MASSIVE TRAINING ARTIFICIAL IMMUNE RECOGNITION SYSTEM FOR LUNG NODULES DETECTION HANG SEE PHENG A thesis submitted in fulfilment of the requirements for the award of the degree of Doctor of Philosophy (Computer Science) Faculty of Computing Universiti Teknologi Malaysia OCTOBER 2014
48
Embed
MASSIVE TRAINING ARTIFICIAL IMMUNE RECOGNITION …eprints.utm.my/id/eprint/78067/1/HangSeePhengPFC2014.pdf · dan penghitungan fitur tidak dilaksanakan dalam pembelajaran mesin ...
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
MASSIVE TRAINING ARTIFICIAL IMMUNE RECOGNITION
SYSTEM FOR LUNG NODULES DETECTION
HANG SEE PHENG
A thesis submitted in fulfilment of the
requirements for the award of the degree of
Doctor of Philosophy (Computer Science)
Faculty of Computing
Universiti Teknologi Malaysia
OCTOBER 2014
iii
To my beloved parent and husband for their support and encouragement during this
graduate program.
iv
ACKNOWLEDGEMENT
First and foremost, I would like to express my deepest gratitude to my main
supervisor, Professor Dr. Siti Mariyam Shamsuddin for her guidance and dedication
towards the completion of my thesis. I would like to extend my sincere appreciation
to my co-supervisor, Dr. Kenji Suzuki for his assistance in achieving the objectives of
this thesis. Besides, I would like to thank Professor Anca L Ralescu for her advice and
suggestions in my thesis.
In addition, I would like to thank the School of Graduate Studies (SPS) for
the financial support in my studies. This study is also partially supported by The
Ministry of Higher Education (MOHE) under Long Term Research Grant Scheme
(LRGS/TD/2011/UTM/ICT/03, VOT 4L805). Furthermore, I would like to express my
gratitude to Dr Jahari bin Mahbar in the Department of Radiology and Imaging, KPJ
Johor Specialist Hospital for his advice in this study. I am also grateful to the Faculty
of Computing and library of UTM for providing various facilities and references.
Besides, I would also like to thank the developers of the utmthesis LATEX project for
making the thesis writing process a lot easier for me.
Last but at least, I would like to thank my beloved family for their great moral
support and to my friends for their contributions in completing my thesis.
v
ABSTRACT
In the early detection and diagnosis of lung nodule, computer aided detection
(CAD) has become crucial to assist radiologists in interpreting medical images
and decision making. However, some limitations have been found in the existing
CAD algorithms for detecting lung nodules, such as imprecision classification
due to inaccurate segmentation and lengthy computation time. In this research,
Massive Training Artificial Immune Recognition System (MTAIRS) is proposed to
detect lung nodules on Computed Tomography (CT) scans. MTAIRS is developed
based on the pixel machine learning and artificial immune-based system-Artificial
Immune Recognition System (AIRS). Two versions of proposed algorithms have been
investigated in the study: MTAIRS 1 and MTAIRS 2. Since segmentation and feature
calculation are not implemented in the pixel-based machine learning, the loss of
information can be avoided during the data training in MTAIRS 1 and MTAIRS 2.
The experiment and analysis find that MTAIRS 1 and MTAIRS 2 have successfully
reduced the computation time and accomplished good accuracy in the detection of
lung nodules on CT scans compared to other well-known pixel-based classification
algorithms. Furthermore, MTAIRS 1 and MTAIRS 2 are investigated to improve
their performance in eliminating the false positives. A weighted non-linear affinity
function is employed in the training of MTAIRS 1 and MTAIRS 2 to replace Euclidean
distance in affinity measurement. The enhanced algorithms named, E-MTAIRS 1 and
E-MTAIRS 2 are capable to reduce the false positives in the non-nodule classification
while maintaining the accuracy in nodule detection. In order to further provide
comparative analysis of pixel-based classification algorithms in lung nodules detection,
a pixel-based evaluation method of Kullback Leibler (KL) divergence is proposed in
this study. Based on the pixel-based quantitative analysis, MTAIRS 1 performs better
in the elimination of false positives, while MTAIRS 2 in lung nodules detection. The
average detection accuracy for both MTAIRS algorithms is 95%.
vi
ABSTRAK
Dalam pengesanan awal dan diagnosis nodul peparu, pengesanan berbantukan
komputer (CAD) telah menjadi penting untuk membantu ahli radiologi dalam
mentafsirkan imej perubatan dan membuat keputusan. Walaubagaimanapun, terdapat
keterbatasan yang telah ditemui dalam algoritma CAD sedia ada dalam mengesan
nodul peparu seperti ketakpersisan pengelasan yang disebabkan oleh segmentasi
yang tidak tepat dan masa penghitungan yang panjang. Dalam kajian ini, Sistem
Pengecaman Imun Buatan Latihan Besar (MTAIRS) telah dicadangkan untuk
mengesan nodul peparu pada imbasan Tomografi Berkomputer (CT). MTAIRS
dibangunkan berdasarkan kepada pembelajaran mesin piksel dan sistem imun buatan -
Sistem Pengecaman Imun Buatan (AIRS). Dua versi algoritma dicadangankan telah
diselidik dalam kajian ini: MTAIRS 1 dan MTAIRS 2. Oleh kerana segmentasi
dan penghitungan fitur tidak dilaksanakan dalam pembelajaran mesin berasaskan
piksel, kehilangan maklumat boleh dielakkan semasa latihan data pada MTAIRS 1
dan MTAIRS 2. Ujikaji dan analisis mendapati MTAIRS 1 dan MTAIRS 2 telah
berjaya mengurangkan masa penghitungan dan mencapai ketepatan yang baik dalam
mengesan nodul peparu pada imbasan CT berbanding dengan algoritma pengelasan
bersandarkan piksel lain yang terkenal. Tambahan pula, MTAIRS 1 dan MTAIRS
2 telah dikaji untuk meningkatkan prestasi mereka dalam menghapuskan positif
palsu. Fungsi afiniti tak linear berpemberat digunakan dalam melatih MTAIRS 1 dan
MTAIRS 2 bagi menggantikan pengukuran afiniti jarak Euclidean. Algoritma yang
dipertingkatkan dinamakan sebagai E-MTAIRS 1 dan E-MTAIRS 2 dan algoritma
ini berupaya untuk mengurangkan positif palsu dalam pengelasan bukan nodul di
samping mengekalkan ketepatan dalam mengesan nodul tersebut. Bagi mendapatkan
analisis perbandingan lanjutan terhadap algoritma pengelasan berasaskan piksel bagi
pengesanan nodul peparu, kaedah penilaian bersandarkan capahan Kullback Leibler
(KL) telah dicadangkan. Daripada analisis kaedah penilaian berasaskan piksel,
MTAIRS 1 menunjukkan prestasi yang lebih baik dalam menghapuskan positif palsu
manakala MTAIRS 2 adalah lebih baik dalam pengesanan nodul peparu. Purata
ketepatan pengesanan bagi kedua-dua MTAIRS algoritma adalah 95%.
vii
TABLE OF CONTENTS
CHAPTER TITLE PAGE
DECLARATION ii
DEDICATION iii
ACKNOWLEDGEMENT iv
ABSTRACT v
ABSTRAK vi
TABLE OF CONTENTS vii
LIST OF TABLES xi
LIST OF FIGURES xiii
LIST OF ABBREVIATIONS xvi
LIST OF SYMBOLS xviii
LIST OF APPENDICES xx
1 INTRODUCTION 1
1.1 Overview 1
1.2 Background of Problem 2
1.3 Problem Statement 6
1.4 Objectives of Study 9
1.5 Contribution of Study 9
1.6 Research Overview 10
1.7 Scope of Study 12
1.8 Thesis Organisation 13
2 LITERATURE REVIEW 15
2.1 Introduction 15
2.2 Computed-Aided Detection in Medical Imaging 15
2.2.1 Role of Computer-Aided Detection for
CT Scans 17
2.2.2 Lung Nodules Detection on CT Scans by
CAD 18
viii
2.3 Recent Review on Lung CT Nodules Detection by
CAD 19
2.4 Review of CAD Classification Algorithms in
Detection of Lung Nodules 22
2.4.1 Feature-based Classifiers 22
2.4.2 Pixel-based Classifiers 23
2.4.3 Advantages and Disadvantages of
Feature-based and Pixel-based Classifiers 25
2.5 Massive Training Artificial Neural Network
(MTANN) as Pixel-based Classifiers 25
2.6 Artificial Immune Recognition System (AIRS) as
Classifier 27
2.6.1 The Procedures of AIRS 29
2.6.2 Affinity Measures in AIRS 37
2.6.2.1 Weighted Affinity Function 39
2.7 Evaluation Methods for Medical Image Classifica-
tion 41
2.7.1 Quantitative Analysis for Classification 41
2.8 Discussion 42
2.9 Summary 44
3 RESEARCH METHODOLOGY 45
3.1 Introduction 45
3.2 Research Framework of Study 46
3.3 Medical Dataset Preparation 48
3.3.1 Reference Image Data to Evaluate Re-
sponse (RIDER) 48
3.3.2 Lung Image Database Consortium
(LIDC) 49
3.4 Pixel Extraction for Massive Training and Teaching
Images Preparation 49
3.5 Pixel Extraction for Massive Training and Teaching
Images Preparation 51
3.5.1 Stage 1: Input of Pixel-based Training
Data 52
3.5.2 Stage 2: Massive Data Training of
Artificial Immune Recognition System 52
3.5.3 Stage 3: Pixels Classification 58
ix
3.5.4 Stage 4: Reconstruction of Pixels for
Nodule and Non-nodule Images 59
3.6 Enhancement of Affinity Function in the Training
of MTAIRS 59
3.7 Performance Measurement and Evaluation 60
3.8 Summary 62
4 THE PROPOSED MASSIVE TRAINING ARTIFICIAL
IMMUNE RECOGNITION SYSTEM FOR LUNG NOD-
ULE DETECTION 64
4.1 Introduction 64
4.2 Pixel-based Learning and Parameters Determina-
tion 65
4.2.1 Adaptive Parameter in Teaching Images -
Gaussian Function 66
4.2.2 Size of Window Extraction for Training
Sub-regions 70
4.3 Procedure of Massive Training Artificial Immune
Recognition System in Lung Nodules Detection 74
4.3.1 Training Data Preparation 74
4.3.2 The Proposed MTAIRS 1 and MTAIRS 2 75
4.4 Experimental Results and Analysis of the Proposed
MTAIRS 1 and MTAIRS 2 78
4.4.1 Performance of MTAIRS Algorithms in
Lung Nodules Detection 79
4.4.1.1 MTAIRS Training Computation
Time 80
4.4.1.2 Visualisation Results of Testing
Images 80
4.5 The Proposed Scheme for Pixel-based Image
Quantitative Evaluation 85
4.5.1 Image Diversity Measurement based on
KL Divergence 86
4.5.2 Procedures of Pixel-based Image Quanti-
tative Evaluation 87
4.5.3 Implementation of Pixel-based Image
Quantitative Evaluation 90
4.6 Discussion 97
4.7 Summary 100
x
5 ENHANCEMENT OF PROPOSED MTAIRS 1 AND
MTAIRS 2 FOR FALSE POSITIVES REDUCTION 102
5.1 Introduction 102
5.2 Enhanced Massive Training Artificial Immune
Recognition System (MTAIRS 1 and MTAIRS 2) 103
5.2.1 Non-linear Distance in Affinity Measures
for E-MTAIRS 1 and E-MTAIRS 2 104
5.3 Experimental Results of E-MTAIRS 1 and E-
MTAIRS 2 108
5.3.1 Computation Time 108
5.3.2 Qualitative Analysis 109
5.3.2.1 Nodules Visualisation 110
5.3.2.2 Non-Nodules Visualisation 110
5.3.3 Quantitative Analysis 113
5.3.3.1 Pixel-based Evaluation for Nod-
ule Cases 113
5.3.3.2 Pixel-based Evaluation for Non-
nodule Cases 116
5.4 Discussion 126
5.5 Summary 128
6 CONCLUSION AND FUTURE WORK 129
6.1 Research Summary 129
6.2 Research Findings 130
6.3 Research Contributions 133
6.4 Limitations of Research 133
6.5 Future Work 134
REFERENCES 135
Appendices A – C 149 – 154
xi
LIST OF TABLES
TABLE NO. TITLE PAGE
2.1 Description of three basics elements in AIS 28
3.1 Initialised parameters in MTAIRS 55
4.1 Estimated standard deviation of teaching images 68
4.2 Regression analysis for parameter estimation 69
4.3 Details of training vector for different sizes of sub-regions 70
4.4 Comparison of training time for different size of sub-regions 73
4.5 Comparison of training processes in MTAIRS 1 and MTAIRS
2 77
4.6 Comparison of training time for s-MTANN and MTAIRS 80
4.7 Sizes of lung nodules in testing images 81
4.8 Paired sample t-test to analyses the significant difference
between MTAIRS 1 and MTAIRS 2 93
4.9 The bin number with maximum divergence measures and the
total number of pixels 97
4.10 The accuracy of pixel classification by MTAIRS 1 and
MTAIRS 2 97
5.1 Comparison of training time for pixel-based training models 109
5.2 Paired sample t-test to analyse the significant difference
between E-MTAIRS 1 and E-MTAIRS 2 116
5.3 The bin number with maximum divergence measures and
total number of pixels 118
5.4 The accuracy of pixel classification by E-MTAIRS 1 and E-
MTAIRS 2 119
5.5 Number of false positives and percentage of reduction of
MTAIRS 1 and MTAIRS 2 122
5.6 Percentage of false positives and reduction of MTAIRS 1 and
E-MTAIRS 1 124
5.7 Percentage of false positive and reduction of MTAIRS 2 and
E-MTAIRS 2 126
6.1 Answers for formulated research questions 132
xii
B.1 Literature review for detection of lung nodules 150
C.1 Literature review for the application of AIRS in different
application domain 154
xiii
LIST OF FIGURES
FIGURE NO. TITLE PAGE
1.1 Architecture of MTANN 4
1.2 The overview of proposed research flow 11
2.1 Overview of literature review for study domain 16
2.2 Hounsfield scales in CT and X-rays diagnosis 18
2.3 The roles of CAD in the different analysis of lung CT scans 19
2.4 The process of CAD system of lung nodules analysis 20
2.5 The characteristics of machine learning algorithm in CAD
classification 22
2.6 Features-based learning 23
2.7 Pixel-based learning 24
2.8 Framework of MTANN 27
2.9 Basic elements of AIS in application domain 28
2.10 Expansion from Immune Network Theory to Artificial
Immune Recognition System 30
2.11 Four main phases of AIRS algorithm 31
2.12 Memory cell and ARB generation 32
2.13 The mutation process 33
2.14 Competition for resources involving stimulation, resource
allocation and ARB removal 34
2.15 Mutation for ARB survivals 35
2.16 Memory cell introduction 35
2.17 Competition for resources involving stimulation, resource
allocation and ARB removal of the revised version of AIRS 2 37
2.18 Mutation for ARB survivals in AIRS 2 38
2.19 The relation of distance between testing point and training
points 40
3.1 Overview of research framework 47
3.2 ROI extraction from publicly available dataset 50
3.3 Research operational procedures for MTAIRS 51
3.4 Research design of the proposed MTAIRS 53
xiv
3.5 The matching of sub-regions with pixels of teaching images 54
3.6 Memory cell identification and ARB generation 56
3.7 Competition for resources 57
3.8 Memory cell introduction 57
3.9 The matching of testing vector with the unknown class in
testing file 58
3.10 Classification result for testing vector and comparison of
original images and visualisation of computed results 60
3.11 Framework for quantitative evaluation 62
4.1 Contents of Chapter 4 that correspond to the objectives of
research 65
4.2 Work flow for pixel machine learning in classification
algorithms 66
4.3 Scatter diagram to depict the relationship between nodule size
and standard deviation of teaching images 69
4.4 The three nodule cases in the testing of algorithms 71
4.5 The three non-nodule cases in the testing of algorithms 71
4.6 The output images with lung lesions generated by using
MTAIRS 1 and MTAIRS 2 72
4.7 The output images without lung lesions generated by
MTAIRS 1 and MTAIRS 2 73
4.8 Training samples for nodule cases 75
4.9 Training samples for non-nodule cases 75
4.10 The training process in MTAIRS 76
4.11 ARB cell refinements in competition of limited resources in
MTAIRS 1 78
4.12 ARB cell refinements in competition of limited resources in
MTAIRS 2 79
4.13 The visualisation of original testing images and output images
for nodule cases 82
4.14 The illustrations of original testing images and output images
for non-nodule cases 84
4.15 Proposed scheme of KL divergence for quantitative analysis 86
4.16 Procedures of proposed pixel-based evaluation 88
4.17 The comparison of KL divergence measures for MTAIRS 1
and MTAIRS 2. 92
4.18 The log-difference measurement among the bins of histogram
for MTAIRS 1 and MTAIRS 2 95
xv
5.1 Contents of Chapter 5 that correspond to Objective 4 of the
research 103
5.2 Research design for the enhancement of MTAIRS 1 and
MTAIRS 2 105
5.3 The implementation of non-linear affinity function in training
mechanism of E-MTAIRS 1 and E-MTAIRS 2 107
5.4 The distance relation among both, testing and training points
in non-linear distance function 108
5.5 Output images of nodule cases produced by MTAIRS 1,
MTAIRS 2, E-MTAIRS 1 and E-MTAIRS 2 compared to
original images 111
5.6 Output images of non-nodule cases produced by MTAIRS
1, MTAIRS 2, E-MTAIRS 1 and E-MTAIRS 2 compared to
original testing images 112
5.7 The comparison of KL divergence measures for E-MTAIRS
1 and E-MTAIRS 2. 114
5.8 The log-difference measurement among the bins of histogram
for E-MTAIRS 1 and E-MTAIRS 2 117
5.9 Procedures of pixel-based evaluation of non-nodule cases 120
5.10 Pixel difference in output images for non-nodule cases by
MTAIRS 1 and MTAIRS 2 121
5.11 Comparison of pixel difference in output images for non-
nodule cases by MTAIRS 1 and E-MTAIRS 1 123
5.12 Comparison of pixel difference of output images for non-
nodule cases by MTAIRS 2 and E-MTAIRS 2 125
6.1 Overview of conducted research 130
xvi
LIST OF ABBREVIATIONS
AI – Artificial Intelligence
AIRS – Artificial Immune Recognition System
AIS – Artificial Immune System
ANN – Artificial Neural Network
ARB – Artificial Recognition Ball
CAD – Computer Aided Detection
CT – Computed Tomography
DICOM – Digital Imaging and Communications in Medicine
DVDM – Discretised Value Difference Metric
E-MTAIRS – Enhanced Artificial Immune Recognition System
FP – False Positive
GA – Genetic Algorithm
GLCM – Grey Level Co-occurrence Matrix
GPU – Graphics Processor Unit
HEOM – Heterogeneous Euclidean-overlap Metric
HU – Hounsfield values
HVDM – Heterogeneous Value Difference Metric
KL – Kullback Leibler
kNN – k-nearest-neighbor
LDA – Linear Discriminant Analysis
LIDC – Lung Image Database Consortium
MATLAB – Matrix Laboratory
MIPAV – Medical Image Processing, Analysis, and Visualisation
MRF – Markov Random Field
MRI – Magnetic Resonance Imaging
MTAIRS – Massive Training Artificial Immune Recognition System
MTANN – Massive Training Artificial Neural Network
NA – Not Available
xvii
NCI – National Cancer Institution
NIBIB – National Institute of Bioimaging and Bioengineering
NMR – Nuclear Magnetic Resonance
PET – Positron Emission Tomography
RIDER – Reference Image Data to Evaluate Response
ROC – Receiver-Operating Characteristic
ROI – Region of Interest
RSNA – Radiological Society of North America
s-MTANN – Standard Massive Training Artificial Neural Network
SOM – Self Organizing Maps
SVM – Support Vector Machine
USG – Ultrasonographic
VDM – Value Difference Metric
WHO – World Health Organisation
XML – Extensible Markup Language
xviii
LIST OF SYMBOLS
ag – Training antigen in AIRS
n – Number of training vector
mcmatch – The best match memory cell
Mc – Memory cell
MCag.c – Memory pool with class c
d – Dimension of vector
Xt – Testing vector in AIRS algorithm
Xi – Training vector in AIRS algorithm
pi – Probability distribution for image P
qi – Probability distribution for image Q
T – Two-dimension teaching image
σT – Standard deviation for teaching image
ν – Training vector in MTAIRS
w – Weight in distance function
δtij – Distance between input and training points along j th feature
Dist – Distance in non-linear affinity function
D – Divergence function
YT – Predicted standard deviation in regression model
X – Size of nodules in regression model
Y – Actual standard deviation in regression model
σest – Standard error in estimation
R – Correlation coefficient
ω – Size of window for training sub-region
O – Output image
HO – Original testing image
HA – Computed image by MTAIRS
h0i – Probability distribution of image HO
hAi – Probability distribution of image HA
xix
Jmax – Argument of the maximum column in KL divergence
Imax – Argument of the maximum value in KL divergence
φ – Null set
z – Distance between concerned training point and each training
instances
xx
LIST OF APPENDICES
APPENDIX TITLE PAGE
A List of Publication 149
B Review of Methods and Performance for Lung Nodules
Detection 150
C Application of AIRS 154
CHAPTER 1
INTRODUCTION
1.1 Overview
The recent statistics reported by World Health Organisation (WHO) shows
that lung cancer continues to contribute the highest number of death among other
types of cancer in the world (WHO, 2014). Lung cancer is also known as the most
common cancer for both men and women in many Asia countries as well as the
developing countries (Liam et al., 2006). In Malaysia, the National Cancer Registry
of Malaysia has shown that lung cancer is the third most common cancer in the
country (Sachithanandan and Badmanaban, 2012). Due to the dramatic statistics of
lung cancer, its diagnoses have always gained global attention from medical experts in
an attempt to improve the patients’ survival rate. The survival rate of lung cancer
patients can possibly be reduced if the cancer is detected in the early pre-clinical
stage (Jemal et al., 2005). In modern medical diagnosis, screening is an effective
manner to discover lung lesion in the early stage, and hence reduces the lung cancer
mortality (Alex, 2005, WHO, 2014). Thus, to view the condition of human body
noninvasively through screening, medical imaging plays an important role to detect
the early symptoms of lung cancer.
Recently, the development of medical imaging by various technologies is
rapidly emerging in the field of medical diagnosis. There are different types of medical
modalities that have been actively used, such as, planar X-ray, computed tomography
(CT) scans, magnetic resonance imaging (MRI), ultrasonographic diagnostics (USG),
positron emission tomography (PET) and nuclear magnetic resonance (NMR) (Alex,
2005). Compared to conventional planar X-ray images and other modalities, CT
images provide accurate visualisation of lung conditions especially in the lung nodules
detection. Lung nodule is a pathological spot, measuring three centimeters or less
in diameter on the lung. The growth of a nodule larger than three centimeters is
2
considered as lung mass which is more likely to represent lung cancer. Therefore, early
detection of lung nodules is crucial as they are likely to be curable lung cancer (Delogu
et al., 2005). With a CT scan, a series of cross-sections is obtained and analysed by a
radiologist to detect lung nodules. The interpretation of CT images requires significant
time from radiologists, as hundreds of multi-sections per patient is acquired from a
CT scan. Consequently, an automated system is crucial in assisting the detection and
classification of disease patterns in the lung medical images.
Computer aided detection (CAD) is an automated system used to enhance the
performance of radiologists and thus, reduces their workload. As reported by Girvin
and Ko (2008), performance of radiologists in the detection of lung nodules has been
significantly increased by mean of CAD schemes when interpreting a large-scale
dataset. In recent decade, research efforts have been focusing on the development of
CAD system that could accurately recognise lesions in CT imagery, and consequently
reduces false positive detection in the existing system. In the detection of lung nodules,
computational intelligence techniques is constantly applied as these techniques provide
promising results (Shiraishi et al., 2011, Suzuki et al., 2005b). It was reported that
a variety of intelligence techniques were also successfully used in different types
of medical images processing (BagcI et al., 2012, Jiang et al., 2010, Shi and He,
2010, Stoitsis et al., 2006). In this chapter, the problem background of computational
intelligent algorithms in CAD in detecting lung nodules will be discussed in the
following section. It is followed by the discussion on the problem statements, the
objectives of the study, the research contributions, the research methodology and the
scope of the study. Lastly, the organisation of chapters is summarised in this thesis.
1.2 Background of Problem
Since the majority of imaging modalities provide a large number of datasets,
the interpretation of medical imaging by automatically classifying medical images is
crucial. As an automatic system, CAD has become one of the main research concerns
in medical imaging and diagnostic radiology. In the current lung nodules assessment
by CAD, the computerised systems are able to detect lung nodules and categorise
the types of nodules, for example malignant and benign. Thus, the improvement of
efficiency and accuracy in lung nodules identification has always gained interest of
radiologists and researchers alike. In order to classify these abnormalities, pattern
recognition techniques are usually adopted in the computerised detection system
(BagcI et al., 2012, Kilic et al., 2009, Shiraishi et al., 2011, Stoitsis et al., 2006).
3
There are several pattern recognition techniques available in CAD scheme,
such as, methods based on decision theory, structural methods and computational
intelligence methods (Ogiela and Tadeusiewicz, 2008). As a subset of pattern
recognition approaches, computational intelligence algorithms are popular since
they provided promising results and more efficient diagnosis (Stoitsis et al., 2006,
Verma and Zakos, 2001). There are a number of typical computational intelligence
approaches used to analyse medical imaging such as artificial neural network