interpretation of plain x-ray images using fuzzy … of plain -ray images... · pengimejan awal...
TRANSCRIPT
INTERPRETATION OF PLAIN X-RAY IMAGES USING FUZZY LOGIC TO DETECT AND CLASSIFY BONE TUMORS
Yeck Yin Ping
Master of Science 2010
Pusat Khidmat Maklumat Akademik UNIVERSITI MALAYSIA SARAWAK
INTERPRETATION OF PLAIN X-RAY IMAGES USING FUZZY LOGIC TO DETECT AND CLASSIFY BONE TUMORS
P. KHIDMAT MAKLUMAT AKADEMIK
111111111 Iii'I'Vi III II1 1000246244
YECK YIN PING
A thesis submitted in fulfillment of the requirements for the degree of Master of Science
Faculty of Computer Science and Information Technology UNIVERSITI MALAYSIA SARAWAK
2010
Acknowledgements
First of all, I gratefully acknowledge Unimas Postgraduate Fellowship for financially
sponsoring my education and research work.
I would like to express my sincere gratitude to my main supervisor, Professor Dr. Wang Yin
Chai. His excellent guidance, advice, continuous supports and insightful comments help me
throughout all the phase of the research work and thesis write-up. This research has been a
very challenging but rewarding experience and I have gained extra knowledge which have
made me more confident and self-independent to participate any particular researches in the
future.
My sincere appreciation is also extended to my co-supervisor, Professor Dr. Pan Kok Long
for providing me with the bone tumor x-ray images that are used in this research. He patiently
guide me through medical knowledge and generously shared his expertise, assisted me to
write several publications and conduct a successful research.
Special thanks are dedicated to all postgraduate committee members from Faculty of
Computer Science and Information Technology for sharing their valuable suggestion, and
great ideas. Furthermore, I would also like to thank my friends in the lab for the fruitful
discussions and invaluable help.
Finally, I am very grateful to my parents for their spiritual support and love. They are always
there for me in any situation to make my life and gained this experience a lot easier.
i
Abstract
C Radiographic is a conventional x-ray image, typically the first imaging test used to diagnose
bone tumor. Probably the most common use of x-ray image is to assist the medical experts in
detecting the early stages of benign tumor growth, identifying tumor suspicious location and
monitoring the progression of degenerative tumor. Reading of x-ray image is usually done by
medical experts visually. The diagnosis process requires human expert's cognitive. It depends
extremely on the knowledge and long term diagnosis experiences of the medical experts. At
the earliest stage of bone tumors, when they are small and difficult to recognize, the
radiological finding can lead to potential misidentification and increase the frequency of
human error. Meanwhile, different medical experts have different perception of bone tumors
because the variable distributions of tumor appeared in x-ray images have presented
ambiguity. To help in overcoming such problems, an x-ray interpretation method based on
fuzzy logic has been developed in this studyýThis method allows the interpretation of plain x-
ray images to be performed semi automatically involving minimum number of input variables
in the detection and classification of bone tumor. In order to ensure that all the abnormalities
present as benign or malignant are classified properly, an image enhancement method has
been developed to refine the input image based on direct manipulation of pixel in the partial
domain. The proposed enhancement method employs image filtering technique with
combination of image registration to increase the contrast of tumor region. The developed
method has been extensively tested and compared against the Mamdani's fuzzy inference
methods in term of accuracy using test samples that were obtained from humeral parts with
various intensities on the x-ray images. The result showed that a 87.36% of accuracy rate was
achieved in bone tumor detection and a 98.38% of sensitivity was achieved in the
11
classification of bone tumor. Demonstrations of the experiment results show the feasibility of
the proposed method in detecting the distributed abnormalities and classifying any
abnormalities present as benign and malignant tumor.
iii
Abstrak
Radiografi merupakan imej x-ray konvensional yang lazimnya digunakan sebagai ujian
pengimejan awal untuk mendiagnos tumor tulang. Biasanya, imej x-ray digunakan untuk
membantu pakar perubatan dalam mengesan peringkat awal pertumbuhan tumor, mengenal
pasti lokasi tumor dan memantau penyebaran tumor yang progresif. Pemeriksaan imej x-ray
lazimnya dilakukan oleh pakar perubatan secara visual. Proses ini memerlukan kepakaran
pakar perubatan. Pada peringkat awal pertumbuhan tumor tulang, pakar perubatan sukar
mengenal pasti tumor tulang yang kecil dan ketaksaan, terutamanya bagi pakar perubatan
yang kurang pengalaman. Ini boleh mengakibatkan peningkatan kesilapan yang dilakukan
oleh pakar perubatan dalam penterjemahan imej x-ray. Bentuk tumor tulang yang pelbagai
juga boleh mengakibatkan pakar-pakar perubatan mempunyai pengertian dan penterjemahan
yang berbeza. Bagi mengatasi masalah tersebut, satu kaedah penafsiran x-ray berdasarkan
fuzzy logik telah dibangunkan dalam kajian ini. Kaedah ini membolehkan tafsiran imej x-ray
dilakukan secara separuh automatik yang melibatkan bilangan pembolehubah yang minimum
dalam pengesanan dan pengelasan tumor tulang. Demi memastikan bahawa kesemua
ketaknormalan sel dapat diklasifikasikan dengan betul, kaedah pembaikan imej telah
dibangunkan untuk membaiki imej berdasarkan manipulasi piksel dalam turnor tulang. Teknik
peningkatan yang dicadangkan dengan menggunakan penapisan imej dan pengabungan imej
untuk meningkatkan kontras rantau tumor. Kaedah-kaedah yang dibangunkan telah diuji dan
dibandingkan dengan kaedah-kaedah Mamdani. Ini adalah untuk menilai kejituan dalam
proses pengesan and pengelasan tumor tulang. Pengkajian ini diuji dengan sampel-sampel
yang didapati daripada pelbagai bahagian humerus yang mempunyai intensiti yang
berlainan. Kadar ketepatan sebanyak 87.36% telah dicapai dalam pengesanan tumor tulang
iv
dan kadar kejituan sebanyak 98.38% telah dicapai dalam pengelasan tumor tulang.
Keputusan eksperimen menunjukkan bahawa kaedah yang dicadangkan mempunyai
kebolehan dalam pengesanan ketaknormalan dan pengelasan tumor tulang.
V
Pusat Khidmat Maklumat Akademik tJNIVERSITI MALAYSIA SARAWAK
Table of Contents
Acknowledgements
Abstract
Abstrak
Table of Contents
List of Figures
List of Tables
List of Abbreviations
Chapter 1 Introduction
1.1 Introduction
1.2 Research Problem
1.3 Research Objectives
1.4 Scope
1.5 Chapters Outline
1
11
iv
V1
X1
xvi
xvii
I
3
4
4
5
Chapter 2 Literature Review
2.1 Introduction 7
2.2 Radiographic Preprocessing and Radiographic Enhancement 7
2.3 Past Investigations on the Detection and Classification of Tumors 11
2.3.1 Neural Networks 11
2.3.2 Fractal Theory 14
2.3.3 Template Matching 16
vi
2.3.4 Wavelet Approaches
2.3.5 Statistical or Texture Analysis Methods
2.3.6 Morphological Operations
2.3.7 Fuzzy Logic
2.4 Comparative Studies
2.5 Summary
Chapter 3 Fuzzy Tumor Detection
3.1 Introduction
3.2 X-ray Images Preprocessing
3.3 Fuzzy Logic Based Knowledge Formalization
3.4 Fuzzy Variable Identification and Rejection
3.4.1 Mean Intensity
3.4.2 Maximum of Gradient Amplitude
3.4.3 Euclidean Distance
3.4.4 Skewness
19
21
23
25
33
35
37
37
39
40
42
43
44
45
3.5 Fuzzification 47
3.5.1 Fuzzy Variables Normalization 47
3.5.2 Quantization 49
3.5.3 Assignment of Fuzzy Membership Functions 49
3.5.3.1 Fuzzy Membership Functions for Mean Intensity 50
3.5.3.2 Fuzzy Membership Functions for Maximum of Gradient 53
3.5.3.3 Fuzzy Membership Functions for Skewness 56
3.6 Fuzzy Inference Rules Setup 59
Vll
3.7 Defuzzification
3.8 Summary
Chapter 4 Fuzzy Tumor Classification
4.1 Introduction
4.2 Partial Tumor Enhancement
4.2.1 Image Filtering
4.2.1.1 Mean Filter
4.2.1.2 Median Filter
4.2.1.3 Gaussian Blur Filter
4.2.2 Image Registration
4.3 Fuzzy Logic Based Knowledge Formalization
4.4 Fuzzy Variable Identification and Rejection
4.4.1 Intensity Difference
4.4.2 Standard Deviation
4.5 Fuzzification
4.5.1 Fuzzy Variables Normalization
4.5.2 Quantization
4.5.3 Assignment of Fuzzy Membership Functions
4.6 Fuzzy Inference Rules Setup
4.7 Defuzzification
4.8 Summary
61
63
66
66
66
67
67
68
68
71
73
74
75
76
77
78
79
87
93
95
viii
Chapter 5 Results and Analysis
5.1 Introduction 98
5.2 Data Acquisitions 98
5.3 X-ray Image Preprocessing Analysis 99
5.3.1 Preprocessing Analysis on High Frequency X-ray Images 99
5.3.2 Preprocessing Analysis on Low Frequency X-ray Images 100
5.3.3 Preprocessing Method Selection 101
5.3.4 Comparison Against Ordinary Images and Enhanced Images 101
5.4 Fuzzy Context Selection 104
5.5 Radiographic Analysis 104
5.5.1 Intensity Characteristics of Normal Radiographic 104
5.5.2 Intensity Characteristics of Abnormal Radiographic 106
5.6 Fuzzy Variables Analysis for Tumor Detection 109
5.6.1 Mean Intensity Analysis on Healthy Samples 109
5.6.2 Mean Intensity Analysis on Abnormal Samples 111
5.6.3 Euclidean Distance Analysis 112
5.6.4 Maximum of Gradient Amplitude Analysis 113
5.6.5 Skewness Analysis 114
5.7 Fuzzy Variables Selection for Tumor Detection 114
5.8 Fuzzy Membership Interval Analysis for Turpor Detection 116
5.9 Fuzzy Defuzzification Method Selection for Tumor Detection 125
5.10 Fuzzy Tumor Detection Benchmark 127
5.11 Partial Filter Selection 128
5.12 Full Filter Analysis 129
ix
5.13 Intensity Difference Analysis for Benign and Malignant Tumors 135
5.14 Standard Deviation Analysis for Benign and Malignant Tumors 137
5.15 Fuzzy Membership Interval Analysis for Tumor Classification 138
5.15 Fuzzy Defuzzification Method Selection for Tumor Classification 147
5.16 Fuzzy Tumor Classification Benchmark 148
5.17 Summary 150
Chapter 6 Conclusions
6.1 Summary
6.2 Research Contributions
6.3 Future Works
152
155
158
References 160
Appendix I Fuzzy Tumor Detection Benchmark for Noise and Noise-free 168
X-ray Images
Appendix II Fuzzy Tumor Classification Benchmark for Noise and Noise- 190
free X-ray Images
Appendix III List of Published Papers 203
X
List of Figures
Figure 2.1 Neural networks structure (Demuth et al., 2008) 11
Figure 3.1 Horizontal projection block diagram 43
Figure 3.2 Block diagram of Euclidean distance measurement on healthy 44
bone image
Figure 3.3 Block diagram of Euclidean distance measurement on bone tumor 45
image
Figure 3.4 Block diagram of symmetric distribution of skew data sets 46
Figure 3.5 Block diagram of asymmetric distribution of skew data sets 46
Figure 3.6 L-function for linguistic value of A, Lo 51
Figure 3.7 Triangular function for linguistic value of A, Med 52
Figure 3.8 F-function for linguistic value of A, Hi 52
Figure 3.9 The membership functions used in classifying the A, value 53
Figure 3.10 L-function for linguistic value of A,, Lo 54
Figure 3.11 Triangular function for linguistic value of A,,, Med 54
Figure 3.12 F-function for linguistic value of AHi 55
Figure 3.13 The membership functions used in classifying the A,, value 56
Figure 3.14 L-function for linguistic value of A.. Lo - 57
Figure 3.15 Triangular function for linguistic value of A, Med 57
Figure 3.16 F-function for linguistic value of A, Hi 58
Figure 3.17 The membership functions used in classifying the A, value 59
X1
Figure 3.18 Cube FAM and sliced cube FAM representations 60
Figure 4.1 The mechanism of image registration 69
Figure 4.2 Subtraction from the image at the left to create the image at right 70
after OR operation is being utilized
Figure 4.3 The enhanced image is created after AND operation is being 70
applied
Figure 4.4 Block diagram of intensity difference between bone lesion and 75
healthy bone
Figure 4.5 Block diagram of selecting and calculating the distributed bone 76
lesion
Figure 4.6 Fuzzy classifier control process 77
Figure 4.7 L-function for linguistic value of AdVLo 80
Figure 4.8 Triangular function for linguistic value of AdLo 81
Figure 4.9 Trapezoidal function for linguistic value of AdMed 81
Figure 4.10 Triangular function for linguistic value of AdHi
Figure 4.11 F-function for linguistic value of AJVHi
Figure 4.12 The membership functions used in classifying the A,, value
Figure 4.13 L-function for linguistic value of AVLo
Figure 4.14 Triangular function for linguistic value of A, Lo
Figure 4.15 Triangular function for linguistic value of A%Med
Figure 4.16 Triangular function for linguistic value of A, Hi
Figure 4.17 F-function for linguistic value of AVVHi
82
83
83
94
85
85
86
87
X11
Figure 4.18 The membership functions used in classifying the Av value 87
Figure 5.1 Comparison between enhanced images and high frequency image 99
Figure 5.2 Comparison between enhanced images and low frequency image 100
Figure 5.3 Comparison between ordinary images and enhanced images 102
Figure 5.4 Intensity characteristics of healthy sample 105
Figure 5.5 Results of threshold value for healthy samples 106
Figure 5.6 Malignant tumor 107
Figure 5.7 Benign tumor 107
Figure 5.8 Intensity characteristics of abnormal sample 108
Figure 5.9 Results of threshold value for normal and abnormal samples 108
Figure 5.10 Mean intensity histogram of healthy sample 1 110
Figure 5.11 Mean intensity histogram of healthy sample 2 110
Figure 5.12 Mean intensity histogram of healthy sample 3 110
Figure 5.13 Mean intensity histogram of abnormal sample 1 111
Figure 5.14 Mean intensity histogram of abnormal sample 2 112
Figure 5.15 Mean intensity histogram of abnormal sample 3 112
Figure 5.16 The comparison of Euclidean distance between healthy bone and 113
bone lesion
Figure 5.17 Maximum of gradient values for normal and abnormal images 113
Figure 5.18 The skewness measurement based on asymmetric and symmetric 114
distribution
Figure 5.19 Construction of fuzzy membership interval for mean intensity 117
Figure 5.20 Optimal frequency for linguistic value of A, Lo 118
X111
Figure 5.21 Optimal frequency for linguistic value of A, Med 118
Figure 5.22 Optimal frequency for linguistic value of A, Hi 119
Figure 5.23 Construction of fuzzy membership interval for maximum gradient 120
Figure 5.24 Optimal frequency for linguistic value of A. Lo
Figure 5.25 Optimal frequency for linguistic value of AmMed
Figure 5.26 Optimal frequency for linguistic value of A�, Hi
121
121
123
Figure 5.27 Construction of fuzzy membership interval for skewness 123
Figure 5.28 Optimal frequency for linguistic value of A,. Lo
Figure 5.29 Optimal frequency for linguistic value of A, Med
Figure 5.30 Optimal frequency for linguistic value of A, Hi
124
124
125
Figure 5.31 Optimal value of mean filter 130
Figure 5.32 Optimal value of median filter 131
Figure 5.33 Optimal value of Gaussian blur filter 132
Figure 5.34 Intensity difference of benign and malignant bone lesion 136
Figure 5.35 Standard deviation of benign and malignant bone lesion 137
Figure 5.36 Construction of fuzzy membership interval for intensity difference 139
Figure 5.37 Optimal frequency for linguistic value of AVLo 140
Figure 5.38 Optimal frequency for linguistic value of AdLo
Figure 5.39 Optimal frequency for linguistic value of AdMed
Figure 5.40 Optimal frequency for linguistic value of Ad Hi
Figure 5.41 Optimal frequency for linguistic value of AdVHi
140
141
142
142
Figure 5.42 Construction of fuzzy membership interval for standard deviation 144
xiv
Figure 5.43 Optimal frequency for linguistic value of A, VLo
Figure 5.44 Optimal frequency for linguistic value of A, Lo
Figure 5.45 Optimal frequency for linguistic value of A, Med
Figure 5.46 Optimal frequency for linguistic value of AVHi
Figure 5.47 Optimal frequency for linguistic value of AVVHi
Figure 5.48 Mean of benign and malignant bone tumor samples
145
145
146
146
147
149
xv
List of Tables
Table 2.1 Comparison among the current studies and clinical applications 34
Table 3.1 Categorization of fuzzy variables into status, ratio range and 50
linguistic value
Table 4.1 Categorization of fuzzy variables into status, ratio range and 79
linguistic value
Table 4.2 The fuzzy rules table for fuzzy tumor classification 88
Table 5.1 Contrast refinement rates for each preprocessing method 101
Table 5.2 Results of difference between ordinary image and enhanced 103
image
Table 5.3 Data overlapped rate for fuzzy variables 115
Table 5.4 Abnormality detection rate for fuzzy defuzzification methods 126
Table 5.5 Benchmarking proposed approach and Pandey approach 127
Table 5.6 Feature enhancement rates for each filter type 128
Table 5.7 Refinement rates from each mask radius of the partial filters 133
Table 5.8 Refinement rates of the partial filters 133
Table 5.9 Benign and malignant refinement rates for each mask radius 134
Table 5.10 Tumor classification rate for fuzzy defuzzification methods 147
Table 5.11 Benchmarking proposed approach and Jain approach 149
Table 5.12 Data overlap rate for fuzzy variable of mean 150
xvi
List of Abbreviations
ANFIS Adaptive Neuro-Fuzzy Inference System
BN Bayesian Networks
CAD Computer Aided Diagnosis
CPFIS Characteristic-point-based Fuzzy Inference System
CT Computed Tomography
DroG Derivative of Gaussian
FAM Fuzzy Associative Memory
fBm fractional Brownian motion
FD Fractal Dimension
FL Fuzzy Logic
FLAIR Fluid-attenuated Inversion-recovery
FPCM Fuzzy Possibilities C-Mean
FT Fractal Theory
GD Gradient Decent
GMM Gaussian Mixture Models
ICA Independent Component Analysis
LSE Least-Squares Estimator
MF Membership Functions
MO Morphological Operations
MRI Magnetic Resonance Imaging
NN Neural Network
PCA Principal Component Analysis
xvii
PSNR Peak Signal-to-Noise Ratio
PTPSA Piecewise Triangular Prism Surface Area
RBF Radial Basis Function
ROI Region of Interest
ROS Region of Suspicious
SGLD Spatial Grey Level Dependency
SOM Self-organizing Map
STA Statistical or Texture Analysis Methods
SVM Support Vector Machine
TM Template Matching
WA Wavelet Approaches
WBC Wisconsin Breast Cancer
WDBC Wisconsin Diagnostics Breast Cancer
xviii
Chapter 1 Introduction
1.1 Introduction
Bone cancer is a case where abnormal cells grow rapidly without any order and destroy the
healthy tissues. Bone metastasis occurs when the cancer cells lose the ability to control
abnormal cells growth and spread to other parts of the body through the bloodstream. The
trend can cause great morbidity including debilitating pain and pathologic fractures (Yao et
al., 2006). It will threaten human's life dangerously. Primary prevention seems impossible
since the causes of this disease still remain unknown as bone cancer is not contagious and
inherited through faulty gene. Hence, early detection at the level of expert recognition of a
medical specialist is more reliable to improve the bone cancer prognosis.
The conventional X-ray image is a useful tool with the highest sensitivity for detecting early
bone cancer and their distributions in bones. It is yielding a significant improvement in bone
cancer survival. The strength of bone x-ray image lies in its unequalled ability to detect bone
cancers in early stages, before tumors have invaded and destroyed nearby healthy tissues and
organs. Through its abilities to detect the density of soft tissue contrast, bone x-ray image is
capable to show the location, size and shape of bone tumor, as well as monitor the progression
of malignant tumor.
At the earliest possible stage of bone tumors, when they are small and difficult to recognize,
the medical experts manually check for the conventional radiographic features of cartilage to
find the distributed abnormalities in bones. The diagnosis requires human expert's cognitive,
which is due to the long term experiences and knowledge of medical expert by visually
I
inspecting the bone x-ray images. The basic studies and clinical applications towards the
development of modern computerized scheme are carried out for the detection and
characterization of lesions radiographic images. This is not only designed to provide an
effective second reader information to the radiologists in the diagnosis of x-ray images, but it
also has the capability to diagnose the presence of lesion on medical images of healthy people
for bone cancer control.
The diagnosis process involves several levels of uncertainty and imprecision and it has been
the major challenge in the field of computer vision. The ultimate challenge is the integration
of knowledge and experience of medical experts with intelligent computer to produce output
of computerized analysis and characterization of medical images for diagnosing purposes.
Current studies toward the investigation for detection and classification of bone lesion is
based on fuzzy logic technique. Fuzzy logic is applied to deal with the problem of knowledge
representation in an uncertain and imprecise interpretation that heavily relies on domain
knowledge and expert analysis which therefore rule out the conventional diagnosis methods.
In this research, it first uses fuzzy logic for tumor detection and then the detected region is
enhanced to extract the desired features for tumor classification. The abnormalities are
decoded to tell whether the tumor is benign or malignant. Expert knowledge is utilized to set
the fuzzy parameters of membership functions based on individual images, and then come up
with rules of data selection and extraction.
2
1.2 Research Problem
Diagnostics on bone tumors researches are relatively rare. The clinical symptoms are usually
unspecific and therefore most tumors are discovered accidentally during routine radiological
exams. Reading of bone x-ray image is generally done by medical experts visually, to detect
and interpret any abnormalities present, as benign or malignant. This is a labour intensive task
as it requires multiple reading of a single x-ray image in order to increase reliability.
The different diagnoses of bone tumors in pathology are difficult and can be based on the
radiological findings, for example define and diffuse the structure of tumor region in the x-ray
images. The abnormalities presented on x-ray images can be extremely small and difficult to
recognize. It can lead to potential misidentification for the medical doctors who have less
experience and do not have specialization on bone tumor diagnosis.
An individual abnormal region may manifest itself differently depending on a benignancy or
malignancy of tumor as tumor appears in different sizes, various densities and irregular
shapes. The variable distributions of tumors appear in bone x-ray images have presented
ambiguity in the field of medical diagnosis. This will cause fuzziness in human perception.
Many bone x-ray images exhibit poor contrast with non-uniform background illumination. In
particular, the abnormal cell boundaries are not sufficiently sharp to be readily extracted. A
major element of this error lies in the failure of the reader to detect signs of abnormality, in
which high intensity pixels in the form of white patches scattering in the image can have a
number of interpretations.
3
1.3 Research Objectives
This research is primarily aimed at the investigation of technique for the interpretation of
plain x-ray images that serves as a diagnostic aid for medical doctors who have less
experience in detecting and classifying the bone tumor. The specific objectives of this
research constitute an important achievement of the long term work outlined as follows:
" To investigate the potential of fuzzy logic and adopt fuzzy logic to develop an x-ray
interpretation method that will be able to detect the bone tumors with various densities
and classify the abnormalities present, as benign and malignant tumors.
" To develop an image enhancement method that will be able to provide the best
separation to distinguish between tumor and its surrounding soft tissue background.
" To identify the appropriate and efficient linguistic variables from domain knowledge
to deal with reasoning into the use of approximate information and uncertainty to
generate decisions in the fuzzy detection and fuzzy classifier.
" To validate the proposed method in term of accuracy through the comparison against
Mamdani's fuzzy inference methods.
1.4 Scope
The overall scope of work reported in this research is focused on techniques for tumor
diagnosis operated semi automatically in uncertainty environment. The detection of tumor is
done simultaneously with normal and abnormal features recognition using fuzzy logic
method. The regions of suspicious with various densities are extracted to differentiate whether
the desired location is normal or abnormal class. Then the abnormal region is selected for
patial enhancement and feature generator as the knowledge of the extent of tumor is important
4