearest-neighbour

classification," Medical Image Analysis, vol. 13, pp. 757-770, 2009.

[88] J. R. F. da Silva Sousa, A. C. Silva, A. C. de Paiva, and R. A. Nunes,

"Methodology for automatic detection of lung nodules in computerized

tomography images," Computer Methods and Programs in Biomedicine, vol.

98, pp. 1-14, 2010.

[89] T. Messay, R. C. Hardie, and S. K. Rogers, "A new computationally efficient

CAD system for pulmonary nodule detection in CT imagery," Medical Image

Analysis, vol. 14, pp. 390-406, 2010.

[90] S. L. A. Lee, A. Z. Kouzani, and E. J. Hu, "Random forest based lung nodule

classification aided by clustering," Computerized Medical Imaging and

Graphics, vol. 34, pp. 535-542, 2010.

[91] M. C. Lee, L. Boroczky, K. Sungur-Stasik, A. D. Cann, A. C. Borczuk, S. M.

Kawut, et al., "Computer-aided diagnosis of pulmonary nodules using a two-

step approach for feature selection and classifier ensemble construction,"

Artificial Intelligence in Medicine, vol. 50, pp. 43-53, 2010.

[92] H. Chen, W. Wu, H. Xia, J. Du, M. Yang, and B. Ma, "Classification of

pulmonary nodules using neural network ensemble," in Advances in Neural

Networks–ISNN 2011, ed: Springer, 2011, pp. 460-466.

142

[93] S. Magalhães Barros Netto, A. Corrêa Silva, R. Acatauassú Nunes, and M.

Gattass, "Automatic segmentation of lung nodules with growing neural gas

and support vector machine," Computers in Biology and Medicine, vol. 42, pp.

1110-1121, 2012.

[94] D. Cascio, R. Magro, F. Fauci, M. Iacomi, and G. Raso, "Automatic detection

of lung nodules in CT datasets based on stable 3D mass–spring models,"


[95] H. Chen, J. Zhang, Y. Xu, B. Chen, and K. Zhang, "Performance comparison

of artificial neural network and logistic regression model for differentiating

lung nodules on CT scans," Expert Systems with Applications, vol. 39, pp.

11503-11509, 2012.

[96] W.-J. Choi and T.-S. Choi, "Genetic programming-based feature transform

and classification for the automatic detection of pulmonary nodules on

computed tomography images," Information Sciences, vol. 212, pp. 57-78,

2012.

[97] A. Teramoto and H. Fujita, "Fast lung nodule detection in chest CT images

using cylindrical nodule-enhancement filter," International Journal of

Computer Assisted Radiology and Surgery, vol. 2, pp. 193-205, 2013.

[98] W.-J. Choi and T.-S. Choi, "Automated pulmonary nodule detection system in

computed tomography images: a hierarchical block classification approach,"

Entropy, vol. 15, pp. 507-523, 2013.

[99] Q. Wang, W. Kang, C. Wu, and B. Wang, "Computer-aided detection of lung

nodules by SVM based on 3D matrix patterns," Clinical Imaging, vol. 37, pp.

62-69, 2013.

[100] A. Tartar, N. Kilic, and A. Akan, "A new method for pulmonary nodule

detection using decision trees," in Engineering in Medicine and Biology

143

Society (EMBC), 2013 35th Annual International Conference of the IEEE,

2013, pp. 7355-7359.

[101] M. Keshani, Z. Azimifar, F. Tajeripour, and R. Boostani, "Lung nodule

segmentation and recognition using SVM classifier and active contour

modeling: A complete intelligent system," Computers in Biology and

Medicine, vol. 43, pp. 287-300, 2013.

[102] H.-s. Jang, W.-J. Choi, and T.-S. Choi, "Optimal Fuzzy Rule based Pulmonary

Nodule Detection," Advanced Science and Technology Letters, vol. 29, pp. 75-

78, 2013.

[103] J. Kuruvilla and K. Gunavathi, "Lung cancer classification using neural

networks for CT images," Computer Methods and Programs in Biomedicine,

vol. 113, pp. 202-209, 2014.

[104] P. Cao, J. Yang, W. Li, D. Zhao, and O. Zaiane, "Ensemble-based hybrid

probabilistic sampling for imbalanced data learning in lung nodule CAD,"


[105] A. O. de Carvalho Filho, W. B. de Sampaio, A. C. Silva, A. C. de Paiva, R. A.

Nunes, and M. Gattass, "Automatic detection of solitary lung nodules using

quality threshold clustering, genetic algorithm and diversity index," Artificial

Intelligence in Medicine, vol. 60, pp. 165-177, 2014.

[106] P. Badura and E. Pietka, "Soft computing approach to 3D lung nodule

segmentation in CT," Computers in Biology and Medicine, vol. 53, pp. 230-

243, 2014.

[107] M. S. Brown, P. Lo, J. G. Goldin, E. Barnoy, G. H. J. Kim, M. F. McNitt-

Gray, et al., "Toward clinically usable CAD for lung cancer screening with

computed tomography," European Radiology, vol. 24, pp. 2719-2728, 2014.

144

[108] E. Taşcı and A. Uğur, "Shape and texture based novel features for automated

juxtapleural nodule detection in lung cts," Journal of Medical Systems, vol. 39,

pp. 1-13, 2015.

[109] S. Akram, M. Y. Javed, U. Qamar, A. Khanum, and A. Hassan, "Artificial

Neural Network based Classification of Lungs Nodule using Hybrid Features

from Computerized Tomographic Images," Applied Mathematics &

Information Sciences, vol. 9, pp. 183-195, 2015.

[110] B. Wang, X. Tian, Q. Wang, Y. Yang, H. Xie, S. Zhang, et al., "Pulmonary

nodule detection in CT images based on shape constraint CV model," Medical

Physics, vol. 42, pp. 1241-1254, 2015.

[111] A. Kaya and A. B. Can, "A weighted rule based method for predicting

malignancy of pulmonary nodules by nodule characteristics," Journal of

Biomedical Informatics, vol. 56, pp. 69-79, 2015.

[112] Z. Shi, B. Xu, M. Zhao, J. Zhao, Y. Wang, Y. Liu, et al., "A joint ROI

extraction filter for computer aided lung nodule detection," Bio-Medical

Materials and Engineering, vol. 26, pp. 1491-1499, 2015.

[113] S. Dai, K. Lu, J. Dong, Y. Zhang, and Y. Chen, "A novel approach of lung

segmentation on chest CT images using graph cuts," Neurocomputing, vol.

168, pp. 799-807, 2015.

[114] H. Han, L. Li, F. Han, B. Song, W. Moore, and Z. Liang, "Fast and adaptive

detection of pulmonary nodules in thoracic CT images using a hierarchical

vector quantization scheme,"IEEE Journal of Biomedical and Health

Informatics, vol. 19, pp. 648-659, 2015.

[115] S. Shen, A. A. Bui, J. Cong, and W. Hsu, "An automated lung segmentation

approach using bidirectional chain codes to improve nodule detection

accuracy," Computers in Biology and Medicine, vol. 57, pp. 139-149, 2015.

145

[116] K.-L. Hua, C.-H. Hsu, S. C. Hidayati, W.-H. Cheng, and Y.-J. Chen,

"Computer-aided classification of lung nodules on computed tomography

images via deep learning technique," OncoTargets and Therapy, vol. 8, 2015.

[117] M. Firmino, G. Angelo, H. Morais, M. R. Dantas, and R. Valentim,

"Computer-aided detection (CADe) and diagnosis (CADx) system for lung

cancer with likelihood of malignancy," Biomedical Engineering Online, vol.

15, p. 2, 2016.

[118] A. Nibali, Z. He, and D. Wollersheim, "Pulmonary nodule classification with

deep residual networks," International Journal of Computer Assisted

Radiology and Surgery, pp. 1-10, 2017.

[119] Q. Dou, H. Chen, L. Yu, J. Qin, and P.-A. Heng, "Multilevel contextual 3-d

cnns for false positive reduction in pulmonary nodule detection," IEEE

Transactions on Biomedical Engineering, vol. 64, pp. 1558-1567, 2017.

[120] G. L. da Silva, O. P. da Silva Neto, A. C. Silva, A. C. de Paiva, and M.

Gattass, "Lung nodules diagnosis based on evolutionary convolutional neural

network," Multimedia Tools and Applications, vol. 76, pp. 19039-19055,

2017.

[121] W. Sun, B. Zheng, and W. Qian, "Automatic feature learning using

multichannel ROI based on deep structured algorithms for computerized lung

cancer diagnosis," Computers in Biology and Medicine, vol. 89, pp. 530-539,

2017.

[122] A. Gupta, T. Saar, O. Martens, and Y. L. Moullec, "Automatic detection of

multisize pulmonary nodules in CT images: Large‐scale validation of the

false‐positive reduction step," Medical Physics, vol. 45, pp. 1135-1149, 2018.

[123] W. Zhang, X. Wang, X. Li, and J. Chen, "3D skeletonization feature based

computer-aided detection system for pulmonary nodules in CT datasets,"


146

[124] M. A. Jaffar, A. B. Siddiqui, and M. Mushtaq, "Ensemble classification of

pulmonary nodules using gradient intensity feature descriptor and differential

evolution," Cluster Computing, vol. 21, pp. 393-407, 2018.

[125] A. Teramoto and H. Fujita, "Automated Lung Nodule Detection Using

Positron Emission Tomography/Computed Tomography," in Artificial

Intelligence in Decision Support Systems for Diagnosis in Medical Imaging,

ed: Springer, 2018, pp. 87-110.

[126] J. Gong, J.-y. Liu, L.-j. Wang, X.-w. Sun, B. Zheng, and S.-d. Nie, "Automatic

detection of pulmonary nodules in CT images by incorporating 3D tensor

filtering with local image feature analysis," Physica Medica, vol. 46, pp. 124-

133, 2018.

[127] I. Ali, G. Hart, G. Gunabushanam, Y. Liang, W. Muhammad, B. Nartowt, et

al., "lung nodule Detection via Deep reinforcement learning," Frontiers in

Oncology, vol. 8, p. 108, 2018.

[128] H. Jiang, H. Ma, W. Qian, M. Gao, and Y. Li, "An automatic detection system

of lung nodule based on multigroup patch-based deep learning network," IEEE

Journal of Biomedical and Health Informatics, vol. 22, pp. 1227-1237, 2018.

[129] C. Li, G. Zhu, X. Wu, and Y. Wang, "False-Positive Reduction on Lung

Nodules Detection in Chest Radiographs by Ensemble of Convolutional

Neural Networks," IEEE Access, vol. 6, pp. 16060-16067, 2018.

[130] R. Huidrom, Y. J. Chanu, and K. M. Singh, "Pulmonary nodule detection on

computed tomography using neuro-evolutionary scheme," Signal, Image and

Video Processing, vol. 13, pp. 53-60, 2019.

[131] F. Shaukat, G. Raja, R. Ashraf, S. Khalid, M. Ahmad, and A. Ali, "Artificial

neural network based classification of lung nodules in CT images using

intensity, shape and texture features," Journal of Ambient Intelligence and

Humanized Computing, vol. 10, pp. 4135-4149, 2019.

147

[132] H. Xie, D. Yang, N. Sun, Z. Chen, and Y. Zhang, "Automated pulmonary

nodule detection in CT images using deep convolutional neural networks,"

Pattern Recognition, vol. 85, pp. 109-119, 2019.

[133] G. Kasinathan, S. Jayakumar, A. H. Gandomi, M. Ramachandran, S. J. Fong,

and R. Patan, "Automated 3-D lung tumor detection and classification by an

active contour model and CNN classifier," Expert Systems with Applications,

vol. 134, pp. 112-119, 2019.

[134] X. Feng, Z. Xiao, B. Zhong, Y. Dong, and J. Qiu, "Dynamic weighted

ensemble classification for credit scoring using Markov Chain," Applied

Intelligence, vol. 49, pp. 555-568, 2019.

[135] J. A. Hartigan and M. A. Wong, "Algorithm AS 136: A k-means clustering

algorithm," Applied Statistics, pp. 100-108, 1979.

[136] A. El-Baz, A. A. Farag, R. Falk, and R. La Rocca, "A unified approach for

detection, visualization, and identification of lung abnormalities in chest spiral

CT scans," in International Congress Series, 2003, pp. 998-1004.

[137] N. R. Pal, K. Pal, J. M. Keller, and J. C. Bezdek, "A possibilistic fuzzy c-

means clustering algorithm," IEEE Transactions on Fuzzy Systems, vol. 13,

pp. 517-530, 2005.

[138] A. Tolouee, H. Abrishami Moghaddam, M. Forouzanfar, M. Gity, and R.

Garnavi, "Image based diagnostic aid system for interstitial lung diseases,"

Expert Systems with Applications, vol. 38, pp. 7755-7765, 2011.

[139] D. Weinland, R. Ronfard, and E. Boyer, "A survey of vision-based methods

for action representation, segmentation and recognition," Computer Vision and

Image Understanding, vol. 115, pp. 224-241, 2011.

[140] H. H. Jo, H. Hong, and J. M. Goo, "Pulmonary nodule registration in serial CT

scans using global rib matching and nodule template matching," Computers in

Biology and Medicine, vol. 45, pp. 87-97, 2014.

148

[141] L. A. Zadeh, "Fuzzy logic, neural networks, and soft computing,"

Communications of the ACM, vol. 37, pp. 77-84, 1994.

[142] M. Sezgin, "Survey over image thresholding techniques and quantitative

performance evaluation," Journal of Electronic Imaging, vol. 13, pp. 146-168,

2004.

[143] D. L. Pham, C. Xu, and J. L. Prince, "Current methods in medical image

segmentation 1," Annual Review of Biomedical Engineering, vol. 2, pp. 315-

337, 2000.

[144] F. Meyer and S. Beucher, "Morphological segmentation," Journal of Vsual

Communication and Image Representation, vol. 1, pp. 21-46, 1990.

[145] J. Lucaya, J. Piqueras, P. García-Peña, G. Enríquez, M. García-Macías, and J.

Sotil, "Low-dose high-resolution CT of the chest in children and young adults:

dose, cooperation, artifact incidence, and image quality," American Journal of

Roentgenology, vol. 175, pp. 985-992, 2000.

[146] A. Khotanzad and Y. H. Hong, "Invariant image recognition by Zernike

moments," IEEE Transactions on Pattern Analysis and Machine Intelligence,

vol. 12, pp. 489-497, 1990.

[147] S. Makridakis, "The forthcoming Artificial Intelligence (AI) revolution: Its

impact on society and firms," Futures, vol. 90, pp. 46-60, 2017.

[148] Y. LeCun, Y. Bengio, and G. Hinton, "Deep learning," Nature, vol. 521, p.

436, 2015.

[149] J. Schmidhuber, "Deep learning in neural networks: An overview," Neural

Networks, vol. 61, pp. 85-117, 2015.

[150] M.-T. Wu, J.-M. Chang, A. A. Chiang, J.-Y. Lu, H.-K. Hsu, W.-H. Hsu, et al.,

"Use of quantitative CT to predict postoperative lung function in patients with

lung cancer," Radiology, vol. 191, pp. 257-262, 1994.

149

[151] A. M. Reza, "Realization of the contrast limited adaptive histogram

equalization (CLAHE) for real-time image enhancement," Journal of VLSI

Signal Processing Systems for Signal, Image and Video Technology, vol. 38,

pp. 35-44, 2004.

[152] S. M. Pizer, R. E. Johnston, J. P. Ericksen, B. C. Yankaskas, and K. E. Muller,

"Contrast-limited adaptive histogram equalization: speed and effectiveness,"

in [1990] Proceedings of the First Conference on Visualization in Biomedical

Computing, 1990, pp. 337-345.

[153] M. Unser, "Sum and difference histograms for texture classification," IEEE

Transactions on Pattern Analysis and Machine Intelligence, vol. PAMI-8, pp.

118-125, 1986.

[154] A. K. Agrawala, A. V. Kulkarni, and E. H. Paul Jr, "System for extracting

shape features from an image," ed: Google Patents, 1980.

[155] N. Dalal and B. Triggs, "Histograms of oriented gradients for human

detection," in Computer Vision and Pattern Recognition, 2005. CVPR 2005.

IEEE Computer Society Conference on, 2005, pp. 886-893.

[156] A. M. Andrew, "An Introduction to Support Vector Machines and Other

Kernel-Based Learning Methods by Nello Christianini and John Shawe-

Taylor, Cambridge University Press, Cambridge, 2000, xiii+ 189 pp., ISBN 0-

521-78019-5 (Hbk,£ 27.50)," Robotica, vol. 18, pp. 687-689, 2000.

[157] M. Kass, A. Witkin, and D. Terzopoulos, "Snakes: Active contour models,"

International Journal of Computer Vision, vol. 1, pp. 321-331, 1988.

[158] R. M. Haralick and K. Shanmugam, "Textural features for image

classification," IEEE Transactions on Systems, Man, and Cybernetics, pp.

610-621, 1973.

150

[159] S. M. Naqi, M. Sharif, and M. Yasmin, "Multistage segmentation model and

SVM-ensemble for precise lung nodule detection," International Journal of

Computer Assisted Radiology and Surgery, vol. 7, pp. 1083-1095, 2018.

[160] G.-Z. Li, H.-L. Bu, M. Q. Yang, X.-Q. Zeng, and J. Y. Yang, "Selecting

subsets of newly extracted features from PCA and PLS in microarray data

analysis," BMC Genomics, vol. 9, p. S24, 2008.

[161] T. Cover and P. Hart, "Nearest neighbor pattern classification," IEEE

Transactions on Information Theory, vol. 13, pp. 21-27, 1967.

[162] C. Elkan, "Boosting and naive Bayesian learning," in Proceedings of the

International Conference on Knowledge Discovery and Data Mining, 1997.

[163] T. G. Dietterich, "An experimental comparison of three methods for

constructing ensembles of decision trees: Bagging, boosting, and

randomization," Machine Learning, vol. 40, pp. 139-157, 2000.

[164] G. M. Foody and A. Mathur, "A relative evaluation of multiclass image

classification by support vector machines," IEEE Transactions on Geoscience

and Remote Sensing, vol. 42, pp. 1335-1343, 2004.

[165] B. K. Samanthula, Y. Elmehdwi, and W. Jiang, "K-nearest neighbor

classification over semantically secure encrypted relational data," IEEE

Transactions on Knowledge and Data Engineering, vol. 27, pp. 1261-1273,

2015.

[166] H. Zhang, "Exploring conditions for the optimality of naive Bayes,"

International Journal of Pattern Recognition and Artificial Intelligence, vol.

19, pp. 183-198, 2005.

[167] G. Rätsch, T. Onoda, and K.-R. Müller, "Soft margins for AdaBoost,"

Machine Learning, vol. 42, pp. 287-320, 2001.

151

[168] E. Cuevas, D. Zaldivar, and M. Pérez-Cisneros, "A novel multi-threshold

segmentation approach based on differential evolution optimization," Expert

Systems with Applications, vol. 37, pp. 5265-5271, 2010.

[169] A. K. Dhara, S. Mukhopadhyay, S. A. Mehre, N. Khandelwal, N. Prabhakar,

M. Garg, et al., "A Study of Retrieval Accuracy of Pulmonary Nodules Based

on External Attachment," in SPIE Medical Imaging, 2017, pp. 101343T-

101343T-6.

[170] J. Canny, "A computational approach to edge detection," IEEE Transactions

on Pattern Analysis and Machine Intelligence, vol. 8, pp. 679-698, 1986.

[171] F. Lafarge and X. Descombes, "Geometric feature extraction by a

multimarked point process," IEEE Transactions on Pattern Analysis &

Machine Intelligence, vol. 32, pp. 1597-1609, 2010.

[172] F. Yang, Y.-Y. Xu, S.-T. Wang, and H.-B. Shen, "Image-based classification

of protein subcellular location patterns in human reproductive tissue by

ensemble learning global and local features," Neurocomputing, vol. 131, pp.

113-123, 2014.

[173] M. Couceiro and P. Ghamisi, "Fractional-order Darwinian PSO," in

Fractional Order Darwinian Particle Swarm Optimization, ed: Springer,

2016, pp. 11-20.

[174] W. Gander, G. H. Golub, and R. Strebel, "Least-squares fitting of circles and

ellipses," BIT Numerical Mathematics, vol. 34, pp. 558-578, 1994.

[175] F. Jiang, A. Grigorev, S. Rho, Z. Tian, Y. Fu, W. Jifara, et al., "Medical image

semantic segmentation based on deep learning," Neural Computing and

Applications, vol. 29, pp. 1257-1265, 2018.

[176] X. Yang, X. Shen, J. Long, and H. Chen, "An improved median-based Otsu

image thresholding algorithm," Aasri Procedia, vol. 3, pp. 468-473, 2012.

152

[177] S. Arora, J. Acharya, A. Verma, and P. K. Panigrahi, "Multilevel thresholding

for image segmentation through a fast statistical recursive algorithm," Pattern

Recognition Letters, vol. 29, pp. 119-125, 2008.

[178] J. Yu, S. Wang, and L. Xi, "Evolving artificial neural networks using an

improved PSO and DPSO," Neurocomputing, vol. 71, pp. 1054-1060, 2008.

[179] M. Couceiro and S. Sivasundaram, "Novel fractional order particle swarm

optimization," Applied Mathematics and Computation, vol. 283, pp. 36-54,

2016.

[180] M. S. Couceiro, R. P. Rocha, N. F. Ferreira, and J. T. Machado, "Introducing

the fractional-order Darwinian PSO," Signal, Image and Video Processing,

vol. 6, pp. 343-350, 2012.

[181] B. Žalik, D. Mongus, N. Lukač, and K. R. Žalik, "Efficient chain code

compression with interpolative coding," Information Sciences, vol. 439, pp.

39-49, 2018.

[182] A. Ananda, I. Purnama, and M. Purnomo, "CT Lung Image filtering based on

Max-Tree method," in Instrumentation, Communications, Information

Technology, and Biomedical Engineering (ICICI-BME), 2011 2nd

International Conference on, 2011, pp. 137-140.

[183] L. Moraru, S. Moldovanu, and A. Biswas, "Optimization of breast lesion

segmentation in texture feature space approach," Medical Engineering &

Physics, vol. 36, pp. 129-135, 2014.

[184] F. Han, H. Wang, G. Zhang, H. Han, B. Song, L. Li, et al., "Texture feature

analysis for computer-aided diagnosis on pulmonary nodules," Journal of

Digital Imaging, vol. 28, pp. 99-115, 2015.

[185] D. Bibicu, L. Moraru, and A. Biswas, "Thyroid nodule recognition based on

feature selection and pixel classification methods," Journal of Digital

Imaging, vol. 26, pp. 119-128, 2013.

153

[186] F. Han, G. Zhang, H. Wang, B. Song, H. Lu, D. Zhao, et al., "A texture

feature analysis for diagnosis of pulmonary nodules using LIDC-IDRI

database," in Medical Imaging Physics and Engineering (ICMIPE), 2013,

IEEE International Conference on, 2013, pp. 14-18.

[187] A. Farag, A. Ali, J. Graham, A. Farag, S. Elshazly, and R. Falk, "Evaluation of

geometric feature descriptors for detection and classification of lung nodules

in low dose CT scans of the chest," in Biomedical Imaging: From Nano to

Macro, 2011 IEEE International Symposium on, 2011, pp. 169-172.

[188] J. Yang, J.-y. Yang, D. Zhang, and J.-f. Lu, "Feature fusion: parallel strategy

vs. serial strategy," Pattern Recognition, vol. 36, pp. 1369-1381, 2003.

[189] D. Sudha and M. Ramakrishna, "Comparative Study of Features Fusion

Techniques," in Recent Advances in Electronics and Communication

Technology (ICRAECT), 2017 International Conference on, 2017, pp. 235-

239.

[190] L. Meng, S. Ding, N. Zhang, and J. Zhang, "Research of stacked denoising

sparse autoencoder," Neural Computing and Applications, vol. 30, pp. 2083-

2100, 2018.

[191] M. F. Møller, "A scaled conjugate gradient algorithm for fast supervised

learning," Neural Networks, vol. 6, pp. 525-533, 1993.

[192] V. A. Moyer, "Screening for lung cancer: US Preventive Services Task Force

recommendation statement," Annals of Internal Medicine, vol. 160, pp. 330-

338, 2014.

[193] C. Jacobs, E. M. Rikxoort, K. Murphy, M. Prokop, C. M. Schaefer-Prokop,

and B. Ginneken, "Computer-aided detection of pulmonary nodules: a

comparative study using the public LIDC/IDRI database," European

Radiology, vol. 26, pp. 2139-2147, 2016.

154

[194] K. McGuinness and N. E. O‘connor, "A comparative evaluation of interactive

segmentation algorithms," Pattern Recognition, vol. 43, pp. 434-444, 2010.

[195] R. Unnikrishnan, C. Pantofaru, and M. Hebert, "Toward objective evaluation

of image segmentation algorithms," IEEE Transactions on Pattern Analysis &

Machine Intelligence, vol. 29, pp. 929-944, 2007.

[196] M. Mignotte, "A label field fusion model with a variation of information

estimator for image segmentation," Information Fusion, vol. 20, pp. 7-20,

2014.

[197] M. A. Jaffar, A. Hussain, and A. M. Mirza, "Fuzzy entropy based optimization

of clusters for the segmentation of lungs in CT scanned images," Knowledge

and Information Systems, vol. 24, pp. 91-111, 2010.

[198] X. Liao, J. Zhao, C. Jiao, L. Lei, Y. Qiang, and Q. Cui, "A segmentation

method for lung parenchyma image sequences based on superpixels and a self-

generating neural forest," PloS One, vol. 11, p. e0160556, 2016.

[199] V. Vapnik and R. Izmailov, "Knowledge transfer in SVM and neural

networks," Annals of Mathematics and Artificial Intelligence, vol. 81, pp. 3-

19, 2017.

[200] S. Zhang, X. Li, M. Zong, X. Zhu, and D. Cheng, "Learning k for knn

classification," ACM Transactions on Intelligent Systems and Technology

(TIST), vol. 8, p. 43, 2017.

[201] X. Liu, R. Lu, J. Ma, L. Chen, and B. Qin, "Privacy-preserving patient-centric

clinical decision support system on naive Bayesian classification," IEEE

Journal of Biomedical and Health Informatics, vol. 20, pp. 655-668, 2016.

[202] S. R. Safavian and D. Landgrebe, "A survey of decision tree classifier

methodology," IEEE transactions on Systems, Man, and Cybernetics, vol. 21,

pp. 660-674, 1991.

155

List of Publications

1. Syed Muhammad Naqi, Muhammad Sharif, Ikram Ullah Lali (2019) "A 3D

nodule candidate detection method supported by hybrid features to reduce

false positives in lung nodule detection," Multimedia Tools and Applications,

vol. 78(18), pp. 26287-26311, (Impact Factor: 2.1)

2. Syed Muhammad Naqi, Muhammad Sharif, Arfan Jaffar (2018) ―Lung

Nodule Detection and Classification based on Geometric Fit in Parametric

Form and Deep Learning,‖ Neural Computing and Applications.

https://doi.org/10.1007/s00521-018-3773-x (Impact Factor: 4.66), published

online.

3. Syed Muhammad Naqi, Muhammad Sharif, Mussarat Yasmin, (2018)

―Multistage Segmentation Model and SVM-Ensemble for Precise Lung

Nodule Detection,‖ International Journal of Computer Assisted Radiology and

Surgery, 13(7) 1083-1095 (Impact Factor: 2.15)

4. Syed Muhammad Naqi, Muhammad Sharif, Mussarat Yasmin, Steven

Lawrence Fernandes (2018) ―Lung Nodule Detection Using Polygon

Approximation and Hybrid Features from Lung CT Images,‖ Current Medical

Imaging Reviews, 14(1) 108-117 (Impact Factor: 0.53)

5. Syed Muhammad Naqi and Muhammad Sharif (2017) "Recent

Developments in Computer Aided Diagnosis for Lung Nodule Detection from

CT images: A Review," Current Medical Imaging Reviews, 13(1) 3-19.

(Impact Factor: 0.53)

156

Appendix:

WHO Cancer Statistics

157

158

Documents

Intelligent Detection and Classification of Lung Nodules