produced, small tumours misdetection and the need for user manual interaction (Andreea et al., 2011). 1.4 Problems ...... Wong, M. T., He, X. & Yeh, W. C. (2011).
! " # $ % " & !$ " ! % "! & " '& ( ) * +( $ ( $ , , - * , , " . / 0 $ "
* &0 1 2 3445 2 2 &6 1 3447 8 ' ! # 1 3495 8 ! # ' ! " 8*
! " # $
%"% & "'
!"#$ % "
! & ' & ( )*! &
" " *)+,-!
./ 0 & ! 10&!23,)+ 45 67 0 8 99 !% 4 3,)* % "
! & ' & ( " " 3,)*
TUMOURS DETECTION IN BREAST MRI IMAGES USING IMPROVED METHODS Improved Computer Image Processing Methods for Tumours Segmentation and Detection in Breast Magnetic Resonance Imaging
by
Author: ALI QUSAY AL-FARIS
Co-authors: Umi Kalthum Ngah Nor Ashidi Mat Isa
$&.12:/('*0(17
Foremost, I would like to express my deepest and sincerest gratitude to God, the most Merciful for letting me through all the difficulties, and for providing me the blessings to complete this work. It is with immense gratitude that I acknowledge the help of my PhD supervisor Associate Professor Dr. Umi Kalthum bt Ngah for the continuous support throughout my study and research, and for her patience, motivation, enthusiasm, knowledge, and kindness. It was a great honor to work under her supervision.I would like to express my deepest appreciation and thanks to my PhD co-supervisor Professor Dr. Nor Ashidi Mat Isa for giving invaluable help, advising support, suggestions and comments. I want to express my gratitude also to my other PhD co-supervisor Professor Dr. Ibrahim Lutfi Shuaib (Advanced Medicaland Dental Institute (AMDI)) for the inspiration and the great assistance in the medical and radiological aspects of the research. This work would not have been possible without the generous support of all my supervisors. I would like to express my deepest appreciation to School of Electrical and Electronics Engineering, Universiti Sains Malaysia (USM) for providing methe necessary facilities and equipments.Last, but not least, this work is specially dedicated to people in my heart; my parents; Qusay Al-Faris & Shurook Mubarak, my beloved wife and soul mate; Shams AlFarees and my sons; Layth and Sam for their love, unconditional support, continues prayers and for all of the sacrifices that they have made throughout my life. I cannot find words to express my gratitude, respect and appreciation for them.
ii
7$%/(2)&217(176 $&.12:/('*(0(17«««««««««««««««««««
LL
7$%/(2)&217(176«««««««««««««««««««
LLL
/,672)7$%/(6«««««««««««««««««««««
YLL
/,672)),*85(6«««««««««««««««««««
L[
/,672)$%%5(9,$7,216«««««««««««««««««
[LLL
$%675$&7«««««««««««««««««
[Y
&+$37(5 ,1752'8&7,21 1.1 Background ««««««««««««««««
1.2 Breast Tumour Imaging Techniques ««««««««««««
1.3 Breast MRI Tumour Segmentation ««««««««««««
1.4 Problems and Motivations ««««««««««««««««
1.5 Research Objectives «««««««««««««««««
1.6 Scope of the Study ««««««««««««««««
1.7 Book Outline ««««««««««««««««
CHAPTER 2 : LITERATURE REVIEW 2.1 Introduction «««««««««««««««««
2.2 Breast Screening Modalities «««««««««««««««
2.2.1
Mammography ««««««««««««««««
14
2.2.2
Ultrasonography ««««««««««««««««
15
2.2.3
MRI Screening ««««««««««««««««
16
2.3 CAD for Breast MRI «««««««««««««««««
iii
2.4 Breast MRI Tumour Segmentation Approaches ««««««««
2.4.1
Supervised Approaches ««««««««««««««
18
2.4.2
Unsupervised Approaches «««««««««««««
19
2.4.3
Semi-Supervised Approaches ««««««««««««
21
2.5 Breast Skin-Line Exclusion Approaches «««««««««««
2.6 Image Processing Techniques «««««««««««««««
Image Thresholding Methods ««««««««««««
29
Automatic Thresholding ««««««««««
30
6HHGHG5HJLRQ*URZLQJ 65* «««««««««««
2.6.2.1
SRG in Medical Images ««««««««««
35
2.6.2.2
Methods of Automatic SRG ««««««««« 36
2.6.1
2.6.1.1
2.6.3
2.6.4
Image Clustering Methods «««««««««««««
39
2.6.3.1
Hierarchical Clustering ««««««««««
39
2.6.3.2
Partitional Clustering «««««««««««
40
Fundamental Morphological Operations ««««««««
45
2.6.4.1
Morphological Thinning Operation ««««««
45
2.6.4.2
0RUSKRORJLFDO'LODWLRQDQG(URVLRQ2SHUDWLRQV«
46
2.6.4.3
0RUSKRORJLFDO2SHQLQJ2SHUDWLRQ««««««
48
2.6.4.4
Connected &RPSRQHQW/DEHOOLQJ«««««««
49
2.7 Summary ««««««««««««««««
CHAPTER 3 : METHODOLOGY 3.1 Introduction «««««««««««««««««
3.2 Data Acquisition Phase ««««««««««««««««
iv
The RIDER Dataset ««««««««««««««««
54
3.3 Pre-Processing Phase «««««««««««««««««
3.4 Breast Skin-Line Exclusion Phase using Proposed Integration 0HWKRGRI/6$&DQG0RUSKRORJLFDO7KLQQLQJ$OJRULWKPV«««
3.4.1
Breast Skin-Line ERUGHU6HJPHQWDWLRQ6WDJH««««««
59
3.4.2
Breast Skin-Line Removal Stage ««««««««««
61
3.5 Image Thresholding Phase using Proposed Mean Maximum Raw Thresholding Method «««««««««««
3.2.1
3.5.1
Mean Maximum Raw Thresholding Algorithm (MMRT «
67
3.6 Breast MRI Tumour Segmentation Phase Using Two Proposed 0HWKRGV««««««««««««««««««««««««« 3.6.1
Tumour Segmentation Preprocessing «««««««««
73
3.6.2
Using Seeded Region Growing «««««««««««
74
3.6.3
First Proposed Method: Modified Automatic Seeded Region Growing (BMRI-MASRG) «««««««««««««
76
3.6.3.1
Automatic SRG Seed Selection of BMRI-MASRG«««««««««««««
76
Automatic SRG Threshold Value Selection of BMRI-MASRG ««««««««««««
81
Second Proposed Method: Integrated method of SRG and PSO Image Clustering (BMRI-SRGPSOC «
83
3.6.4.1
3DUWLFOH6ZDUP2SWLPL]DWLRQ,PDJH&OXVWHULQJ«
83
3.6.4.2
Automatic SRG Seed Selection of BMRI-SRGPSOC .««««««««««««
85
Automatic SRG Threshold Value Selection of BMRI-SRGPSOC «««««««««««
88
3.7 Evaluation Criteria «««««««««««««««««
Skin-line Exclusion Phase Evaluation ««««««««
90
3.6.3.2
3.6.4
3.6.4.3
3.7.1
v
3.7.2
Image Thresholding Phase Evaluation ««««««««
92
3.7.3
Tumour Segmentation Phase Evaluation «««««««
93
3.8 Summary «««««««««««««««««««««««
CHAPTER 4 : RESULTS AND DISCUSSION 4.1 Introduction ««««««««««««««««
4.2 Results of Breast Skin-Line Exclusion «««««««««««
4.2.1
Results of Breast Skin-/LQH%RUGHU6HJPHQWDWLRQ6WDJH«
101
4.2.2
Results of Breast Skin-Line Removal Stage ««««««
103
4.3 Results of Image Thresholding Using MMRT ««««««««
4.4 Results of Tumour Segmentation Phase «««««««««««
4.4.1
Results of Modified Automatic Seeded Region Growing (BMRI-MASRG) «««««««««««««
117
Results of Integrated Method of SRG and PSO Image Clustering «««««««««««««
121
Comparison of Proposed Segmentation Approaches (BMRIMASRG and BMRI-65*362& DQG2WKHU$SSURDFKHV«
125
4.5 Summary «««««««««««««««««««««««
4.4.2
4.4.3
CHAPTER 5 : CONCLUSION AND FUTURE WORKS 5.1 Conclusions and Research Contributions ««««««««««
5.2 Suggestions and Future Works ««««««««««««««
REFERENCES ««««««««««««««««««««««
vi
/,672)7$%/(6
Table 2.1
Comparison of breast MRI tumour segmentation approaches .....
Table 2.2
Comparison of breast skin-line exclusion methods ««««
Table 2.3
Comparison of automatic Thresholding methods ««««««
Table 2.4
Comparison of SRG methods «««««««««««««
Table 2.5
Comparison of image clustering methods «««««««««
Table 3.1
Breast skin-line thickness by different studies in mm and pixel units «««««««««««««««««««««««.
Table 4.1
Summary results of skin-line segmentation for RIDER MRI breast images using evaluation measures (TPF, FNF, FPF, TNF and STVF) «««««««««««««««««««««
Table 4.2
Summary results of skin-line segmentation for RIDER MRI breast images using evaluation measures (Jaccard MCR and Dice) ««««««««««««««««««««««.
Table 4.3
Summary results of skin-line removal for RIDER MRI breast images using evaluation measures (TPF, FNF, FPF, TNF and STVF) ««««««««««««««««««««««
Table 4.4
Summary results of skin-line removal for RIDER MRI breast images using evaluation measures (Jaccard MCR and Dice) «
Table 4.5
Summary results of the pixel based evaluation approach (Jaccard and Dice measures) for MMRT «««««««««
Table 4.6
Summary results of the quality evaluation approach (PSNR and MSE measures) for MMRT ««««««««««««««
Table 4.7
Results of evaluating the Jaccard and Dice measures for thresholding using the proposed method and other methods (Iterative Thresholding, Grey level Histogram, Entropy based, Fuzzy Thresholding and Multilevel Thresholding) «««««.
Table 4.8
Results of evaluating the PSNR and MSE measures using the thresholding of the proposed method and other methods (Iterative Thresholding, Grey level Histogram, Entropy based, Fuzzy Thresholding and Multilevel Thresholding) «««««.
vii
Table 4.9
Summary of the ANOVA tests analysis for the proposed DSSURDFK¶V UHVXOWV FRPSDUHG ZLWK WKH UHVXOWV RI WKH RWKHU approaches (Iterative Thresholding, Grey level Histogram, Entropy based, Fuzzy Thresholding and Multilevel Thresholding) «««««««««««««««««««
Table 4.10 Evaluation results of BMRI-MASRG using TPF (Sensitivity), FNF, FPF, TNF (Specificity) and STVF ««««««««« Table 4.11 Evaluation results of BMRI-MASRG using Relative Overlap (Jaccard) and MCR «««««««««««««««««
Table 4.12 Evaluation results of BMRI-MASRG of automatically selected LQLWLDOVHHGSL[HO¶Vcoordinates compared with the manually VHOHFWHGSL[HO¶VFRRUGLQDWHV «««««««««««««« Table 4.13 Evaluation results of BMRI-SRGPSOC using TPF (Sensitivity), FNF, FPF, TNF (Specificity) and STVF «««««««««. Table 4.14 Evaluation results of BMRI-SRGPSOC using Relative Overlap (Jaccard) and MCR «««««««««««««««««..
Table 4.15 Evaluation results of BMRI-SRGPSOC of automatically VHOHFWHGLQLWLDOVHHGSL[HO¶Vcoordinates compared with the PDQXDOO\VHOHFWHGSL[HO¶VFRRUGLQDWHV «««««««««« Table 4.16 Segmentation results for the proposed approaches (BMRIMASRG and BMRI-SRGPSOC) and other approaches (KNN, SVM, Bayesian, FCM and IMPST) ««««««««««« Table 4.17 Summary of the ANOVA tests analysis for BMRI-MASRG results compared with the results of the other approaches (KNN, SVM, Bayesian, FCM and IMPST) ««««««««««« Table 4.18 Summary of the ANOVA tests analysis for BMRI-SRGPSOC results compared with the results of the other approaches (KNN, SVM, Bayesian, FCM and IMPST) ««««««««««« Table 4.19 Area under the Curve for the proposed approaches (BMRIMASRG and BMRI-SRGPSOC) compared to previous methods (KNN, SVM, Bayesian, FCM and IMPST) ««««««««
viii
/,672)),*85(6
Figure 1.1
Estimated number of cancer diagnosed cases in the world based on IARC study «««««««««««««««««
Figure 1.2
Estimated number of cancer deaths in the world based on IARC study ««««««««««««««««««««
Figure 1.3
Estimated number of cancer diagnosed cases in Malaysia based on IARC study «««««««««««««««««
Figure 1.4
Estimated number of cancer deaths in Malaysia based on IARC study ««««««««««««««««««««
Figure 1.5
An example of a breast MRI image shows the similarity in the grey level intensity of skin-line and the tumour ««««««
An example of a breast MRI image shows tumour, skin-line and other tissues regions ««««««««««««««««««
Figure 2.1
Examples of mammogram images «««««««««««
Figure 2.2
Examples of breast ultrasound images ««««««««««
Figure 2.3
Examples of breast MRI «««««««««««««««
Figure 2.4
Morphological thinning operation ««««««««««««
Figure 2.5
Morphological erosion operation ««««««««««««
Figure 2.6
Morphological dilation operation ««««««««««««
Figure 2.7
Morphological opening operation ««««««««««««
Figure 2.8
Connected Component Labelling ««««««««««««
Figure 3.1
Flowchart of the proposed segmentation approach for breast MRI tumour ««««««««««««««««««««
Figure 3.2
Malignant breast image from RIDER dataset «««««««««
Figure 3.3
Benign breast image from RIDER dataset ««««««««««
Figure 3.4
Results of image splitting «««««««««««««««
Figure 3.5
Applying Median filter «««««««««««««««««««
Figure 1.6
ix
Figure 3.6
Different results after the application of LSAC algorithm ZLWKGLIIHUHQWYDOXHVRIıDQGܰௌ ««««««««««««««
Figure 3.7
The pixel p and its eight neighbours pixels ««««««««
Figure 3.8
Results after applying Morphological Thinning algorithm with three different iteration numbers on the resultant image of LSAC algorithm ««««««««««««««««««
Figure 3.9
3VHXGRFRGHRI0057«««««««««««««««««
Figure 3.10 The process of automatic selection of the threshold value using MMRT ««««««««««««««««««««
Figure 3.11 The initial seed pixel and their eight neighbor pixels ««««
Figure 3.12 Pseudo FRGHRI65*DOJRULWKP«««««««««««««
Figure 3.13 Block diagram which illustrates the processes of Automatic SRG Seed Selection of BMRI-MASRG «««««««««
Figure 3.14 5HJLRQUDQNLQJDFFRUGLQJWRWKHLUSL[HOV¶GHQVLW\YDOXHV «««
Figure 3.15 Block diagram which illustrates the processes of Automatic SRG Threshold Value Selection of BMRI-MASRG ««««...
Figure 3.16 Block diagram which illustrates the processes of Automatic SRG Seed Selection of BMRI-SRGPSOC «««««««« Figure 3.17 Pseudo code of BMRI-SRGPSOC threshold value selection ««
Figure 3.18 Figure 3.18 Diagram showing the definitions of TPF, TNF, FPF DQG)1)LQWKHHYDOXDWLRQRIVHJPHQWDWLRQUHVXOWV««««««
Breast skin-line exclusion processes on three malignant RIDER image «««««««««««««««««««
Figure 4.2
Breast skin exclusion processes on two benign RIDER images..
Figure 4.3
The ROC curve for MRI breast skin-line segmentation «««
Figure 4.4
The ROC curve for MRI breast skin-line removal «««««
Figure 4.5
Results of applying MMRT and the standard thresholding methods on malignant test image ««««««««««««.
Figure 4.1
x
Figure 4.6
Figure 4.7
Figure 4.8
Figure 4.9
Results of applying MMRT and the standard thresholding methods on benign test image «««««««««««««
Statistical ANOVA graphs for MMRT in comparison with the standard methods using results of Jaccard measure «««««
Statistical ANOVA graphs for MMRT in comparison with the standard methods using results of Dice measure ««««««
Statistical ANOVA graphs for MMRT in comparison with the standard methods using results of PSNR measure «««««
Figure 4.10 Statistical ANOVA graphs for MMRT in comparison with the standard methods using results of MSE measure «««««« Figure 4.11 The proposed (BMRI-MASRG) approach processes on one of miliganant RIDER image «««««««««««««« Figure 4.12 The proposed (BMRI-MASRG) approach processes on one of benign RIDER image «««««««««««««««
Figure 4.13 The proposed (BMRI-SRGPSOC) approach processes on one of benign RIDER image «««««««««««««««
Figure 4.14 The proposed (BMRI-SRGPSOC) approach processes on one of maliganant RIDER image ««««««««««««««
Figure 4.15 Comparison of segmented tumour using proposed approaches (BMRI-MASRG and BMRI-SRGPSOC) by testing five RIDER images with their GT ««««««««««««««««« Figure 4.16 Results of applying BMRI-MASRG and BMRI-SRGPSOC in comparison to previous approaches malignant WHVWLPDJH««
Figure 4.17 Results of applying BMRI-MASRG and BMRI-SRGPSOC in comparison to previous approaches benign WHVWLPDJH««««
Figure 4.18 Statistical ANOVA graphs for BMRI-MASRG and BMRISRGPSOC in comparison with the previous methods using results of TPF measure «««««««««««««««« Figure 4.19 Statistical ANOVA graphs for BMRI-MASRG and BMRISRGPSOC in comparison with the previous methods using results of TNF measure «««««««««««««««« Figure 4.20 Statistical ANOVA graphs for BMRI-MASRG and BMRI- SRGPSOC in comparison with the previous methods using
xi
results of STVF measure ««««««««««««««« Figure 4.21 Statistical ANOVA graphs for BMRI-MASRG and BMRISRGPSOC in comparison with the previous methods using results of Jaccard measure ««««««««««««««« Figure 4.22 Statistical ANOVA graphs for BMRI-MASRG and BMRISRGPSOC in comparison with the previous methods using results of MCR measure ««««««««««««««« Figure 4.23 The ROC curves for the proposed method and the previous methods ««««««««««««««««««««««««««
xii
/,672)$%%5(9,$7,216
ANN
Artificial Neural Networks
ANOVA
Analysis of Variance
AUC
Area Under the Curve
BMRI-MASRG
Breast Magnetic Resonance Imaging Tumour using Modified Automatic Seeded Region Growing
BMRI-SRGPSOC
Breast Magnetic Resonance Imaging Tumour using Hybrid Automatic Method of Seeded Region Growing and Particle Swarm Optimization Image Clustering
CAD
Computer Aided Detection
CCL
Connected Component Labelling
EM
Expectation Maximization
FCM
Fuzzy C-Means
FNF
False Negative Fraction
FPF
False Positive Fraction
GT
Ground Truth
IARC
International Agency for Research on Cancer
ICM
Iterative Conditional Mode
IMPST
Improved Self-Training
KNN
K-Nearest Neighbours
LSAC
Level Set Active Contour
Maxp
Maximum Possible Pixel
MCET
Minimum Cross Entropy Thresholding
MCR
Misclassification Rate
xiii
MMRT
Mean Maximum Raw Thresholding
MRI
Magnetic Resonance Imaging
MSE
Mean Square Error
PSNR
Peak Signal to Noise Ratio
PSO
Particle Swarm Optimization
RIDER
Reference Image Database to Evaluate Therapy Response
ROC
Receiver Operating Characteristic
ROI
Region Of Interest
SR
Main Suspected Region
SRG
Seeded Region Growing
SRGFE
Seeded Region Growing Feature Extraction
STVF
Sum of True Volume Fraction
SVM
Support Vector Machine
SYNERACT
Synergistic Automatic Clustering Technique
TNF
True Negative Fraction
TPF
True Positive Fraction
WMMR
Window Mean Maximum Raw
xiv
$%675$&7
Breast cancer is the leading cause of death amongst cancer patients afflicting women and the second most common cancer around the world. Magnetic Resonance Imaging (MRI) is one of the most effective radiology tools to screen breast cancer. However, image processing techniques are needed to help radiologists in interpreting the images and segmenting tumours regions to reduce the number of false-positive. In this study, a segmentation approach with automatic features is developed for breast MRI tumours. The methodology starts with data acquisition followed by pre-processing. This is then followed with breast skin-line exclusion using integrated method of Level Set Active Contour and Morphological Thinning. Next, regions of interests are detected using proposed Mean Maximum Raw Thresholding method (MMRT). In the tumour segmentation phase, two modified Seeded Region Growing (SRG) methods are proposed; i.e. Breast MRI Tumour using Modified Automatic SRG (BMRI-MASRG) and Breast MRI Tumour using SRG based on Particle Swarm Optimization Image Clustering (BMRI-SRGPSOC). The RIDER breast MRI dataset was used for evaluation and the results are compared with the ground truth of the dataset. From analysing the evaluation results, it can be noticed that the proposed approaches scored high results using various measures comparing to previous methods. The results of skin-line exclusion scored high average performance in both stages; border segmentation stage (sensitivity = 0.81 and specificity = 0.94) and removal stage (sensitivity = 0.86 and specificity = 0.97). The quality evaluation of MMRT showed improved results with average of PSNR = 69.97 and MSE = 0.01. In the tumour segmentation phase, the sensitivity results of the two proposed methods; BMRI-MASRG and BMRI-SRGPSOC showed more accurate segmentation with averages of 0.82 and 0.84 respectively. xv
Similarly, the specificity results also scored better performance compared to previous methods. The averages of BMRI-MASRG and BMRI-SRGPSOC are 0.90 and 0.91 respectively.
xvi
CHAPTER 1 INTRODUCTION
1.1 Background Breast cancer is the second most common cancer in the world and is the leading cancer amongst women. According to a study conducted by the International Agency for Research on Cancer (IARC) (an intergovernmental agency forming part of the World Health Organization of the United Nations), an estimation of 1.677 million new breast cancer cases have been diagnosed in 2012 (794,000 in developed countries and 883,000 cases in the third world countries), making 25.2 % of total new cancer cases in the world. Figure 1.1 shows the ten most commonly diagnosed cancers in the world, the figure estimates total number and percentage of new cases diagnosed per year. Similarly, the death rates among breast cancer patients are the most amongst cancer cases, as shown in Figure 1.2 (Ferlay et al., 2013).
Figure 1.1 Estimated number of cancer diagnosed cases in the world based on IARC study (Ferlay et al., 2013). 1
Figure 1.2 Estimated number of cancer deaths in the world based on IARC study (Ferlay et al., 2013). In Malaysia, breast cancer is the leading diagnosed cancer among women where the estimated number of this disease is around 38.74 per 100,000 populations. Close to 5,410 new cases are reported annually, making 28.0 % of total new cancer cases for women in Malaysia. Figure 1.3 shows the estimated number of cancer diagnosed cases in Malaysia based on IARC study (Ferlay et al., 2013). Breast cancer is also the first common cause of death between women cancer patients with 2,572 death cases per year, making 24.7 % of total cancer death cases in Malaysia. Figure 1.4 shows estimated number of cancer deaths in Malaysia based on IARC study (Alias et al., 2008).
2
Figure 1.3 Estimated number of cancer diagnosed cases in Malaysia based on IARC study (Ferlay et al., 2013).
Figure 1.4 Estimated number of cancer deaths in Malaysia based on IARC study (Ferlay et al., 2013).
3
Breast screening techniques are an essential way for cancer detection. Accurate segmentation for suspected tumours using computer algorithms increases significantly the possibilities of breast cancer patients to survive.
1.2 Breast Tumour Imaging Techniques Tumours are painful or painless lumps that could appear as dense masses with regular or irregular shapes in screening techniques images (Hussain et al., 2011). Besides the physical examination, several commonly used modalities for breast screening are mammography,
ultrasonography
and
Magnetic
Resonance
Imaging
(MRI).
Mammography and ultrasonography are difficult to interpret because the sensitivity of VFUHHQLQJ LV DIIHFWHG E\ WKH LPDJH TXDOLW\ DQG WKH UDGLRORJLVW¶V OHYHO RI H[SHUWLVH (Rangayyan et al., 2007; Bronstein et al., 2014).
On the other hand, MRI images are clear and have better contrast between most of the breast regions (Croshaw et al., 2011). For that reason, MRI is used for breast screening. However, some normal healthy breast tissue regions in MRI breast images have similar intensity values of tumour regions which may affect segmentation processes. Therefore, the MRI breast images need to be enhanced using image processing techniques. These help to assist radiologists in detecting the masses also. Image processing techniques are important to help radiologists in interpreting the images and to reduce the number of false positive and false negative diagnoses.
4
1.3 Breast MRI Tumour Segmentation Breast MRI CAD are computer programs that have been created using a combination of computer algorithms and image processing methods to provide help in detecting of suspected tumour regions (Andreea et al., 2011). The advantages of using CAD systems of detection breast tumours in screening technologies have been proven in several studies (Andreea et al., 2011; Leichter I et al., 2000; Arora et al., 2008; Li et al., 2002). CAD advantages could be concluded in fast detection, accuracy, assisting radiologists in finding dense breasts that might be overlooked. However, CAD systems still need improvements in overcoming the disadvantages of existing systems. The most common disadvantages of breast CAD systems are false-positive results in many breast images are produced, small tumours misdetection and the need for user manual interaction (Andreea et al., 2011).
1.4 Problems and Motivations Tumour segmentation approaches that have been used in Breast MRI CAD systems are categorized into supervised, unsupervised and semi-supervised. The performance of previous work using supervised segmentation scored higher accuracy than other approaches (Azmi et al., 2011b). However, prior knowledge is required and the process becomes difficult, expensive, and involves a lot of time. On the other hand, unsupervised methods need no prior knowledge but their performance is lower than other methods.
One of the key challenges in breast MRI segmentation approaches is developing automatic algorithms for different stages of the segmentation process. The majority of the
5
approaches required user manual inputs in one or more phases; either pre-processing phases or the tumour segmentation phase. The user inputs usually need a manual selection process of the initial and threshold variable values, selecting initial seeds pixels from the images or drawing windows around the tumour region of interests.
Another problem is the breast skin-line exclusion, which is an important pre-process in breast tumour segmentation. This process is important because of the similarity in the intensity level between the skin-line and tumour regions. Figure 1.5 shows an example of a breast MRI where the grey level intensity of skin-line and tumour are similar. From Figure 1.5, it can be observed that the average of grey level intensity for skin-line region is 194.7, while the average of grey level intensity for tumour region is 199.3. The close grey level intensity averages between skin-line regions and tumour regions could lead the tumour segmentation process to unsuccessful results. This process would be more important in the next stage of breast tumour segmentation.
6
and Multilevel Thresholding (Yen et al., 1995; Arora et al., 2008; Liao et al., 2001; Yan et al., 2005). These methods recorded high performances with images that have contrast between background and objects distributed equally in the whole image such as document images, human and nature images. However, these methods failed to achieve high performance in distinguishing the suspected tumour regions from most of unwanted tissues regions in the breast MRI images (Xiao et al., 2013).
Seeded Region Growing (SRG) (Adams & Bischof, 1994) is another vital process that has been used frequently in tumour segmentation approaches. SRG assigns neighbouring pixels based on certain characteristics. SRG needs two important values to complete the process. The first value is the initial seed pixel for starting the process and second value is the threshold value to end the process. These values are usually predefined or selected manually by the user. The challenge in SRG is to select a suitable initial seed and determining suitable threshold value automatically without the need for user input.
1.5 Research Objectives Based on the above-mentioned problems in Section 1.4, this study focuses on efforts to provide solutions to the subject of segmenting breast MRI tumours automatically. Therefore, the main goal of this research is to develop a computer aided segmentation approach to detect and segment breast MRI tumours automatically. This goal will be achieved by the following objectives:
9
1-
To develop a computer aided segmentation approach to segment breast MRI tumours automatically.
2-
To develop integrated method to exclude MRI breast skin-line from the image automatically.
3-
To propose automatic global image Thresholding method to breast tumours regions of interests.
4-
To modify SRG method by selecting initial seed and determine threshold value automatically to segment the tumour regions in breast MRI.
1.6 Scope of the Study In order to achieve the objectives of the research, it is important to identify the scopes. This study aims to develop an approach to segment tumour regions from breast MRI images. The study covers segmentation and pre-processing methods that lead to the main goal. All proposed methods are applied and tested on the registered dataset, i.e. Reference Image Database to Evaluate Therapy Response (RIDER)(U.S. National Cancer Institute, 2007).
10
1.7 Book Outline The book consists of five chapters and the outline of the book is structured as follows:
Chapter 1 - Introduction Chapter 1 presents an overview of MRI breast tumour and computer aided approaches for breast tumour segmentation. Problem statement, research scope as well as research objectives are presented in this chapter.
Chapter 2 - Literature Review In this chapter, a thorough review is covered on Breast MRI tumour segmentation approaches and their categories such as supervised, unsupervised and semi-supervised. Followed by, a review on previous works used for breast skin-line exclusion. In addition, a review on the different but relevant image processing methods in this study are also covered such as image thresholding, Seeded Region Growing algorithms and different image clustering methods, and a number of fundamental morphological operations.
Chapter 3 ± Methodology This chapter presents the overall approach used in this study to perform MRI breast tumour segmentation. Main contributions of the research are highlighted in this chapter which include data acquisition phase followed by pre-processing phase. In breast skinline exclusion phase, an integration approach of two methods is proposed to achieve the segmentation and removal of the skin-line. Subsequently, the tumour searching areas are minimized by eliminating non-suspected tumour regions using proposed general Thresholding method. Tumour segmentation phase is divided into two new methods that
11
automate region growing seed. Both proposed methods are presented to automatically select the necessary variable factors (initial seed and threshold value) of SRG. This chapter ends with describing the different evaluation approaches that used to obtain the results.
Chapter 4 - Results and Discussion The results and discussion for the proposed MRI breast tumour segmentation approach are discussed. The performance of the proposed enhancement and segmentation techniques are compared with previous techniques. The performance is measured in terms of evaluation approaches.
Chapter 5 - Conclusions and Future Work Finally, the last chapter draws the conclusions and highlights the contributions of the research. Some areas and number of interesting directions to be pursued are detailed as future work.
12
CHAPTER 2 LITERATURE REVIEW
2.1 Introduction This chapter reviews the methods and techniques that have been used in breast segmentation approaches in general and MRI breast tumour segmentation specifically. It starts with a description of the different breast screening modalities and comparing MRI to other techniques such as mammography and ultrasonography. The role of CAD systems for breast MRI is then highlighted.
Breast MRI tumour segmentation approaches are reviewed according to their categories such as supervised, unsupervised and semi-supervised. Subsequently, the importance of breast skin-line exclusion process is explained along with a review of previous works used to segment and remove the skin-line from breast images. Next, different image processing methods are investigated, especially methods that have been used or enhanced such as thresholding, seeded region growing algorithms and different image clustering methods, and a number of fundamental morphological operations.
2.2 Breast Screening Modalities Besides self-checking and physical examination for potential breast cancers, different screening techniques are used for more accurate examination. Three breast screening
13
modalities have been widely used in hospitals and clinical centers. The modalities are mammography, ultrasonography and Magnetic Resonance Imaging (MRI).
2.2.1 Mammography Mammography is a non-invasive X-ray technique, widely used for mass screening. This involves the process of exposing the breast to a small amount of ionizing radiation to produce details of the image (Hussain et al., 2011). Mammography is widely available as an image screening modality as it is able to produce acceptable imagining of abnormalities and its capability to show indirect calcifications (Croshaw et al., 2011; Bronstein et al., 2014). However, mammography has its weaknesses and limitations. These weaknesses appear in restricted dynamic range, contrast features and identifying very small tumours (Andreea et al., 2011). Examples of mammogram images (Suckling et al., 1994) are shown in Figure 2.1.
Figure 2.1 Examples of mammogram images (Suckling et al., 1994) 14
2.2.2 Ultrasonography Ultrasonography is also a non-invasive screening modality whereby sound waves are used to produce visualisation of the breast. The ultrasound image may be useful if a lump contain solid or fluid (Hussain et al., 2011). Ultrasonography has many advantages to offer such as it may detect tumours that may not be clearly determined whether it is solid or liquid in mammography. Besides, the technique causes less discomfort, incurs less expenditure and does not cause any side effects healthwise (Lee et al., 2010).
On the other hand, ultrasound images could have false positive results that lead to incorrect diagnosis (Sheng & Pei-Hong, 2013). Further, ultrasonography is not widely available in clinical centres and it requires highly trained and skilled experts to operate the devices (Elmore et al., 2005; Sheng & Pei-Hong, 2013). Examples of breast ultrasound images (Prapavesis et al., 2003) are shown in Figure 2.2.
Figure 2.2 Examples of breast ultrasound images (Prapavesis et al., 2003)
15
2.2.3 MRI Screening MRI screening is a non-invasive imaging technique. It has been widely used for medical imaging, including brain, spine, bones, joints and breast screening. It is based on both radio frequency and magnetic fields. The radio frequency pulses influence the arrangement of the resonant nuclei and thereby produce a detectable signal (Weaver et al., 1991). While the produced images by mammogram demonstrate the contrast between soft tissue and bone density, MRI on the other hand, produces clear and crisp images, which provides a better contrast between different kinds of soft tissues. For that reason, MRI is used for breast screening, i.e. to explore the small details between breast tissues. Although this is valuable information, the presented data still needs to be interpreted by the radiologist (Gardiner, 2010). Image processing techniques are used to help MRI radiologists to interpret the breast MRIs and to reduce the number of false-positive diagnoses (Lehman et al., 2006).
Figure 2.3 Examples of breast MRI (U.S. National Cancer Institute, 2007) 16
2.3 CAD for Breast MRI CAD systems are used with image processing algorithms to assist MRI radiologists in improving the quality of the breast MRIs, detecting the tumour masses and to reduce the number of false-positive detection (Lehman et al., 2006). CAD algorithms are developed to detect tumours within body organs (Li et al., 2002; Verma & Zakos, 2001). Several segmentation and classification techniques are used for the different modalities of medical images such as X-Ray and ultrasound. In recent years, statistical methods, vision based methods, wavelets, and fractals were proposed for breast mass detection (Chen et al., 2002; Chen et al., 2005; Kuo et al., 2001). Furthermore, artificial intelligence methods too have been proposed for classifications such as Artificial Neural Networks (ANN) and Fuzzy Logic (Mousa et al., 2005; Cheng & Cui, 2004; Chen et al., 2002). However, the researches on breast MRI CAD systems are still few compared to the other parts of the human anatomy.
2.4 Breast MRI Tumour Segmentation Approaches Image segmentation techniques could be categorized into two main approaches which are supervised and unsupervised approaches. Breast MRI tumour segmentation approaches likewise, follow the same categories. In addition, some researchers have also proposed semi-supervised approaches or mixed systems (Azmi et al., 2011b), which will be reviewed later in this section.
17
2.4.1 Supervised Approaches In the supervised approach (also known as model based segmentation), the numerical characteristics such as mean and variance of the classes of objects in the image are known in advance by the analyst and it is used in the training step. The training step is used to learn the specific objects to be detected. Then, the system has to be able to detect and classify new images depending on the presence or absence of similar objects. The training step cover examples with and without the object present (Oliver et al., 2010). There are several popular supervised algorithms such as; K-Nearest Neighbours (KNN), Support Vector Machine (SVM) and the Bayesian Method.
Yao et al. (2009) proposed supervised method for breast MRI tumour segmentation based on SVM classifier. In this method, firstly the breast region is segmented, excluding the chest and out-of-body regions from further processing. For every pixel, the texture features are extracted. Wavelet transform is applied to extract frequency features. In the training stage, a progressive feature selection is conducted to choose effective features, and a committee of SVM is generated as the classifier. This classifier is applied to new data for pixel-by-pixel classification. This method can be used to address different image protocols. In addition, it reduces the number of selected features. However, it needs to be trained on at least ten different cases to achieve the required results.
KNN classifier was used by Rabiei et al. (2007). They used contextual information based on the shape of the objects of interest and the temporal kinetic signal in their work. The approach was demonstrated on the application of using machine vision to classify breast disorders into four classes using KNN classifier. The approach scored high tumour
18
segmentation results in complex backgrounds.The main disadvantage of this approach is that the user needs to start the initial segmentation by detecting a binary window manually to select the breast region and to disregard the rest of the regions of the images that is related to tissues of the chest and heart.
Another supervised approach is based on the Bayesian method and Markov random field model proposed by Wu et al. (2006). This approach analysed the features of the breast MRI images and classified them as tumour or non-tumour regions. The estimation of the class membership is made using the Iterative Conditional Mode (ICM) method. The prior distribution of the class membership is modelled as a multi-level logistic model, a Markov Random Field model in which the class membership is assumed to depend upon the nearest neighbours only. The likelihood distribution is assumed to be Gaussian. This approach could be efficiently applied for real-time segmentation in health centres. However, the parameters of each Gaussian distribution are manually selected as an estimated representative of the class.
2.4.2 Unsupervised Approaches Unsupervised segmentation consists of partitioning the image into a group of regions which are distinguished and constant with respect to particular features, such as intensity level, size or texture (Fu & Mui, 1981; Oliver et al., 2010; Omran, 2005). Image clustering approaches, region based approaches, thresholding approaches and contour approaches, are all members of unsupervised segmentation family. The unsupervised methods have several advantages over the supervised methods. Supervised methods require an analyst
19
to identify the characteristics of the images in the dataset in advance before starting the segmentation. On the other hand, unsupervised methods automatically find distinct classes, which dramatically reduce the work of the analyst. In addition, some characteristics of objects may not be identified in advance for the supervised methods. However, unsupervised methods automatically flag these characteristics in the image (Davies, 2004; Omran, 2005).
Chen et al. (2006) presented a fuzzy c-means (FCM) clustering based method for the segmentation of breast tumours in MRI images. The proposed tumour segmentation method needs a human input to select the region of interest (ROI), followed by image enhancement within the selected ROI. Then FCM was used to classify the enhanced ROI. Finally, the tumour was segmented by applying the thresholding on the tumour membership map and connected component labelling, and hole filling on the selected object. The method has the potential for accurate, efficient, and consistent segmentation of breast MRI tumours. The main weakness of this method is the need of user input to recognize the ROI as a rectangle shape to start the segmentation process. Markercontrolled watershed method is proposed by Cui et al. (2009) to segment malignant tumours from breast MRI images. The semi-automatic approach is started by determining ROI ellipse manually. Gaussian mixture modelling is then applied to detect the internal and external markers for watershed segmentation for the tumour. Initial findings show good segmentation results that correlate wHOOZLWKWKHUDGLRORJLVW¶VPDQXDOGHVFULSWLRQRI tumour volume. The main drawback in this approach is the use of a computer mouse to draw an elliptical ROI on a selected region containing the target tumour.
20
2.4.3 Semi-Supervised Approaches Semi-supervised classification method is proposed by Azmi et al. (2011a) to segment breast tumour in MRI based on texture analysis in order to achieve a high performance. This method has two main stages; in the first stage, Improved Self-Training (IMPST) classifier is trained only with a labelled image. In the next stage, nondeterministic unlabelled data is obtained through simple Thresholding, and this classifier is retrained with them to reach high accuracy. Although the accuracy and precision of segmented images are improved based on initial results, the weakness of this approach appears in the need of the user to draw a small window to identify the tumour ROI region.
The supervised, unsupervised and semi-supervised methods are explored in Azmi et al¶V study (2011a). In their comparison study on the MRI Breast RIDER dataset (U.S. National Cancer Institute, 2007), they found that the supervised segmentation methods such as; KNN, SVM and Bayesian and the semi-supervised method such as; self-training and improved self-training (IMPST) lead to high accuracy. However, prior knowledge is required. Hence, the process becomes difficult, expensive, and involves a lot of time. On the other hand, unsupervised methods such as; FCM needs no prior knowledge but their performance is low (Azmi et al., 2011a).
The following sections review various image processing techniques that involve in MRI breast tumour segmentation approaches, starting with reviewing breast skin-line exclusion methods. Meanwhile, Table 2.1 concludes comparison of breast MRI tumour segmentation approaches.
21
Table 2.1 Comparison of breast MRI tumour segmentation approaches Method
Description
FCM clustering segmentation based (Chen et al., 2006)
FCM is used to classify ROI. Then tumour is segmented by applying thresholding, connected component labelling and hole filling methods.
The method has the The need of user potential for input to recognize accurate, efficient, ROI as rectangle and consistent shape to start the segmentation of segmentation breast MRI tumours. process.
Advantages
Disadvantages
Bayesian and Markov random field based (Wu et al., 2006).
Features are analysed and classified as tumour or non-tumour regions using combination of methods such as Bayesian and Markov random field and Gaussian distribution.
This approach could be efficiently applied for real-time segmentation in health centres.
The parameters of Gaussian distribution are estimated manually selected as representative of the class.
KNN classifier based approach (Rabiei et al., 2007).
By using shape and temporal kinetic signal information, the approach is demonstrated using KNN to classify breast disorders into four classes using
The approach scored high tumour segmentation results in complex backgrounds.
Manually drawing window to select the breast region and to disregard the rest of regions of the images that contain related tissue of chest and heart.
SVM based Breast region is x It can be used to approach segmented, texture address different (Yao et al., 2009) features are extracted image protocols. and selected, and then x The number of SVM is generated as the selected features is classifier. reduced.
The approach needs to be trained on at least ten different cases to achieve the required results.
MarkerGaussian mixture controlled modelling is applied on watershed ROI to detect internal method (Cui et and external markers for al., 2009) watershed segmentation for the tumour.
Initial findings show The use of a good segmentation computer mouse to results that correlate draw an elliptical well with the ROI on a selected UDGLRORJLVW¶VPDQXDO region containing description of the target tumour. tumour volume.
Improved SelfTraining IMPST (Azmi et al., 2011a)
The accuracy and precision of segmented images are improved based on initial results
IMPST has two stages; classify trained labelled image and classify unlabelled data is obtained through simple Thresholding.
22
The weakness of this approach appears in the need of the user to draw a small window to identify the tumour ROI.
From analysing the number of Breast MRI tumour segmentation approaches, i.e.; supervised and unsupervised. It is found that the performance of supervised segmentation approaches has higher accuracy than unsupervised approaches (Azmi et al., 2011b). However, prior knowledge is required and the process becomes difficult, expensive, and involves a lot of time. On the other hand, unsupervised methods need no prior knowledge but their performance is lower than other methods. In addition, for the majority of both approaches, the main weakness is that manual ROI selection or other user inputs are necessary steps to complete the segmentation processes successfully.
2.5 Breast Skin-Line Exclusion Approaches Breast skin-line exclusion is a significant pre-process in breast tumour computersegmentation. Limiting the area to be processed into a specific target region in a breast image would increase the accuracy and efficiency of tumour segmentation algorithms (Yapa & Harada, 2007).
The intensity level between the breast skin-line and the tumour regions is similar in the majority of MRI breast cases. Thus, segmenting and removing the breast skin-line region methods will be discussed in this study. Without this, segmentation errors may occur if not managed correctly (Solves et al., 2012a). A number of approaches have been developed in order to exclude skin-line regions from breast images.
Yapa and Harada (2007) proposed fast-marching method to estimate and segment breast skin-line from mammogram images. The method grouped intensity level information and
23
gradient information on fast marching speed function, and end point constraint are then introduced to ensure that the boundary expanded within the proposed region and stopped when the boundary reached the end-point. The experiments showed that in most cases, the method managed to segment breast skin-line. This algorithm is also capable of extracting the nipple approximately from the mammogram when it is available in the profile. The weakness of this method is the need of user manual interaction to select an initial seed region and an end point on the breast edge (Yapa & Harada, 2007).
Zhang et al. (2010) proposed automatic segmentation approach for breast skin-line by developing seed determination method and region growing. To determine the location of seed point, two points are located manually near the start and end of the diagonal reference line, and then the proposed algorithm is applied based on calculating the mean of the intensity of local area around these seeds to determine the seed. The technique of object recognition is used to detect the size of the initial segmented area to decide one or two seed points used in the region growing process. A 3x3 neighbour matrix is applied to measure the mean of a local area in the process of region growing. This approach managed to deal with the low contrast, noisy mammogram images. However, the weakness of this approach is that two points have to be selected manually to start the process. In addition, the approach is suitable for mammogram images only.
Solves et al. (2012b) presented a novel method of skin-line segmentation that aimed to obtain as many pixels belonging to the real skin-line as possible. This method uses a curvature flow filter to minimize MRI noise and to prepare it for image clustering analysis. Image clustering with four clusters is able to differentiate image parts, and two
24
of those parts contain skin-line pixels. Although the experiments demonstrate that, in most of the cases, the method of skin-line segmentation is accurate, the two clusters contain some dense tissue that needs to be reanalysed with another clustering process to extract skin-line pixels correctly.
Marti et al. (2007) presented a novel methodology to obtain the breast skin-line in mammographic images. The clue behind the proposed approach is based on finding the skin-line by using a contour growing technique. The growing process is based on subsequent similar ideas of attraction and regularisation found in active contours. The approach started by computing a scale space representation of the image in order to perform edge detection using different scales. Subsequently, an initial seed point lying in the skin-line contour is located based on an estimation procedure. The initial seed is used for contour growing process, which is started based on enlarging and adapting a contour using different criteria. The researchers tested their approach on mammogram dataset. Most of the segmentation results were considered satisfactory when the noise in the images were in minimum background. On the other hand, in some cases the approach did not obtain acceptable segmentation when the images had large amount of noise. This had led to poor estimation of the initial seed and to non-uniform breast intensity distribution which yielded under segmented images. Another weakness of this approach is the weighting factors of the growing process were established empirically, experiencing the fact that extreme values of any of the factors did not obtain satisfactory results.
Raba et al (2005) presented an approach to segment the breast skin-line. In this approach, global histogram is calculated and smoothed with a Gaussian operator. N consecutive
25
percentage of bright pixels is tested to obtain N thresholds. Each value is used in order to threshold the image and obtains masks which are overlapped. The region is defined by the boundary of the smallest threshold to the boundary of the largest one which is statistically evaluated to calculate the mean of the grey level which is used as the final threshold value. The result of applying this threshold is a collection of different regions. The largest one is the union of the breast and the pectoral muscle. The largest region used is extracted using the Connected Component Labelling algorithm. The region of interest of the breast has been extracted from the pectoral muscle using the region growing algorithm.
Nie at el. (2008) used dynamic searching in their research to eliminate breast skin-line. This work is based on the assumptions of the darkness of the air signal, the brightness of breast fat tissue and the intermediate grey level intensity between them for the skin-line. The dynamic searching is based on the change of grey level gradient. After the breast-air boundary is determined, the coordinates of each pixel along the boundary are recorded. The slope of the tangential line at each position is calculated from the nearby three pixels along the curve, and the dynamic searching is achieved along the vertical direction. The upper border of the skin-line is detected after the negative gradient from skin-line to air is determined; the lower border of the skin-line is detected once the positive gradient from skin-line to fatty tissue is determined. The region between the two margins is considered as skin-line. However, this method assumed that the thickness of the skin-line is about three pixels and is searched in this range. If the skin-line is thicker than three pixels, only a certain thickness of that amount will be excluded.
26
From reviewing the previous work, it can be noticed that even though in most cases the goal of skin-line exclusion is achieved, still most of the approaches have two main weaknesses, i.e. the first weakness is that they are designed for mammogram images only and not for MRI images. While the second weakness is that the user interaction is essential to accomplish the goal. The breast skin-line exclusion approaches are summarized in Table 2.2.
The next section discusses the various image processing techniques that are used in breast MRI tumour segmentation approaches such as image thresholding, image clustering, region based methods and morphological operations.
27
Table 2.2 Comparison of breast skin-line exclusion methods Method
Description
Advantages
Disadvantages
New approach of breast skin-line segmentation (Raba et al., 2005)
The skin-line is segmented by using combination of global histogram and thresholding methods.
Most of experimental results on mammogram dataset show accurate segmentation.
This method developed and applied on mammogram images not on MRI.
Fast-marching and region based method (Yapa & Harada, 2007).
By grouping intensity level and gradient information on fast marching speed function, to be used in region based method for segmenting skinline. The approach is based on finding the skin-line by using a contour growing technique.
x In most cases, the x User interaction to
A novel methodology to obtain the breast skin-line in mammographic images (Martí et al., 2007).
method segmented select the initial breast skin-line. seed and an end x It is capable of point on the breast extracting nipple edge. approximately if it x It applied on is available in the mammogram profile. images not on MRI. Most of the results x Poor estimation of were considered the initial seed if satisfactory when the images are the noise is in noisy. minimum in the x The weighting image background. factors of the growing were established empirically.
Dynamic searching for skin-line segmentation (Nie et al., 2008).
Dynamic searching x Applied on MRI is based on the images. change of intensity x Combined two differences between approaches to the skin-line and enhance the other regions. segmentation.
Region growing segmentation approach for breast skin-line (Zhang et al., 2010)
Based on developing seed determination for region growing on the skin-line.
Image based (Solves 2012b)
clustering Based on cluster analysis. Image et al., clustering with 4 clusters is applied, 2 of them contain skin-line.
If the skin-line is thicker than three pixels, only a thickness of that amount will be excluded.
This approach x Manual selection managed to deal of two points to with the low start the process. contrast, noisy x It is suitable for images. mammogram images only and not for MRI. Experiments The 2 clusters demonstrate that, in contain some dense most of the cases, tissue that needs the method of skin- reanalysed with a line segmentation is new clustering to accurate. extract skin-line.
28
2.6 Image Processing Techniques Many image processing techniques could be involved in the breast tumour segmentation. In this section, matters pertaining to the following will be explained. These are image thresholding and the related approaches of selecting the threshold value automatically, the uses of SRG in medical image segmentation, and the methods of SRG automatic selection for initial seeds. The different image clustering methods will also be explored and compared to address the most suitable method to be used. The section ends with a description of a number of fundamental Morphological operations such as image thinning, erosion, dilation, opening and connected component labelling.
2.6.1 Image Thresholding Methods Image thresholding is one of the fundamental operations in most image processing DSSOLFDWLRQV,QPRVWFDVHVLWLVXVHGWRFDWHJRUL]HWKHLPDJH¶VSL[HOVLQWRWZRFODVVHV namely, the foreground and the background. The pixel considers the portion as the IRUHJURXQG LI WKH SL[HO¶V LQWHQVLW\ H[FHHGV D FHUWDLQ WKUHVKROG YDOXH RWKerwise it is considered as background pixels (Gonzalez, 2008).
Manual or automatic selection of a threshold value separates the image into two intensity regions which are white (intensity value = 255) and black (intensity value = 0). Selecting the appropriate threshold value is subject to change, according to the type of image and the goal of the thresholding process. For example, one of the uses of thresholding is in the process of transforming a grey level image into its binary form. Hence, thresholding has an important role to play, whether it is for segmentation, pre-processing, and detecting
29
the suspected regions or subjects in the specific image in question, according to the application and problem to be solved.
The challenge in applying thresholding on breast MRI is how to automatically select the suitable threshold value that could detect the suspected regions. The difficulty of selecting WKHVXLWDEOHWKUHVKROGYDOXHLVEHFDXVHRIWKHVLPLODULW\RIWKHWXPRXUUHJLRQ¶VLQWHQVLW\ with the other regions. Therefore choosing an incorrect value could cause either detecting VHYHUDOXQZDQWHGUHJLRQVDVDWXPRXU¶VUHJLRQRILQWHUHVWRUGHWHFWLQJLQFRPSOHWHWXPRXU regions.
2.6.1.1 Automatic Thresholding Various methods have been explored to find the best approach for thresholding techniques and how to select the best threshold value automatically.
Iterative thresholding approach is proposed by Ridler and Calvard (1978) after an initial threshold is estimated, i.e. mean image intensity. Pixels above and below the threshold are assigned to the white and black classes respectively. The threshold is iteratively reestimated as the mean of the two class means. Succeeding iterations thresholding provides increasingly cleaner extractions of the image areas (Ridler & Calvard, 1978). However, this method may face difficulties when dealing with very noisy images, because the weighing scale may be misplaced (Anjos et al., 2010). Another misclassification difficulty may happen to small foreground objects in the images. Grey Level Histogram selection is one of the most popular automatic thresholding. Otsu (1979) presented a
30
method to select a threshold automatically from a grey level histogram that was derived from the viewpoint of discriminant analysis. An optimal threshold is selected by the discriminant criterion. The proposed method is characterized by its nonparametric and XQVXSHUYLVHGQDWXUHRIWKUHVKROGVHOHFWLRQ$OWKRXJK2WVX¶VPHWKRGUHPDLQVRQHRIWKH most popular choices for global thresholding techniques, it does not work well for many real world images where a significant overlap exists between the pixel intensity values of the objects and the background for un-even and poor illumination (Ray & Saha, 2007).
Entropy based concept is used also in image thresholding. Kapur et al. (1985) has developed an algorithm for choosing a threshold from the level histogram derived by using the entropy concept from information theory. The optimal thresholded is calculated by finding the maximum of the summation of foreground and background entropies (Kapur et al., 1985). Brink and Pendock (1996) and Li & Lee (1993) proposed sequential methods to enhance the method of finding optimal threshold using minimum cross entropy thresholding technique (MCET) based on Gaussian distribution. Alayli and ElZaart (2013) enhanced the prior works by developing a fast iterative algorithm for MCET. This method is applied on mammographic images and results obtained were encouraging. The advantage of entropy thresholding methods is the use of global and objective property RIWKHLPDJH¶VKLVWRJUDP7KLVDOJRULWKPFDQEHXVHGIRUVHJPHQWDWLRQSXUSRVHVEHFDXVH of its general nature. On the other hand, entropy thresholding methods, when images are corrupted with noise or irregular illumination they produce multimodal histograms that does not guarantee the optimum threshold selection process, because no spatial correlation is considered (Prasad et al., 2011).
31
Fuzzy set image thresholding proposed by Huang and Wang (1995), is an image thresholding method which is based on the concept of fuzzy sets and the definition of membership function. It utilizes the measure of fuzziness to evaluate the fuzziness of an image and to determine an adequate threshold value (Huang & Wang, 1995). The advantage of Fuzzy image thresholding is that it is the automatic threshold value which is generated without prior knowledge; it is not based on the minimization of a criterion function. As a result it is proper for image intensity values distributed gradually (Li et al., 2012). However Fuzzy set based thresholding methods achieved poor results when the images have high vagueness, low level of contrast, and grey level ambiguity, where the methods need extra optimize fuzziness to reduce the ambiguity (Arifin et al., 2009).
Multilevel thresholding approaches are another automatic thresholding. Yen et al. (1995) presented new criterion for multilevel thresholding based on the consideration of two factors. The first one is the difference between the thresholded and original images and the second one is the number of bits required to represent the thresholded image. Multilevel thresholding is developed in number of studies such as Arora et al. (2008), Liao et al. (2001) and Yan et al. (2005). Analyses show that the required mathematical processes in these methods are fewer than that are needed in entropy criterion methods. However, multilevel method still has the problem of time consuming processing when the number of threshold increases (Hidayat et al., 2013). Table 2.3 concludes the different methods of image thresholding with their advantages and disadvantages.
32
Table 2.3 Comparison of automatic Thresholding methods Type
Description
Advantages
Disadvantages
Iterative thresholding approach (Ridler & Calvard, 1978; Anjos et al., 2010).
Pixels are assigned as white and black based on initial threshold, the threshold is iteratively reestimated as the mean of the two class means.
Succeeding iterations thresholding provides increasingly cleaner extractions of the image areas.
Difficulties with very noisy images, because the weighing scale may be misplaced. Another difficulty may happen with images having small foreground objects.
Grey level histogram selection (Otsu, 1979).
An optimal threshold is selected by the discriminant criterion. The proposed method is characterized by its nonparametric and unsupervised nature of threshold
This method is one of the most popular choices for global thresholding methods.
Entropy based thresholding (Kapur et al., 1985; Prink & Pendock, 1996; Li & Lee, 1993; Alayli & ElZaart, 2013)
Based the entropy concept from information theory. optimal thresholded is calculated by finding the maximum of the summation of foreground and background entropies
Fuzzy set thresholding. (Huang & Wang, 1995; Li et al., 2012; Arifin et al., 2009).
Based on the concept of fuzzy. It utilizes the measure of fuzziness to evaluate the fuzziness of an image to determine an adequate threshold value
Automatic threshold value is generated without prior knowledge. It is proper for image intensity values distributed gradually.
Poor results when the images have high vagueness, low level of contrast, and grey level ambiguity.
Multilevel thresholding (Yen et al., 1995; Arora et al., 2008; Liao et al., 2001; Yan et al., 2005).
Based on the Analyses show that consideration of two the required factors; the difference mathematical between the processes in these thresholded and methods are fewer original images, and than that are needed in the number of bits entropy criterion required to represent methods. the thresholded image.
Multilevel method still has the problem that lasts a long time process when the number of threshold increases
It does not work well for many real world images where a significant overlap exists between the pixel intensity values of the objects and the background for uneven and poor illumination. The uses of global and If images are noisy or objective property of have irregular WKH LPDJH¶V illumination, it does not histogram. It can be guarantee the optimum used for segmentation threshold selection purposes because of process. its general nature.
33
The previously mentioned methods of automatic thresholding such as Iterative Thresholding, Grey level Histogram, Entropy based, Fuzzy Thresholding and Multilevel Thresholding are studied here. These methods perform well when they are applied on images that have contrast between background and objects distributed equally in the whole image such as document images, human and nature images. However, besides the weaknesses stated in Table 2.3, these methods are less successful in achieving high performance in distinguishing suspected tumour regions from other regions in breast MRI images.
In the next section, the various Seeded Region Growing methods will be discussed. The uses of SRG in medical images and its automatic versions will be reviewed too.
2.6.2 Seeded Region Growing (SRG) SRG is one of the region based segmentation approaches. The process starts from initial seed pixel and assigns neighbouring pixels to the segmented region, based on the examination of the characteristics of the pixels in the region which is defined in advance. The region is grown step by step by adding more similar neighbouring pixels to the initial ones, the process ends when the growing region pixels reach the defined threshold value. SRG algorithm which was proposed by Adams and Bischof (1994) is widely used in medical images today because it is effectively segments different types of images. The most critical parameters of this method is selecting the suitable initial seed pixel for starting the process and selecting the threshold value to end the process (Petrou & Petrou, 2010).
34
2.6.2.1 SRG in Medical Images According to empirical studies of the efficiency of the Seeded Region Growing on MRI brain segmentation, the results are promising for both light and dark abnormalities. Nevertheless, a lower performance in dark abnormalities segmentation is produced as it has slightly lower correlation values in all conditions as compared to light abnormalities (Khalid et al., 2010; Ibrahim et al., 2011).
,Q0HLQHO¶VVWXG\RQ05,EUHDVWVHJPHQWDWLRQ(Meinel, 2005), SRG was also used. The experiments on breast tumour segmentation returned robust results. However, this approach needed an initial seeds to start the SRG process besides SRG threshold value. Both values are to be specified by the user which is then used to find the anticipated locations of the tumours.
The SRG Feature Extraction (SRGFE) algorithm has been proposed on cervical cancer screening. This algorithm extracted four cervical cells features which are size of nucleus, cytoplasm, grey level of nucleus and cytoplasm. Correlation test was applied between data extracted using the proposed SRGFE algorithm with the data extracted manually. The performance is described as high if the correlation value is higher than 0.8, moderate if equal or more than 0.5 but less than 0.8 and weak if less than 0.5 (Devore, 2000). The data extracted using SRGFE algorithm gave high correlation values (0.89 - 0.97) for different features. Still, the user needs to determine the region of interest to select the initial seed pixel. The user also needs to determine the threshold value (Mat-Isa et al., 2005).
35
Khalid et al. (2010) used SRG in segmenting the tumours regions of brain MRI images. The seed pixel is chosen close to the centre of the region. The growing process yields a particular region having pixels with comparable features as the seed pixel. The application of this SRG includes of a specification of two variables. These are the window size and the absolute difference of the grey level value between the grown regions with the seed pixel. Although this work produces consistent results for both light and dark abnormalities in overall brain abnormalities segmentation performance, but it still has drawbacks which are the SRG seed point and threshold point are predefined with estimated values.
2.6.2.2 Methods of Automatic SRG Numerous researches have worked on developing algorithms that generate initial seed pixel automatically. Wu et al. (2008) proposed texture feature-based automated SRG approach on abdominal organ segmentation. The approach comprises of two stages. After the MRI images are loaded into the approach, each pixel in the ROI is processed and three features: co-occurrence, semi variogram (spatial dependence) and Gabor texture feature are extracted. Automated SRG algorithm is applied on texture feature space and in the end, a region grown out of the seed is obtained. The advantage of this algorithm is that it allows minimum user intervention. This is helpful for batch work. However, this approach does have drawbacks. Texture feature based methods all have the assumption that the region should have texture homogeneity. For tissues with complex texture, this approach may not work well.
36
Shan et al. (2008) developed an automatic seed point selection method for SRG on ultrasound breast images. The method needs no prior information or training processes. Both the homogeneous texture features and spatial features of the breast tumours are taken into account. The method is composed of five steps; speckle reduction, iterative threshold selection, boundary-connected regions deletion, regions ranking and seed point determination. However, some cases failed because of the shadowing effects of areas which have similar intensity as the tumour and also right below the tumour.
Fan & Elmagarmid (2001) proposed an automatic edge-oriented seed generation technique to automate SRG algorithm. The colour edge detection technique is first performed to obtain the simplified geometric structures of a colour image. The centroids of the neighbouring labelled colour edges are then taken as the initial seeds for region growing. The method results demonstrate high accuracy results when it is applied on coloured images and video dataset. However, the edge oriented SRG algorithm may induce over segmentation problem because the colour edges may be over detected for the texture images and thus result in redundant seeds.
Image clustering methods are reviewed in next section. Meanwhile, Table 2.4 summarized the different approaches of SRG methods with their advantages and disadvantages.
37
Table 2.4 Comparison of SRG methods Method
Description
Advantages
Disadvantages
Automatic edge oriented SRG algorithm (Fan & Elmagarmid, 2001).
Colour edge detection is performed, and then centroids of the neighbouring labelled colour edges are then taken as the initial seeds for region growing.
The method results demonstrate high accuracy results when it applied on coloured images and video dataset.
This method may induce over segmentation problem because the colour edges may be over detected for the texture images and thus result in redundant seeds.
Breast MRI tumour The SRG was used segmentation by to segment the (Meinel, 2005). tumour from breast MRI images in the proposed approach.
The experiments on breast tumour segmentation returned robust results.
Initial seed and SRG threshold value need to be specified by the user.
The SRGFE (Mat-Isa et al., 2005).
This algorithm applied on cervical cancer screening by extracted cells features; size of nucleus, cytoplasm, grey level. Automated SRG algorithm is applied on extracted texture features from abdominal region MRI images. Both the homogeneous texture features and spatial features of the breast tumours are taken into account.
The data extracted using SRGFE algorithm gave high correlation value when compared with data extracted manually. This approach allows minimum user intervention. This is helpful for batch work.
The user needs to determine the region of interest to select the initial seed pixel. The user also needs to determine SRG threshold value. This approach may not work well if tissues have complex texture, or heterogeneity.
This method needs no prior information or training processes.
Shadowing effects of areas which have similar intensity as the tumour and also right below the tumour.
SRG includes two variables; the window size, the absolute difference of the grey level value between the grown regions with the seed pixel.
Consistent results for both light and dark abnormalities in overall brain abnormalities segmentation performance.
The SRG seed point and threshold point are predefined with estimated values.
Texture featurebased automated SRG approach (Wu et al., 2008).
Automatic seed selection method for SRG on ultrasound breast images (Shan et al., 2008). SRG segmentation for brain MRI abnormalities (Khalid et al., 2010).
38
From reviewing SRG methods, it can be noticed, SRG is very effective in terms of segmenting specific object from the image, especially medical images (Gómez et al., 2007; Heimann et al., 2004). The performance of the method is depending on the accuracy of initial seed point and threshold values are determined for starting and ending the segmentation process. However, the limitations of SRG studies that covered in this chapter are range between the manual or predefined selection of the parameters (initial seed and threshold value) to the problems of shadowing effects and over detection regions.
2.6.3 Image Clustering Methods Image clustering process is classifying the image into groups that have comparable properties. The resulting clusters should contain pixels that are similar between themselves (high intra cluster similarities) and dissimilar to pixels of other clusters (low inter cluster similarities) (Puzicha et al., 1999; Saraswathi & Allirani, 2013). Image clustering is divided into two main families; hierarchical and partitional.
2.6.3.1 Hierarchical Clustering Hierarchical clustering is based on classifying the similarity between each pair of pixels to be clustered (Saraswathi & Allirani, 2013). Hierarchical clustering organizes the results into a sequence of tree shape. The whole image array is the root of the tree; the leaf node represents individual pixel clusters. The middle nodes in the tree signify the similarity between the pixels in the clusters. The Hierarchical clustering algorithms could be
39
grouped into three main groups; single-link methods, complete-link methods and minimum-variance methods.
In single-link hierarchical clustering (Sneath & Sokal, 1973), it merges two clusters based on the smallest distance between the pixels from two clusters. Complete-link hierarchical clustering (King, 1967), it incorporates the maximum distance between the pixels in clusters. Minimum-variance hierarchical clustering (Ward, 1963), merges two clusters to reduce the cost function to produce a different cluster (Jain & Dubes, 1988; Jain et al., 1999; Saraswathi & Allirani, 2013).
As for the advantages of hierarchical clustering methods, it is not necessary to define the number of clusters before starting the clustering process. The methods are independent of the initial condition. However, Hierarchical clustering methods have disadvantages such as; when pixels allocated to a cluster, they could not reallocated to a different cluster. Another disadvantage is the clusters might be overlapped because of the absence of information (Frigui & Krishnapuram, 1999).
2.6.3.2 Partitional Clustering Partitional clustering divides the image pixels into partitions. Each portion organizes group of pixels into one cluster. This type of clustering identifies a specific number of clusters before starting the procedure.
One of the advantages of partitional clustering is treating the large data in easier way than hierarchical clustering (Saraswathi & Allirani, 2013). Another advantage is dealing with 40
the clustering as an optimization problem by reducing the searching criteria. In general, the advantages of the Hierarchical clustering could be considered as the disadvantages of the partitional clustering. In the same time, the advantages of partitional could be considered as disadvantages of hierarchical clustering (Omran, 2005).
One of the most popular clustering methods is K-Means. This method is based on squared error. K-Means selects random initial clusters to start assigning the pixels to the closest cluster in the image based on the similarity. The assigning process repeats for every pixel until a convergence criterion is met (Jain et al., 1999; MacQueen, 1967). Numerous researchers derived their methods from K-Means such as ISODATA clustering which was introduced by Ball & Hall (1967). Then, developed the integrating and dividing of the extended classes in K-Means. Huang (2002) proposed integration method of K-Means with
hierarchical
methods
in
Synergistic
Automatic
Clustering
Technique
(SYNERACT). Many enhanced clustering versions of K-Means method are proposed over the years such as Rosenberger & Chehdi (2000), Hamerly & Elkan (2002), Fahim et al. (2006), Park et al. (2008) and Nazeer et al. (2011). K-Means is simple and effective and when variables are many, K-Means could be computationally less time consuming than hierarchical clustering methods. However, K-Means has some disadvantages such as it depends on the initial condition and the number of the portions needs to be identified initially (Davies, 2004).
Fuzzy clustering was presented for the first time by Ruspini (1969) based on Fuzzy set theory by Zadeh (1965). FCM was developed by Dunn (1973). FCM assigns the pixel in the image clusters according to the degree of Fuzzy membership function. By using this
41
advantage, FCM can overcome the problem of the clusters overlapping is some images. Fuzzy clustering have been developed and applied on different applications by several researchers such as Bezdek (1980), Pham & Prince (1999), Timm & Kruse (2002), Ahmed et al. (2002) and Cai et al. (2007). Although FCM is better than K-Means
method, it has some drawbacks such as; it is time-consuming in the computational process, it is sensitive to the initial values (speed, local minima) and it is sensitive to noise (Suganya & Shanthi, 2012).
Expectation Maximization (EM) method estimated the required factors using some existing data (Hamerly & Elkan, 2002). EM uses combination of Gaussian to estimate the clusters of the image dataset. According to Alldrin et al. (2003), the results of using this method show that the performance of EM is comparable to K-means results. However, they found that EM is not successful when it is applied on high dimensional data because of numerical precision problems. Another weakness of this method is related to Gaussians which are frequently collapsed to delta functions.
One of the recent methods is Particle Swarm Optimization (PSO) image clustering, which is initially proposed by Omran (2005) and updated by the same researcher in Omran (2005) and Omran (2006). In PSO image clustering, a group of particles is preserved. Every individual particle denotes a contender solution for the optimization. The particles are distributed in the image array, each particle created its position according to the distance between its best position and the distance from the best particle of the swarm (Kennedy & Eberhart, 1995; Omran et al., 2005; Yuhui & Eberhart, 1998). The performance of the particle is calculated using fitness function established for problem.
42
Varied versions of fitness functions for PSO image clustering are presented in (Ouadfel et al., 2010a), (Ouadfel et al., 2010b) and (Wong et al., 2011).
Table 2.5 summarizes the different approaches of image clustering methods with their advantages and disadvantages. From Table 2.5, it can be noticed that all discussed clustering methods have their features and weaknesses. However, PSO image clustering produces better outcomes in performance compared to other methods. This conclusion is based on the results of comparison studies between PSO image clustering and other methods. Omran (2005) in his study compared number of image clustering methods such as Fuzzy image clustering, K-Means, Expectation Maximization, Hierarchical clustering and PSO image clustering using MRI images and standard grey level images. In addition, Wong et al. (2011) compared PSO image clustering with K-Means using the standard images, both comparable studies results conclude that PSO image clustering produces better outcomes in performance of inter- and intra-cluster distances.
43
Table 2.5 Comparison of image clustering methods Type
Description
Advantages
Disadvantages
Hierarchical clustering; Singlelink (Sneath & Sokal, 1973), Complete-link (King, 1967), Minimumvariance (Ward, 1963)
Based on classifying the similarity between each pair of pixels to be clustered into a sequence of tree shape.
There is no requiring defining clusters number in advance. The methods are independent of the initial condition.
Pixels could not reallocate to a different cluster. Clusters might be overlapped because of the absence of information.
K-Means (Jain et al., 1999; MacQueen, 1967; Ball & Hall, 1967; Huang, 2002; Fahim et al., 2006; Hamerly & Elkan, 2002; Nazeer et al., 2011; Park et al., 2008; Rosenberger & Chehdi, 2000).
Selection random initial clusters for assigning pixels to the closest cluster in the image based on the similarity, the assigning process repeats for every pixel until a convergence criterion is met.
It is simple and effective. K-Means could be computationally less time consuming than hierarchical clustering methods.
It depends on the initial condition, and the number of the portions needs to be identified initially.
Fuzzy Clustering (Ruspini, 1969; Dunn, 1973; Ahmed et al., 2002; Bezdek, 1980; Cai et al., 2007; Pham & Prince, 1999; Timm & Kruse, 2002).
Fuzzy Clustering assigns pixels in image clusters according to the degree of Fuzzy membership function.
Fuzzy Clustering can overcome the problem of the clusters overlapping.
Time-consuming in the computational process, sensitivity to the initial values (speed, local minima) and sensitivity to the noise.
Expectation Based on factors Maximization estimation using method some existed data. (Hamerly & Elkan, Combination of 2002; Alldrin et al., Gaussian is used to 2003). estimate clusters of the image.
Results of using this method show that the performance is comparable to KMeans results.
x It is not successful when it is applied on high dimensional data. x Gaussians are frequently collapsed to delta functions.
PSO Image Clustering (Ouadfel et al., 2010a; Ouadfel et al., 2010b; Wong et al., 2011; Omran, 2005; Omran et al., 2005; Omran et al., 2006)
PSO image clustering produces better outcomes in performance of interand intra-cluster distances than other methods.
Performance of the method is based on selecting the right fitness function for problem.
Particles are distributed; each particle creates its position according to the distance between its best position and distance from best particle of swarm.
44
2.6.4 Fundamental Morphological Operations Different essential morphological operations are used in this study which are thinning, erosion, dilation, opening and connected component labelling. These methods would be adapted in the actual proposed study. The fundamentals of morphological operations are explained in the following sub-sections.
2.6.4.1 Morphological Thinning Operation Thinning is an iterative neighbourhood process that produces a skinny demonstration of an object in the image. The theory of thinning a binary image of an object is associated with medial axis transformations in that it generates a representation of an approximate axis of symmetry of a shape by successive removal of pixels from the border of the object. Thinning can be defined as a logical neighbourhood operation where object pixels are removed from an image. Figure 2.4 illustrates the thinning operation before and after applying thinning operation.
The pixels removal must be controlled to a certain degree because of that, group of conditions should be followed to achieve the right thinning. The first limit is that the pixel must lie on the border of the object. The second limit is that the removal of a pixel should not affect the object's connectedness, i.e. the number of skeletons after thinning should be the same as the number of objects in the image before thinning. This problem depends on the manner in which each pixel in the object is connected to every other pixel. The third limit is to preserve an object's length, where object pixels which are neighbouring to just one other pixel, must not be deleted (Vernon, 1991).
45
(a)
(b)
Figure 2.4 Morphological thinning operation. (a) Original binary image, (b) after applying thinning
2.6.4.2 Morphological Dilation and Erosion Operations Dilation operation causes the border of foreground regions in binary image to expand or grow in size. Erosion operation causes the border of foreground regions to shrink. The amount and the manner that they dilate or erode are subjected to the choice of the structuring element which is necessary part of dilation and erosion operations. Structure element is an array mask with arbitrary shape of 1s and 0s that are used in morphological operations (Young et al., 1998). The dilation and erosion are defined morphologically as in Equation (2.1) and (2.2) respectively.
ܵ ْ ܫൌ ڂאௌ ܫ
(2.1)
ܵ ٓ ܫൌ ൛หܵ ܫ كൟ
(2.2)
where I is binary image array, S is the structure element array, p is pixel in the array, ٓ is erosion and ْ is dilation.
46
From the definitions and the equations, it can be noticed that the dilation operation aim is the opposite of erosion operation, i.e. a growth of the object into those entire background pixel which edging the object. Thus, erosion shrinks an object while dilation enlarges it (Vernon, 1991). Figure 2.5 and 2.6 show erosion and dilation operations, respectively.
(a) (b) (c) Figure 2.5 Morphological erosion operation; (a) Original binary image ܫ, (b) Structure element ܵ, (c) the image after applying erosion ܵ ٓ ܫ
(a) (b) (c) Figure 2.6 Morphological dilation operation; (a) Original binary image ܫ, (b) Structure element ܵ, (c) the image after applying Dilation ܵ ْ ܫ
47
2.6.4.3 Morphological Opening Operation The combination of dilation and erosion operations produces an important process; which is opening operation (Young et al., 1998). Opening operation aims to smooth the border of a foreground and excludes tiny objects from the image (Gonzalez & Woods, 2007). The opening is defined morphologically as in Equation (2.3).
ܵ ל ܫൌ ሺܵ ٓ ܫሻ ْ ܵ
(2.3)
Where I is binary image array, S is the structure element array, ٓ is erosion, ْ is dilation and לis opening. The complete procedure of Opening operation is illustrated in Figure 2.7.
(a)
(b)
(c)
Figure 2.7 Morphological opening operation; (a) Original binary image ܫ, (b) Structure element ܵ, (c) the image after applying opening ܵ ל ܫൌ ሺܵ ٓ ܫሻ ْ ܵ
48
2.6.4.4 Connected Component Labelling This standard technique assigns to each connected component of the binary image a different label. The labels are usually natural numbers starting from one to the total number of connected components in the input image. The algorithm scans the image from left-to-right and top-to-bottom. On the first line containing foUHJURXQG¶V SL[HOV RI WKH REMHFWDXQLTXHODEHOLVDVVLJQHGWRHDFKFRQWLJXRXVUXQRIREMHFW¶VSL[HOV)RUHDFKSL[HO of next and succeeding lines, the neighbouring pixels on the previous line and the pixel to the left are tested. If any of these neighbouring pixels has been labelled, the same label LVDVVLJQHGWRWKHFXUUHQWREMHFW¶VSL[HO. Otherwise, the next unused label is used. This procedure continues to the bottom line of the image. Upon completion of this process, a connected component may contain pixels having different labels because when they are H[DPLQHGWKHQHLJKERXUKRRGRIDQREMHFW¶VSL[HOWKHSL[HOWRLWVOHIWDQGWKRVHRQWKH previous line might have been labelled differently. Such a situation must be detected and remembered. After the scanning process, the labelling is completed by unifying conflicting labels and reassigning unused labels (Haralick, 1981). The complete procedure is illustrated in Figure 2.8.
(a) (b) Figure 2.8 Connected Component Labelling; (a) Original binary image, (b) after applying Connected Component Labelling 49
2.7 Summary This chapter has focused on a review of previous studies in MRI breast tumour segmentation approaches with the related methods and algorithms of image processing.
The literature covered several aspects such as breast screening, CAD systems, breast MRI tumour segmentation approaches, breast skin-line exclusion approaches, image thresholding methods, SRG approaches, image clustering methods and Morphological operations.
In Chapter 3, number of image processing techniques are developed to overcome the limitations addressed in the literature. An automatic approach is proposed based on combination of unsupervised algorithms, the new approach attempts to overcome previous approaches weaknesses by achieving high accuracy segmentation results without the need to prior knowledge. An integrated method is proposed to exclude skinline regions. The integrated approach tries to segment skin-line automatically from MRI images without the need for user interferes. A new method of general thresholding is presented; this method attempts to achieve better results than previous methods in finding WXPRXU¶V UHJLRQV RI interest in breast MRI images by considering tumour intensity features. For tumour segmentation, two different automated versions of SRG are proposed. The proposed approaches intend to avoid user interaction; the approach automatically generates the regions of interest, SRG initial seed to start the segmentation process and SRG threshold values to use as end point.
50
CHAPTER 3 METHODOLOGY
3.1 Introduction This chapter discusses the different phases of computer aided segmentation approach for breast MRI tumour. The proposed methods are introduced in detail as depicted in the methodology flowchart in Figure 3.1.
The methodology is comprised of six main phases; data acquisition, pre-processing which includes two processes; image filtering and image splitting. Followed by, breast skin-line exclusion phase. After that, image thresholding phase, tumour segmentation phase and evaluation phase.
Modified SRG method is used in the segmentation phase; the modification comprises two proposed methods of selecting SRG variable factors. Each method tries to achieve the same goals. These goals are automated searching for the most suitable seed point value and threshold value. The methodology ends with the discussion of the evaluation criteria that are used to evaluate the performance of each process. The whole perspective of the methodology is shown as in Figure 3.1.
After the initial pre-processing phase of the methodology is applied, an integration method for skin-line regions exclusion is presented. The elimination of skin-line regions is necessary for this approach because of the close intensity level of the skin-line to the
51
WXPRXU¶VOHYHORILQWHQVLW\ZKLFKFRXOGOHDGWRIDOVHWXPRXUVHJPHQWDWLRQ7KHSURSRVHG approach integrates two algorithms, which are Level Set Active Contour (LSAC) algorithm is used to segment the breast skin-line border; the Morphological Thinning algorithm is used to delete the detected breast skin-line.
Then, a new automatic global method is presented during the stage of image thresholding. This method is used to detect the regions of interest of the suspected tumour in the breast MRI; the theory of mean maximum raw thresholding method will be described in detail supported by the pseudo codes of the algorithm.
Subsequently, breast tumour segmentation stage is applied. Two methods of modified SRG are proposed in this chapter. The first proposed method is breast MRI tumour using modified automatic SRG, while the second proposed method merges SRG with PSO image clustering in one integrated method. Both approaches are presented as automated methods to select the best variable factors (initial seed and threshold value) for seeded region growing. The proposed methods are compared later with previous work.
Lastly, the evaluation criteria is discussed, where it is divided into three sub sections. Each section explains the evaluation approaches and equations used for each stage of the methodology.
52
Start
Data Acquisition Proccess
If the image is Axial
Yes
No
Pre-Processing
Breast Skin-Line Exclusion
Splitting the image vertically if it is Axial
Noise Filtering
Skin-Line Segmentation Using Level Set Algorithm
Skin-Line Removal Using Thinning Algorithm
Image Thresholding using MMRT
Tumour Segmentation using modified Seeded Region Growing
BMRI-MASRG
BMRI-SRGPSOC
Automatic Seed Selection
Automatic Seed Selection
Automatic SRG Threshold Value Selection
Automatic SRG Threshold Value Selection
Segmented Tumor
Evaluation Criteria
End
Figure 3.1 Flowchart of the proposed segmentation approach for breast MRI tumour
53
3.2 Data Acquisition Phase The methodology explained earlier is applied and tested on Reference Image Database to Evaluate Therapy Response (RIDER) MRI Breast, registered for U.S. National Cancer Institute (U.S. National Cancer Institute, 2007).
3.2.1 The RIDER Dataset The RIDER Breast MRI dataset is downloaded from the National Biomedical Imaging Archive (NBIA) (U.S. National Cancer Institute, 2007). This website belongs to the U.S. National Cancer Institute. The dataset also include Ground Truth (GT) segmentation, which have been identified manually by expert radiologists. 40 images (24 malignant and 16 benign) from the dataset with their GTs are used in the experiments as test images. GT is used as a benchmark for performance evaluation of segmentation methods in our experiments. All images are Axial 512 X 512 pixels. The original benign and malignant images with their GTs are shown in Figures 3.2 and 3.3.
RIDER is chosen for this study because it is registered dataset and has been used by number of previous approaches. The evaluation results of previous work compared to GT of the dataset are published (Azmi et al., 2011b). Those results are assisting in investigating the significant of the proposed approach of this study. The images are read in DICOM format and convert into 8-bit grey level scale, thus each pixel of image has intensity value (0 to 255). The converting process is using Matlab function rgb2gray(). Where, Grey Level = 0.299 × Red + 0.587 × Green + 0.114 × Blue (Blahuta et al., 2012).
54
(a)
(b)
Figure 3.2 Malignant breast image from RIDER dataset;(a) Original malignant image. (b) Thresholding GT of the image.
(a)
(b)
Figure 3.3 Benign breast image from RIDER dataset; (a) Original benign image. (b) Thresholding GT of the image.
55
3.3 Pre-Processing Phase The pre-processing phase is the first phase that is executed after the data collection. At this study, the image is split into two sub-images; the right breast image and the left breast image. This process is used only if the MRI breast image is Axial (i.e. the image is taken from the perspective of the patient from head to toe). This process is skipped if the image is Sagittal (i.e. the image is taken according to the lateral view) because of the fact that the sagittal image shows only one side of the breast i.e. either the right or the left. Therefore, the splitting process in the case of the sagittal image is not necessary.
The splitting process can be done by finding the middle of the X-coordinate of the image and splitting the image vertically from that point. Figures 3.4 shows example of the image splitting process applied on sample images. The median filter is then applied which used in number of studies to reduce noise from MRI and breast MRI images as standard filter, these studies such as Bohare et al. (2011), Sadri et al. (2012) , Li et al. (2013), Krishnakumar & Parthiban (2014) and Alshanbari et al. (2015). The purpose of using the median filter here is to reduce the speckle noise, while the boundaries and features are kept intact (Yu et al., 2009). Figure 3.5 shows examples of breast MRI images before and after applying median filter.
56
(a)
(b) Figure 3.4
(c)
Results of image splitting; (a) Original image. (b) Left side. (c) Right side
57
Figure 3.5
(a)
(b)
(c)
(d)
(e)
(f)
Applying Median filter, (a), (c) and (e) samples of breast MRI images. (b), (d) and (f) after applying median filter on the same samples.
58
3.4 Breast Skin-Line Exclusion Phase using Proposed Integration Method of LSAC and Morphological Thinning Algorithms This process is a vital process. Similar intensity levels of the skin-line and the other parts of the breast image such as dense tissue and tumour could possibly lead to faulty tumour segmentation if the skin-line is not removed correctly. The integrated approach of MRI breast skin-line exclusion consists of two main stages. These are skin-line border segmentation using LSAC which is applied on the resultant image of the pre-processing phase. Followed by, skin-line removal using morphological thinning. The following is a detailed description of the processes involved. 3.4.1 Breast Skin-Line border Segmentation Stage To segment the breast skin line border, LSAC algorithm (Li et al., 2005) is used. Level set formulation without re-LQLWLDOL]DWLRQ ZKLFK LV VSHFLDOO\ NQRZQ DV &KXQPLQJ¶V algorithm) is included as one of the active contour's detection algorithms. Active contours are dynamic curves that move toward the mass border. An external energy moves the zero level curves toward the mass border using the edge indicator function g that is defined in Equation (3.1). ൌ
ͳ ͳ ȁ כȁଶ
(3.1)
Where I is the image, ܩఙ is the Gaussian kernel with a standard deviation ı. By changing the ı parameter value and the number of the iterations ܰௌ , the resultant detection is changed. Figure 3.6 shows the different results after processing with LSAC algorithm whereby different values of ı and ܰௌ are applied.
59
(a)
(b)
(c)
(d)
(e)
(f)
Figure 3.6 Different results after the application of LSAC algorithm with different YDOXHVRIıDQGܰௌ : (a) ı=3, ܰௌ =100; (b) ı =1.5, ܰௌ =100; (c) ı =3, ܰௌ =300; (d) ı =1.5, ܰௌ =300; (e) ı =3, ܰௌ =700; (f) ı =1.5, ܰௌ =700.
60
Experiments have been done using trial and error method on the dataset to select the best parameters. Different combination of ıand ܰௌ are used such as (ı=3, ܰௌ =100), (ı =1.5, ܰௌ =100), (ı =3, ܰௌ =300), (ı =1.5, ܰௌ =300), (ı =3, ܰௌ =700) and (ı =1.5, ܰௌ =700). From Figure 3.6, it can be seen that with using parameters (a) ı=3, ܰௌ =100 and (b) ı =1.5, ܰௌ =100, the borders were not segmented correctly and more pixels from background were included. While the parameters used in (c) ı =3, ܰௌ =300 and (d) ı =1.5, ܰௌ =300; (e) ı =3, ܰௌ =700, the segmentation ignored parts of the breast regions. The most accurate segmentation was using the parameters have been selected in (f). Where, the best value for ܰௌ is 700 while the best value of ı is 1.5.
3.4.2 Breast Skin-Line Removal Stage Morphological Thinning algorithm is used in the proposed method to delete amount of pixels equals the reference thickness (6.4 pixels). In Morphological Thinning algorithm as described in Lam et al. (1992). The image is divided into two distinct sub-fields. Then in the first sub-iteration, pixel p from the first sub-field is deleted if and only if the first three conditions are true. In the second sub-iteration, pixel p from the second sub-field is deleted if and only if, the first two and fourth conditions are true. The eight neighbours of p, are ݔଵ , ݔଶ , ݔଷ « ଼ݔstarting with the east neighbour and numbered in counterclockwise order as shown in Figure 3.7. The following are the conditions.
First condition: ܺு ሺሻ ൌ ͳ
(3.2)
61
Where ସ
ܺு ሺሻ ൌ ܾ
(3.3)
ୀଵ
ൌ ൜
ͳǡʹǦͳ ൌͲሺʹ ൌͳʹͳ ൌͳሻ Ͳǡ
(3.4)
When i=4 then ଽ = ଵ .
Figure 3.7 The pixel p and its eight neighbours pixels
Second condition: ʹ ݉݅݊ሼ݊ଵ ሺሻǡ ݊ଶ ሺሻሽ ͵
(3.5)
Where ସ
݊ଵ ሺሻ ൌ ݔଶିଵ ݔڀଶ ୀଵ
62
(3.6)
ସ
݊ଶ ሺሻ ൌ ݔଶ ݔڀଶାଵ
(3.7)
ୀଵ
When k=4 then ଽ = ଵ .
Third condition: ሺݔଶ ݔ שଷ ݔ שҧ଼ ሻ ݔ רଵ ൌ Ͳ
(3.8)
Forth condition: ሺ ש שത ସ ሻ רହ ൌ Ͳ
(3.9)
The two sub iterations formulate the main iteration of the thinning algorithm. The thinning level depends on the number of iterations. Whenever the number of iterations is increased, there would be more shrinking of the LPDJH¶V border.
The Morphological Thinning algorithm only accepts a binary version of the image. Therefore, the resultant image after the LSAC algorithm would be converted to binary image using Mean Maximum Raw Thresholding method which will be described later in this chapter.
63
Furthermore, after the thinning procedure, the binary image is reconverted into its original grey level representation. This process can be done by multiplying the binarized image matrix by the original image matrix. Morphological Thinning algorithm could be applied in multiple iterations. In each iteration, the algorithm deletes one pixel width layer from the boundary of the object. The number of layers deleted from the boundary of the object is equal to the number of iterations (்ܰ ).
In order to experiment the choice of ்ܰ value, the thickness of skin-line in pixels needs to be investigated first to give initial knowledge of how many pixel layer need to be removed. Number of researchers tried to measure the thickness of breast skin-line in their studies. The results of five studies are used in this work to identify the breast skin-line thickness range. These studies are; Willson et al. (1982), Pope et al. (1984), Huang et al. (2008), Paredes (2009) and Molleran and Mahoney (2014). Table 3.1 shows the different VWXGLHV¶UHVXOWVRIthickness of each study in millimetre (mm) and pixel.
Table 3.1 Breast skin-line thickness by different studies in mm and pixel units Average range (mm)
Average range (pixel)
Willson et al., (1982)
1.9
7.2
Pope et al., (1984)
1.7
6.4
Huang et al., (2008)
1.6
6.1
Paredes, (2009)
1.9
7.2
Molleran & Mahoney, (2014)
1.3
4.9
Studies
64
From Table 3.1, it can be noticed that the five studies have achieved close results of skinline thickness. The results were ranged between 1.3 mm (4.9 pixel) to 1.9 mm (7.2 pixel). Based on the range of measures in pixels shown in Table 3.1, number of values of ்ܰ experimented on the dataset to select the best iteration number. Different ்ܰ values tested such as (்ܰ =1, ்ܰ =2, ்ܰ =3,«, ்ܰ =7) to cover the range shown in Table 3.1. Experiments proved that only ୦ =7 deleted the right amount of pixel layers from the boundary of skin-line. Figure 3.8 shows example results after the Morphological Thinning algorithm is applied using three different iteration values (்ܰ =1, ்ܰ =3 and ்ܰ =7). From Figure 3.8, it can be seen that ்ܰ =7 is the best iteration value to remove the whole skin-line comparing to other ்ܰ values which left parts of skin-line unremoved.
In the next section, a proposed automatic image thresholding method is presented in order to minimize the searching area for tumours by distinguishing the suspected than unsuspected regions.
65
(a)
(b)
(c)
(d)
(e)
(f)
Figure 3.8 Results after applying Morphological Thinning algorithm with three different iteration numbers on the resultant image of LSAC algorithm; (a) after applying thinning (்ܰ =1) on binary image; (b) after reconverting thinning image (்ܰ =1) to its original grey level; (c) after applying thinning (்ܰ =3) on binary image; (d) after reconverting thinning image (்ܰ =3) to its original grey level; (e) after applying thinning (்ܰ =7) on binary image; (f) after converting thinning image (்ܰ =7) to its original grey level. 66
3.5 Image Thresholding Phase using Proposed Mean Maximum Raw Thresholding Method Image thresholding is one of the most important image processing processes as an essential pre-processing phase along with tumour segmentation process. The importance RIWKHSURFHVVLVWRGLVWLQJXLVKWKHSRWHQWLDOWXPRXU¶VUHJLRQVDQGWRUHPRYHPRVWRIWKH unwanted parts in the breast MRI in order to prepare the image for more accurate tumour detection in the next stage of segmentation.
Mean Maximum Raw Thresholding method (MMRT) considers the tumour intensity features in order to propose a new algorithm that uses the sub-window concept. This is to automatically seek the optimum threshold value that divides the image into two classes, i.e. ROI and the background.
3.5.1 Mean Maximum Raw Thresholding Algorithm (MMRT) The method is based on the hypothesis that the tumour region is one of the highest intensity regions in the breast images (Asselin et al., 2012; Nagarajan et al., 2013; Just, 2014). The proposed method in this work is an automatic selection for the suitable threshold value. The algorithm divides the image into sub-windows. Sub-windows produce more accurate prediction for the best value in each part. For each window, the algorithm searches for the maximum value (M) in each of the row in the window. This is because the tumour region is the region having the highest intensity area amongst other regions in the breast images. Subsequently, the algorithm saves (M) temporarily. This process is repeated for all of the rows until the last. Then, a summation of the temporarily
67
stored values is calculated. The window mean maximum raw (WMMR) is then calculated by dividing the summation value by the number of rows in the window as described in Equation (3.10).
¦ M N 1
WMMR
i 0
(3.10)
i
N
where i LVWKHZLQGRZ¶VURZN is the number of rows, M is the maximum intensity pixel value in the row.
This process is repeated for all of the sub-windows of the image and the WMMR of each window is calculated and stored in a matrix. The matrix elements are ranked and the maximum WMMR among all the windows will be considered as the threshold value for the thresholding process. This divides the image into two intensity regions, i.e. the black region are those that are below the threshold value while the white region would be those that are above the same value.
Figure 3.9 demonstrates the process of automatic selection of the threshold value using MMRT. Figure 3.10 (a) shows example array divided into nine windows. Figure 3.10 (b) shows selection of maximum intensity values in each row of each window. Figure 3.10 (c) illustrates the calculations of MMRT as described in Equation (3.10) to find the maximum value of all windows in the array. Figure 3.10 (d) shows applying the selected threshold value on the array. The pseudo code of the proposed MMRT automatic threshold value selection and the thresholding process is given in the Figure 3.9.
68
The size of the window is important in this method and it is related to the size of the LPDJH¶VDUUD\7ZRSRLQWVULVHIURPWKHH[SHULPHQWVRIFKRRVLQJWKHZLQGRZ¶VVL]H7KH ILUVW SRLQW LV YHU\ VPDOO ZLQGRZ¶V VL]H LV QRW SUHIHUDEOH LQ WKLV PHWKRG EHFDXVH DV described, the method searches for the maximum pixels in each row of the window to calculate WMMR, with small windows such as; (3 x 3), (5 x 5) or (7 x 7) the possibility of having WMMR ؆ 255 is higher with any small white dot in the image. Especially, if WKHLPDJH¶VVL]HLVPRUHWKDQ ; That will lead to the second point, which is WKHUHODWLYHQHVVEHWZHHQWKHLPDJH¶VVL]HDQGWKHZLQGRZ¶VVL]HMMRT would not give WKH VDPH SHUIRUPDQFH ZLWK GLIIHUHQW LPDJHV¶ VL]HV LI the size of the window is fixed. 7KHUHIRUHWKHZLQGRZ¶VVL]HLVFDOFXODted in this method based on Equation (3.11).
ܹ݅݊݀ ݓᇱ ݁ݖ݅ܵݏൌ ݀݊ݑݎሺο ݁݃ܽ݉ܫ כԢ݁ݖ݅ܵݏሻ
(3.11)
From experiments, it found the proper value of ο is in range of 0.05 and 0.1.
In the next section, the proposed methods used for tumour segmentation are described. Two approaches are proposed to automate the selection of initial seed and threshold values of SRG.
69
REPEAT UNTIL last window in the image SET zero to Max_Row_Intensity, Sum_Max FOR each row in the window FOR each column in the window IF 3L[HO¶V,QWHQVLW\!0D[B5RZB,QWHQVLW\ THEN 6(73L[HO¶V,QWHQVLW\WR0D[B5RZB,QWHQVLW\ ENDIF ENDFOR Sum_Max=Sum_Max+Max_Row_Intensity SET zero to Max_Row_Intensity ENDFOR :LQGRZB0HDQB0D[B5RZ 6XPB0D[QXPEHURIZLQGRZ¶VURZV Put Window_Mean_Max_Row in matrix END REPEAT LOOP SET 7KUHVKROGB9DOXHWRPD[LPXPHOHPHQWLQ:LQGRZB0HDQB0D[B5RZ¶VPDWUL[ FOR each row in the image FOR each column in the image IF 3L[HO¶V,QWHQVLW\7KUHVKROGB9DOXH 6(7WR3L[HO¶V,QWHQVLW\ ELSE 6(7]HURWR3L[HO¶V,QWHQVLW\ ENDIF ENDFOR ENDFOR Figure 3.9 Pseudo code of MMRT
70
(a)
(b) Figure 3.10 The process of automatic selection of the threshold value using MMRT; (a) example array divided into nine windows. (b) Selection of maximum intensity values in each row of each window. (c) The calculations of MMRT to find the maximum value of all windows, (d) applying the selected threshold value on the array.
71
(c)
(d) Figure 3.10 continued.
72
3.6 Breast MRI Tumour Segmentation Phase Using Two Proposed Methods In this study, the SRG algorithm (Adams & Bischof, 1994) for tumour segmentation is chosen because it is fast, simple and robust (Gómez et al., 2007). To apply the SRG, two variable factors should be determined first. These factors are usually selected manually. The first factor is the initial seed pixel that the SRG can start growing. The second factor is the threshold value for measuring the difference between the pixel and its neighbours. Two different methods of the automated SRG initial seed and SRG threshold selection are proposed in this chapter.
The first proposed method is Breast MRI Tumour using Modified Automatic Seeded Region Growing (BMRI-MASRG), while the second proposed method is based on particle swarm optimization image clustering (BMRI-SRGPSOC). Both methods are tested and evaluated based on their performance on the same dataset. The analysis of the performance results are also compared in Chapter 4.
3.6.1 Tumour Segmentation Preprocessing Both methods of automated SRG initial seed and SRG threshold selection start with the same pre-processing stages explained earlier in this chapter. This consists of splitting the image into two images if the image is axial. The breast skin is detected and deleted using an integrated approach of LSAC algorithm with Morphological Thinning algorithm (as explained in sections 3.3, 3.4 and 3.5).
73
3.6.2 Using Seeded Region Growing This method starts with an initial seed pixel and tries to compare its neighbourhood pixels with the seed according to some attributes, such as the intensity or the texture. It then merges them if they are similar enough.
In this study, after the proper initial pixel seed is selected for the two different approaches (in the next few sections in this chapter), the eight neighbouring pixels will be tested according to SRG threshold value (also in the two different approaches which will be explained in next few sections). The neighbouring pixel is considered to be in the segmented region if it is above the threshold value. Subsequently, the eight neighbours of the new pixel will be tested and sorted too. The process then continues in the same manner. Figure 3.11 shows the initial seed pixel and the eight neighbours.
Figure 3.11 The initial seed pixel and their eight neighbor pixels
The pseudo code of SRG algorithm is described by Adams & Bischof (1994) as shown in Figure 3.12.
74
Label seed points according to their initial grouping Put neighbours of seed point in the S list Delta to be a measure of how different pixel is from their neighbours DO Remove first pixel P from the S. FOR The eight neighbours of P IF neighbour of P which is already labelled (other than boundary label) has the same label THEN SET P to this label Update running mean of corresponding region IF neighbours of P is neither already set nor already in the S THEN Add it to S according to their value of Delta ELSE Flag P with the boundary label ENDIF ENDFOR UNTIL S is empty
Figure 3.12 Pseudo code of SRG algorithm is described by Adams & Bischof (1994).
The proposed automatic SRG methods (BMRI-MASRG and BMRI-SRGPSOC) are described in the next sections.
75
3.6.3 First Proposed Method: Modified Automatic Seeded Region Growing (BMRIMASRG) In this section, a modified version of SRG with automatic features is proposed. The modified method involves two algorithms, which are automatic seed selection to start SRG process and automatic threshold value selection to end SRG process. The seed selection algorithm consists of seven steps while the threshold value comprises of four steps.
3.6.3.1 Automatic SRG Seed Selection of BMRI-MASRG The method is based on the intensity and area features to select the best seed in the breast image automatically as illustrated in the block diagram of the method shown in Figure 3.10. This method is comprised of seven steps, which are explained below in detail.
1- Minimize the selection search regions by finding the suspected regions first by The
Mean Maximum Raw Thresholding algorithm (MMRT) (as explained in detail in section 3.5).
2- Apply morphological opening operation (erosion followed by dilation operations)
which remove the unwanted small white speckles in the image that does not belong to the tumour regions and also to enhance and smooth the boundary of the suspected regions (erosion, dilation and opening operations are explained in detail in Chapter 2 (sections 2.6.4.2 and 2.6.4.3).
76
3- /DEHOHDFKUHJLRQIRUHDV\DFFHVVWRWKHUHJLRQV¶SURSHUWLHV&RQQHFWHG&RPSRQHQW
Labelling method (CCL) is used to label the regions. CCL groups the connected pixels together and numbers each pixel in the same region with the same number (CCL is explained in detail in Chapter 2 (section 2.6.2.4)).
4- Measure two of the region properties for each suspected region which results from step
2. The properties are; (a) the mean intensity value of pixels and (b) the area which refers to the number of the pixels of the region. These selected properties are related to the tumour features. Before the property measuring process, it is necessary to revert the original grey level representation to the suspected regions (which has become white after the thresholding process). This process can be done by multiplying the binarized image matrix with the original image matrix.
5- Calculate the densities of the UHJLRQV¶LQWHQVLW\E\ GLYLGLQJWKH DUHDRIHDFKUHJLRQ
ZLWKWKHPHDQLQWHQVLW\RIWKHVDPHUHJLRQWKHUHJLRQ¶s density is calculated using Equation (3.12).
ܴ݁݃݅݊Ԣ ݕݐ݅ݏ݊݁ܦݏൌ
ܽ݁ݎܣ ݕݐ݅ݏ݊݁ݐ݊݅݊ܽ݁ܯ
(3.12)
6- Rank the suspected regions by their density values that have been calculated in step 5.
7KHKLJKHVWUHJLRQ¶VGHQVLW\YDOXHZLOOEHFKRVHQDVWKHPDLQVXVSHFWHGUHJLRQ 65 to search for the seed pixel within it. This assumption is made based on number of studies such as Asselin at el. (2012), Nagarajan at el. (2013) and Just (2014). Figure
77
3.14 (a) shows an example for the ranked regions based on their density value. Figure 3.14 (b) shows the bar chart for the ranked regions that indicate region number 6 ZKHUHWKHLQWHQVLW\YDOXHLVDURXQG DVWKHKLJKHVWUHJLRQ¶VGHQVLW\YDOXH
7- Search for the maximum intensity value pixel from the main suspected region. This is
because the tumour region is the region having the highest intensity area amongst other regions in the breast images.
It is important to be mentioned that after seed selection process is achieved. The breast image reverts to its form resultant from skin-line exclusion phase, where SRG starts from selected seed position to grow the segmentation region. SRG tests the pixels according to the threshold value.
The SRG threshold value of BMRI-MASRG is determined automatically in proposed method described in next section. Meanwhile, the block diagram and illustration figures for the automatic SRG seed selection of BMRI-MASRG are shown in Figure 3.13 and 3.14 respectively.
78
Start
Minimize the regions by MMRT Thresholding
Remove unwanted small speckles using Morphological Open
Label each region using Connected Component Labelling
Find the mean intensity and the area of each region
&DOFXODWHWKHGHQVLW\RIHDFKUHJLRQ¶V intensity
Rank the regions by their density value and determine the main suspected region
Select the maximum intensity pixel from the main suspected region
End
Figure 3.13 Block diagram which illustrates the processes of Automatic SRG Seed Selection of BMRI-MASRG
79
(a)
(b) Figure 3.14 5HJLRQUDQNLQJDFFRUGLQJWRWKHLUSL[HOV¶GHQVLW\YDOXHV D ,PDJH region flags. (b) Bar chart for the ranked regions according to the highest mean intensity values. 80
3.6.3.2 Automatic SRG Threshold Value Selection of BMRI-MASRG The Automatic SRG threshold value selection approach is based upon an estimated value that is use as end point for growing region process. This is in accordance to the difference between the mean intensity of the main suspected region and the mean intensity for the other whole breast regions that are neighboring the main suspected region, except for the breast skin-line regions. The approach comprises of four steps which are shown in Figure 3.15. The algorithm is described in four steps as below:
1- Find the mean intensity value (ݏݐ݊ܫ݊ܽ݁ܯௌோ ) for the main suspected region (SR) (that has been detected in the previous section, step 6).
2- Isolate the neighboring area of the SR from the other unwanted regions such as the black background, the breast skin-line (which was detected earlier) and SR itself. This is as stated in Equation (3.13).
݄ܾܰ݁݅݃݃݊݅ݎݑ ൌ ܹ݄݈݁ூ െ ሺ ݀݊ݑݎ݃݇ܿܽܤ݈݇ܿܽܤ ݊݅݇ܵݐݏܽ݁ݎܤ ܴܵሻ
(3.13)
3- Find the mean intensity for the
݄ܾܰ݁݅݃݃݊݅ݎݑ which is called here
(ݏݐ݊ܫ݊ܽ݁ܯே ).
4- Calculate the SRG threshold (்ܴܵܩ ) using the following Equation (3.14). ்ܴܵܩ ൌ ݏݐ݊ܫ݊ܽ݁ܯௌோ െ ݏݐ݊ܫ݊ܽ݁ܯே
81
(3.14)
Start
Find the mean intensity for the main suspected region
Determine the neighboring area of the main suspected region
Find the mean intensity for the neighboring area
Calculate the SRG threshold by subtract neighboring area from the main suspected UHJLRQ¶VPHDQLQWHQVLW\
End
Figure 3.15 Block diagram which illustrates the processes of Automatic SRG Threshold Value Selection of BMRI-MASRG
82
3.6.4 Second Proposed Method: Integrated Method of SRG and PSO Image Clustering (BMRI-SRGPSOC) BMRI-SRGPSOC is an integrated method that merges SRG with PSO image clustering to produce another modified version of SRG with automatic features is proposed. As explained for BMRI-MASRG method earlier in (section 3.63), BMRI-SRGPSOC is also involves two algorithms, which are; automatic seed selection to start SRG process and automatic threshold value selection to end SRG process.
3.6.4.1 Particle Swarm Optimization Image Clustering As explained earlier in Chapter 2 (section 2.6.3), PSO image clustering produces better outcomes compared with other clustering methods such as the K-means and Fuzzy Image clustering (Omran, 2005; Ouadfel et al., 2010a; Wong et al., 2011).
Applying the PSO image clustering would be organizing the image into groups whose members have similar intensity range. Therefore, each cluster represents a different intensity range of pixels in the image. Various versions of image clustering based on PSO have been proposed in Omran (2005), Ouadfel et al. (2010a), Wong et al. (2011). The method used in this study is proposed in Omran (2005) and enhanced by Wong et al. (2011) as below:
ܰ is the number of clusters to be formed, ܼ is the p-th pixel, ܥ is the subset of pixel vectors that form cluster j, ݔ ൌ ൫݉ଵ ǡ ǥ ݉ ǡ ǥ ǡ ݉ே ൯where ݉ refers to the j-th cluster centroid vector of the i-th particle.
83
1- Initialize each particle to contain ܰ randomly selected cluster means with random positions and velocities. 2- FOR t = 1 to ݐ௫ (maximum number of iterations) FOR each particle i -
FOR each pixel ܼ Calculate ݀൫ܼ ǡ ݉ ൯for all clusters ܥ Assign ܼ to ܥ where ݀൫ܼ ǡ ݉ ൯ ൌ ୀଵǡǥǡே ൛݀൫ܼ ǡ ݉ ൯ൟ ݀൫ܼ ǡ ݉ ൯represents the Euclidean distance between the p-th pixel ܼ and the centroid of j-th cluster of particle i.
-
Calculate the fitness function ݂ሺݔ ሺݐሻǡ ܼሻwhere Z is a matrix representing the assignment of pixels to clusters of particle i.
Update the personal best and the global best positions. Update the cluster centroids, velocity and position of the particle using Equations (3.15) and (3.16).
ݒ ሺ ݐ ͳሻ ൌ ݒݓ ሺݐሻ ܿଵ ݎଵ ሺݐሻ൫ݕ ሺݐሻ െ ݔ ሺݐሻ൯ ܿଶ ݎଶ ሺݐሻ ቀݕ ሺݐሻ െ ݔ ሺݐሻቁ
ݔ ሺ ݐ ͳሻ ൌ ݔ ሺݐሻ ݒ ሺ ݐ ͳሻ
(3.15)
(3.16)
Where ݔ is the current position of the particle, ݒ is the current velocity of the particle,ݕ is the best position that particle has achieved so far, ݕ is the location of overall best value,
84
w is the inertia weight, ܿଵ and ܿଶ are the acceleration constants, and ݎଵ ሺݐሻand ݎଶ ሺݐሻ are random numbers generated in the range between 0 and 1. 3- Segment the image using the optimal number of clusters and the optimal clusters centroids given by the best global particle.
All the parameters settings of PSO image clustering such as acceleration constants and the inertia weight are based on the recommendation by Wong et al. (2011). The common choice of number of particles varies from 20 to 50 (Eberhart & Yuhui, 2001; Poli et al., 2007). 20 particles are used for PSO clustering as smaller number of particles can reduce computation time and 20 particles can provide good clustering performance (Wong et al., 2011). For the inertia weight w, the initial weight value is 0.9 and w decreases linearly with the number of iterations (Yuhui & Eberhart, 1999). The acceleration constants ܿଵ and ܿଶ are both set to 2 (Yuhui & Eberhart, 1999). Next section describes the proposed algorithm that merges SRG with PSO image clustering for automatic seed selection.
3.6.4.2 Automatic SRG Seed Selection of BMRI-SRGPSOC After applying the PSO image FOXVWHULQJ WKH FOXVWHUV¶ LQWHQVLWLHV ZRXOG EH UDQNHG LQ ascending order. Subsequently, the cluster which has the highest intensity would be chosen. Then the centre of the chosen cluster region is selected as the initial seed. The steps taken for this process are shown in Figure 3.16. The algorithm is described in five steps as below:
85
1- Apply PSO Image clustering using the parameters and equations (explained earlier in 3.6.4.1) on the MRI breast image.
2- Rank the PSO clusters according to their intensity values in ascending order.
3- 6HOHFWWKHUHJLRQVZLWKWKHKLJKHVWFOXVWHUV¶LQWHQVLW\YDOXHV DQGHOLPLQDWHWKHRWKHU cluster regions.
4- Find the position ((x, y) coordinates) of the centre pixel of the maximum area in the selected regions.
5- Set the selected position in step 4 as the position of the initial seed.
Next section describes the proposed algorithm that merges of SRG with PSO image clustering for Threshold Value Selection.
86
Start
Apply PSO Image clustering
Rank the PSO clusters according to their intensity values in ascending order
Select the regions with the highest FOXVWHUV¶LQWHQVLW\YDOXHV
Eliminate the other cluster regions
Set the position of the centre pixel of the maximum area in the selected regions as the position of the initial seed
End
Figure 3.16 Block diagram which illustrates the processes of Automatic SRG Seed Selection of BMRI-SRGPSOC
87
3.6.4.3 Automatic SRG Threshold Value Selection of BMRI-SRGPSOC The importance of this process is because of the ranges of the grey level representations for the tumour and the other parts of the breast are not consistent from one image to another. Therefore, the proposed method has the capability of changing the SRG WKUHVKROGYDOXHDFFRUGLQJWRWKHUHVSHFWLYHLPDJH¶VJUH\OHYHOGLVWULEXWLRQ7KHPHWKRG is baseG XSRQ ILQGLQJ WKH RSWLPXP HVWLPDWHG YDOXH IURP WKH 362 FOXVWHUV¶ LQWHQVLWLHV PHDQYDOXHV7KHDYHUDJHIRUFOXVWHUV¶LQWHQVLWLHVH[FHSWWKHKLJKHVWFOXVWHU¶VLQWHQVLW\ (which contains the tumour region) has to be calculated first using Equation (3.17).
݃ݒܣܥൌ
ே ିଵ σୀ ܫܥ ܰ െ ͳ
(3.17)
Where i is the PSO clusters counter, ܰ is the maximum clusters number, CI is the cluster LQWHQVLW\DQG&$YJLVWKHFOXVWHUV¶LQWHQVLW\DYHUDJHThe CAvg is used to examine the optimum SRG threshold value as described in the pseudo code of BMRI-SRGPSOC threshold value selection in Figure 3.17. The evaluation approaches for the methodology phases are explained in the next sections.
88
SET Zero to (IntensitySum), Set DefaultValue (the default value is value between CAvg-255) FOR i=1 to ୡ Do IntensitySum= IntensitySum+CI (i) SRGThreshold= - (IntensitySum/i) IF (SRGThreshold < CAvg) THEN SET SRGThreshold as the chosen threshold value Break ELSE SRGThreshold=DefaultValue ENDIF END FOR
Figure 3.17 Pseudo code of BMRI-SRGPSOC threshold value selection.
3.7
Evaluation Criteria
Several approaches and measures are used in this study in order to evaluate the different processes in the methodology. These processes are skin-line exclusion, MMRT thresholding and both tumour segmentation methods BMRI-SRGPSOC and BMRIMASRG. Some of the processes share the same evaluation measures while other processes used different measures. First, the number of pixels of (ܴ௦ ) and (ܴ௧ ) have to be found, where (ܴ௦ ) represents the segmented region by the proposed approach, while (ܴ௧ ) represents the GT regions segmented by the experts.
89
3.7.1 Skin-line Exclusion Phase Evaluation For this process, two evaluation approaches are used. Both are pixel based. The first evaluation approach used in this study contains five measures; which are True Positive Fraction (TPF) (also called Sensitivity), False Negative Fraction (FNF), False Positive Fraction (FPF), True Negative Fraction (TNF) (also called Specificity) and Sum of True Volume Fraction (STVF). The calculations are made using the Equations (3.18-3.22). (Chalana & Kim, 1997; Fenster & Chiu, 2005; Metz, 1986; McNeil & Hanley, 1984). The concept of TPF, TNF, FPF and FNF is showing in Figure 3.18.
Figure 3.18 Diagram showing the definitions of TPF, TNF, FPF and FNF in the evaluation of segmentation results. The segmented object (larger, rotated square) is compared with a reference ground truth object (smaller square) (Berry, 2007).
90
ܶܲܨሺሻ ൌ
ܴ௦ ܴ ת௧ ܴ௧
(3.18)
ൌ
୲ െ ୱ ୲
(3.19)
ܨܲܨൌ
ܴ௦ െ ܴ௧ ܴ௧
(3.20)
ܶܰܨሺ
ሻ ൌ ͳ െ
ܴ௦ െ ܴ௧ ܴ௧
ܸܵܶ ܨൌ ܶܲ ܨ ܶܰܨ
(3.21)
(3.22)
As the basis of measurements: for TPF, TNF and STVF measures, whenever the results are higher, the performance is better. Meanwhile for the rest of the basis of measurements which are FNF and FPF, the lower results indicate better performance. The Sensitivity denotes the proportion of positives which the proposed method is correctly identified. While the Specificity denotes the proportion of negatives which the proposed method is correctly identified.
The second approach used contains measures, i.e. Jaccard (also called Relative Overlap), Misclassification Rate (MCR) and Dice which have been used before for brain segmentation (Song et al., 2006) and in breast segmentation (Azmi et al., 2011a; (UWDú et al., 2008). The Relative Overlap is calculated as in Equation (3.23).
ൌ
ୱ ୲ ת ୱ ୲
(3.23)
91
The Misclassification Rate (MCR) is calculated as in Equation (3.24) ൌ ͳ െ
ୱ ୲ ת ୲
(3.24)
And Dice measure is calculated as in Equation (3.25)
ൌ
ʹȁ ୱ ୲ תȁ ȁ ୱ ȁ ȁ ୲ ȁ
(3.25)
The higher values of Jaccard and Dice indicate better results.
3.7.2 Image Thresholding Phase Evaluation Two evaluation approaches are used in this work to evaluate the thresholding performance; pixel based evaluation approach which has been used before for evaluating image thresholding (Prasad et al., 2011; Rosin & Ioannidis, 2003)and quality evaluation approach. Pixel based evaluation approach is used to evaluate the accuracy of the proposed method by comparing between pixels of thresholded image (ܴ௦ ) and the GT (ܴ௧ ). Two measures are used in this work to build the comparison i.e. the Jaccard measure and Dice measure. The formulas of the two measures are given in Equations (3.23) and (3.25).
The second evaluation approach; i.e. the quality evaluation approach (Huynh-Thu & Ghanbari, 2008) contains two measures; Mean Square Error (MSE) and the Peak Signal to Noise Ratio (PSNR). MSE measures the cumulative squared error between the
92
thresholded image and the GT image, while PSNR measures the peak error between the two images. MSE and PSNR are calculated using Equations (3.26) and (3.27) (Huynh-Thu & Ghanbari, 2008).
¦ >I r, c I r, c @
2
MSE
R ,C
t
g
RuC
(3.26)
where, ܫ௧ is the thresholded image and ܫ is GT image, R is the number of rows; C is the number of columns in the images.
§ Maxp 2 PSNR 10 log 10 ¨¨ © MSE
· ¸¸ ¹
(3.27)
Where Maxp is the maximum possible pixel value.
The higher values obtained from the PSNR calculated gives a better quality thresholded image whereas a lower MSE value indicates lower error. Typical values for the PSNR are between 60 and 80 (Yun-Fei et al., 2012).
3.7.3 Tumour Segmentation Phase Evaluation To evaluate breast tumour segmentation accuracy, the two pixels based approaches used in (section 3.7.1) are used here too, where the formulation of the first approach including; TPF (Sensitivity), FNF, FPF, TNF (Specificity) and STVF are given in Equations (3.18-
93
3.22). On the other hand, the second approach formulas include the Jaccard and Dice as given in Equations (3.23) and (3.25).
To evaluate the automatic seed selection, the positions of the seed pixel that have been VHOHFWHG E\ WKH SURSRVHG DSSURDFK DUH FRPSDUHG ZLWK WKH UHIHUHQFH¶V SRVLWLRQ 7KH UHIHUHQFH¶VSRVLWLRQLVPDQXDOO\VHOHFWed according to the GT. The position accuracy is calculated as given in Equation (3.28)
ܲ ݕܿܽݎݑܿܿܽ݊݅ݐ݅ݏൌ ͳͲͲ െ
ȁି ȁ ൈ ͳͲͲΨ
(3.28)
Where is the (x, y) coordinates of the manual selected seed pixel, is the (x, y) coordinates of the automatic selected seed pixel.
3.8 Summary In this chapter, the proposed methodology of computer aided segmentation approach for breast MRI tumour is explained. Number of image processing methods are presented to achieve the tumour segmentation goal.
The methodology starts with data acquisition followed by pre-processing phase which includes image splitting and filtering using median filter. An integration method is proposed to be used for excluding the skin-line regions from MRI breast images, which is a necessary process because of the similarity of the intensity level of the skin-line and tumour. The proposed method integrates LSAC algorithm and morphological thinning
94
algorithm. Then, image thresholding phase is discussed using proposed mean maximum raw thresholding method.
Tumour segmentation phase is divided into two new methods that automate region growing seed, both proposed methods are presented to automatically select the nessesary variable factors (initial seed and threshold value) of SRG. The methods are breast MRI tumour using modified automatic SRG (BMRI-MASRG), the second method is integrated SRG-PSO image clustering (BMRI-SRGPSOC). This chapter ends with describing the different evaluation approaches that are used to obtain the results and support the proposed methodology. Next chapter focuses on describing the results obtained after applying the proposed methodology on the dataset. The results are then evaluated and analyzed based on GT of the dataset.
95
CHAPTER 4 RESULTS AND DISCUSSION
4.1 Introduction In the previous chapter, the methods proposed for this study were explained. Therefore, in this chapter, the results of the proposed techniques and methods of the approach are presented and discussed. The results of each process are obtained by applying suitable evaluation approach on the same dataset samples. Then statistical analysis is applied to find the significance of the proposed methods in this work. Different evaluation approaches and measures are used here based on equations in section 3.7.
For breast skin-line exclusion method which has been discussed in section 3.4, the results presented here are based on two approaches of pixel evaluation. First approach used measures such as; TPF, FNF, FPF, TNF and STVF, and the second approach used measures such as; Dice, Jaccard and MCR.
For MMRT thresholding method which has been presented in section 3.5, results of this method are obtained by applying pixel based (Jaccard and Dice measures) and quality based evaluations (PSNR and MSE measures).
In the tumour segmentation section, the proposed methods; BMRI-MASRG and BMRISRGPSOC which were presented in sections 3.6.3 and 3.6.4, are tested using the same
96
evaluation approaches that have been used to test the results of skin-line exclusion method. In addition, the accuracy of SRG seed selection position is tested.
The evaluation results are analysed after that by applying Analysis of Variance test (ANOVA) and Receiver Operating Characteristic (ROC). ANOVA is applied when the method is compared with previous methods. ANOVA is a statistical comparison test that indicates the degree of significance between groups of different results when it tests on the same dataset. ANOVA is applied by finding the level of statistical significance (p). The difference between the groups are considered as significance if (p < 0.05) and if (F statistic > F critical).
For a more statistically evaluation, ROC curve analysis is used to illustrate the true positive fraction rate compared with the false positive fraction rate for the segmented images of proposed methods. The high Area under the Curve (AUC) indicates improved segmentation performance, where higher value of TPF is achieved at each value of FPF.
4.2 Results of Breast Skin-Line Exclusion Integration method of LSAC algorithm and morphological thinning algorithm is evaluated in two stages. Two images sets are prepared manually for the purpose of the evaluation; the first set is to evaluate the segmentation process while the second set is to evaluate the removal process.
97
In the breast skin-line segmentation stage, the chosen parameters for LSAC algorithm are ı =1.5 and ܰௌ =700 while ்ܰ =7 is the chosen parameter for the thinning algorithm for the removal stage. The parameters have been selected using the trial and error method (section 3.4).
Figure 4.1 shows samples of malignant RIDER images and the resultant images of breast skin-line segmentation and removal processes. On the other hand, Figure 4.2 shows sample of benign RIDER images and the resultant images of the same processes.
The proposed approach is applied on the 40 RIDER dataset images. Then, two of the pixel based evaluation approaches are used to evaluate the performance of skin-line exclusion; these approaches are explained in section 3.7.1. The calculations are made by comparing the segmented images with the GT using the measures of first and second approach. The measures of the first approach are TPF, FNF, FPF, TNF and STVF for (Equations (3.18), (3.19), (3.20), (3.21) and (3.22)), whereby, the measures of the second approach are Dice, Jaccard and MCR (Equations (3.23), (3.24) and (3.25)).
ROC curve analysis is applied on both stages of the method, i.e. skin-line border segmentation and removal in order to measure the significance improvement statistically.
98
(a)
(b)
(c)
(d)
(e)
(f)
(g)
(h)
(i)
Figure 4.1 Breast skin-line exclusion processes on three malignant RIDER images; (a, b, and c) original images, (d, e, and f) after applying LSAC algorithm with ı and ܰௌ =700, (g, h and i) after applying the thinning algorithm with ்ܰ =7. 99
(a)
(b)
(c)
(d)
(e)
(f)
Figure 4.2 Breast skin-line removal exclusion on two benign RIDER images; (a and b) original images, (c and d) after applying LSAC algorithm with ı DQGܰௌ =700, (e and f) after applying the thinning algorithm with ்ܰ =7. 100
4.2.1 Results of Breast Skin-Line Border Segmentation Stage The summary results of first evaluation approach for the segmentation stage are tabled as in Table 4.1, while the summary results of second evaluation approach are tabled in Table 4.2.
Table 4.1 Summary results of skin-line segmentation for RIDER MRI breast images using evaluation measures (TPF, FNF, FPF, TNF and STVF).
TPF
FNF
FPF
TNF
STVF
Mean
0.8128
0.0926
0.0231
0.9401
1.7530
Stddev
0.0874
0.0406
0.0087
0.0695
0.1413
Variance
0.0076
0.0017
0.0001
0.0048
0.0200
Min
0.7192
0.0917
0.0182
0.9083
1.6275
Max
0.9065
0.0935
0.0280
0.9720
1.8785
Table 4.2 Summary results of skin-line segmentation for RIDER MRI breast images using evaluation measures (Jaccard, MCR and Dice).
Jaccard
MCR
Dice
Mean
0.7708
0.1367
0.8404
Stddev
0.0839
0.0424
0.1039
Variance
0.0070
0.0018
0.0108
Min
0.6597
0.0935
0.7436
Max
0.8818
0.1800
0.9372
101
From Table 4.1 of the first evaluation approach, the means of results were high with TPF = 0.8128, FNF = 0.0926, FPF = 0.0231, TNF = 0.9401 and STVF = 1.7530. It can be noticed that the high sensitivity and specificity of proposed approach indicate high performance in the segmentation stage.
For the second evaluation approach, the mean results are Jaccard = 0.7708, MCR = 0.1367 and Dice = 0.8404. The high Dice and Jaccard results indicate high performance while low result of MCR indicates low error rate in the segmentation.
ROC is applied to illustrate by drawing curve the True Positive Fraction compared with the False Positive Fraction. Where, AUC for the skin-line segmentation stage is 0.9902. The ROC for segmentation processes is shown in Figure 4.3.
Figure 4.3
The ROC curve for MRI breast skin-line segmentation.
102
4.2.2 Results of Breast Skin-Line Removal Stage The summary results of the two evaluation approaches of the skin-line removal stage are tabled in Table 4.3 and Table 4.4.
Table 4.3 Summary results of skin-line removal for RIDER MRI breast images using evaluation measures (TPF, FNF, FPF, TNF and STVF).
TPF
FNF
FPF
TNF
STVF
Mean
0.8583
0.0062
0.0484
0.9708
1.8291
Stddev
0.0950
0.0364
0.0122
0.0229
0.1027
Variance
0.0090
0.0013
0.0001
0.0005
0.0105
Min
0.7168
0.0003
0.0465
0.9535
1.7049
Max
0.9997
0.0120
0.0503
0.9880
1.9532
Table 4.4 Summary results of skin-line removal for RIDER MRI breast images using evaluation measures (Jaccard, MCR and Dice). Jaccard
MCR
Dice
Mean
0.8740
0.1418
0.9321
Stddev
0.0788
0.0950
0.0922
Variance
0.0062
0.0090
0.0085
Min
0.7927
0.0003
0.8871
Max
0.9553
0.2832
0.9771
103
From Table 4.3, it can be noticed that the proposed approach scored good results using the various measures. For the first evaluation approach, the means of results were high with TPF = 0.8583, FNF = 0.0062, FPF = 0.0484, TNF = 0.9708 and STVF = 1.8291. From the results, it noticed that the high sensitivity and specificity of proposed approach indicate high performance in the removal stage.
For the second evaluation approach, the means results are Jaccard = 0.8740, MCR = 0.1418 and Dice = 0.9321. The high Dice and Jaccard results indicate high removal performance while low result of MCR indicates low error rate in the removal stage.
The applying of ROC curve for breast-skin removal process is shown in Figure 4.4 AUC for this process is 0.9507.
Figure 4.4
The ROC curve for MRI breast skin-line removal.
104
The evaluation results for both stages of the method show that the performance is significantly high compared with GT. The statistical ROC and AUC analysis show a good performance too. Hence, the high AUC indicates better segmentation performance. However, skin-line exclusion has not been applied by previous approaches on the same particular breast MRI images dataset i.e. RIDER dataset, therefore the proposed method is not compared to other published work.
4.3 Results of Image Thresholding Using MMRT MMRT is tested on the RIDER dataset. The evaluation criteria explained earlier in section 3.7.2 is applied to evaluate the performance of the proposed thresholding method in comparison with dataset GT. The summary results of the pixel based evaluation are tabled in Table 4.5. This approach includes Jaccard measure (Equation (3.23)) and Dice measure (Equation (3.25)). On the other hand, quality evaluation approach includes MSE and PSNR which are calculated using Equations (3.26) and (3.27) are tabled in Table 4.6.
Table 4.5 Summary results of the pixel based evaluation approach (Jaccard and Dice measures) for MMRT.
Jaccard
Dice
Mean
0.6429
0.6828
Stddev
0.1719
0.1546
Variance
0.0296
0.0239
Min
0.3073
0.9306
Max
0.5550
0.9640
105
Table 4.6 Summary results of the quality evaluation approach (PSNR and MSE measures) for MMRT.
PSNR
MSE
Mean
69.97
0.01
Stddev
2.79
0.005
Variance
7.81
0.00
Min
65.13
74.10
Max
0.002
0.02
In order to evaluate the performance of the proposed method, five standard thresholding methods were applied on the same 40 test images of RIDER dataset. The approaches are Iterative Thresholding (Ridler & Calvard, 1978; Anjos et al., 2010), Grey level Histogram (Otsu, 1979), Entropy based (Kapur et al., 1985; Prink & Pendock, 1996; Li & Lee, 1993; Alayli & El-Zaart, 2013), Fuzzy Thresholding (Huang & Wang, 1995; Li et al., 2012; Arifin et al., 2009) and Multilevel Thresholding (Yen et al., 1995; Arora et al., 2008; Liao et al., 2001; Yan et al., 2005). The five methods are described in Chapter 2 (2.6.1.1)
Figure 4.5 illustrates the results of applying the thresholding as in the proposed method in comparison with the five standard methods on one of the malignant test images of the RIDER dataset. On the other hand, the application of the thresholding methods on a benign test image is shown in Figure 4.6.
106
(a)
(b)
(c)
(d)
(e)
(f)
Figure 4.5 Results of applying MMRT and the standard thresholding methods on malignant test image; (a) Iterative Thresholding. (b) Grey level Histogram. (c) Entropy based. (d) Fuzzy Thresholding. (e) Multilevel Thresholding. (f) MMRT.
107
(a)
(b)
(c)
(d)
(e)
(f)
Figure 4.6 Results of applying MMRT and the standard thresholding methods on benign test image: (a) Iterative Thresholding. (b) Grey level Histogram. (c) Entropy based. (d) Fuzzy Thresholding. (e) Multilevel Thresholding. (f) MMRT
108
Visual observations of Figures 4.5 and 4.6 show that the proposed thresholding method has detected more accurate suspected regions than the other methods. The proposed method and the standard methods are applied on all of the RIDER dataset images. Then, calculations are made, with the process of comparing these results with the GT of the dataset using the Jaccard and Dice measures. The results are then tabled as in Table 4.7. Subsequently, the calculated results of the PSNR and MSE measures are compared with the GT. These are then listed as in Table 4.8.
109
Table 4.7 Results of evaluating the Jaccard and Dice measures for thresholding using the proposed method and other methods (Iterative Thresholding, Grey level Histogram, Entropy based, Fuzzy Thresholding and Multilevel Thresholding). The method Iterative Thresholding
Grey level Histogram
Entropy Based
Fuzzy Thresholding
Multilevel Thresholding
MMRT
Statistic Mean Stddev Variance Min Max Mean Stddev Variance Min Max Mean Stddev Variance Min Max Mean Stddev Variance Min Max Mean Stddev Variance Min Max Mean Stddev Variance Min Max
Jaccard 0.401 0.173 0.03 0.124 0.579 0.399 0.166 0.028 0.117 0.586 0.551 0.225 0.051 0.140 0.877 0.319 0.167 0.028 0.065 0.495 0.558 0.205 0.042 0.140 0.878 0.643 0.172 0.03 0.307 0.931
110
Dice 0.444 0.123 0.015 0.221 0.598 0.452 0.124 0.015 0.209 0.589 0.620 0.200 0.04 0.245 0.935 0.379 0.148 0.022 0.122 0.541 0.616 0.212 0.045 0.245 0.935 0.683 0.155 0.024 0.555 0.964
Table 4.8 Results of evaluating the PSNR and MSE measures using the thresholding of the proposed method and other methods (Iterative Thresholding, Grey level Histogram, Entropy based, Fuzzy Thresholding and Multilevel Thresholding). The method Iterative Thresholding
Grey level Histogram
Entropy Based
Fuzzy Thresholding
Multilevel Thresholding
MMRT
Statistic Mean Stddev Variance Min Max Mean Stddev Variance Min Max Mean Stddev Variance Min Max Mean Stddev Variance Min Max Mean Stddev Variance Min Max Mean Stddev Variance Min Max
PSNR 56.30 1.04 1.07 54.55 58.89 56.81 1.27 1.62 54.69 59.34 66.38 2.03 4.11 62.97 69.78 53.84 0.88 0.78 52.57 55.38 64.56 2.67 7.12 59.47 69.75 69.97 2.79 7.81 65.13 74.10
111
MSE 0.070 0.067 0.005 0.011 0.221 0.064 0.073 0.005 0.009 0.221 0.0198 0.008 0.001 0.007 0.033 0.080 0.006 0.001 0.071 0.090 0.035 0.020 0.001 0.007 0.074 0.011 0.005 0.001 0.003 0.020
ANOVA test has been applied to make statistical comparison between the proposed 0057¶VUHVXOWVDQGWKHVWDQGDUGPHWKRGV¶UHVXOWV RQWKHVDPHGDWDVHW E\ ILQGLQJWKH level of statistical significance (p). The difference between the groups are considered of significance if (p < 0.05) and if (F statistic > F critical), where F critical is 2.253 (Moore & McCabe, 2003; Casella, 2008). The summary of ANOVA tests using the results of Jaccard, Dice, PSNR and MSE measures of MMRT compared with the standard methods are stated in Table 4.9.
Table 4.9 Summary of the $129$ WHVWV DQDO\VLV IRU WKH SURSRVHG DSSURDFK¶V UHVXOWV compared with the results of the other approaches (Iterative Thresholding, Grey level Histogram, Entropy based, Fuzzy Thresholding and Multilevel Thresholding). Jaccard
Dice
PSNR
MSE
F statistic
17.66
22.25
453.21
19.19
P value
0.781 E-14
0.312 E-17
0.36 E-117
0.547 E-15
From Table 4.9, it can be seen that there are a statistically significant differences between methods (p < 0.05) and (F statistic < F critical (2.253)) as determined by one-way ANOVA for Jaccard (F = 17.66, p = 0.78 E-14), Dice (F = 22.25, p = 0.31 E-17), PSNR (F = 453.21, p = 0.36 E-117) and MSE (F = 19.19, p = 0.55 E-15). The statistical ANOVA graphs for the average and the variance of the evaluation measures are shown in Figures 4.7, 4.8, 4.9 and 4.10.
112
Ϭ͘ϳ Ϭ͘ϲ
:ĂĐĐĂƌĚ
Ϭ͘ϱ Ϭ͘ϰ Ϭ͘ϯ Ϭ͘Ϯ
ǀĞƌĂŐĞ
Ϭ͘ϭ
sĂƌŝĂŶĐĞ
Ϭ
dŚĞŵĞƚŚŽĚƐ
Figure 4.7 Statistical ANOVA graphs for MMRT in comparison with the standard methods using results of Jaccard measure.
Ϭ͘ϴ
Ϭ͘ϳ Ϭ͘ϲ
ŝĐĞ
Ϭ͘ϱ Ϭ͘ϰ Ϭ͘ϯ Ϭ͘Ϯ
ǀĞƌĂŐĞ
Ϭ͘ϭ
sĂƌŝĂŶĐĞ
Ϭ
dŚĞŵĞƚŚŽĚƐ
Figure 4.8 Statistical ANOVA graphs for MMRT in comparison with the standard methods using results of Dice measure.
113
Figure 4.9 Statistical ANOVA graphs for MMRT in comparison with the standard methods using results of PSNR measure.
Ϭ͘Ϭϵ Ϭ͘Ϭϴ DĞĂŶ^ƋƵĂƌĞ ƌƌŽƌ
Ϭ͘Ϭϳ Ϭ͘Ϭϲ Ϭ͘Ϭϱ Ϭ͘Ϭϰ Ϭ͘Ϭϯ Ϭ͘ϬϮ
ǀĞƌĂŐĞ
Ϭ͘Ϭϭ
sĂƌŝĂŶĐĞ
Ϭ
dŚĞŵĞƚŚŽĚƐ
Figure 4.10 Statistical ANOVA graphs for MMRT in comparison with the standard methods using results of MSE measure.
114
The pixel based evaluation approach demonstrates improved results when compared with the other methods as stated in Table 4.7. The higher values of Jaccard and Dice measures represent better results. For the Jaccard measure results, the mean of the proposed method is 0.643 which is superior when compared with the results of the standard methods; Iterative Thresholding = 0.401, Grey level Histogram = 0.399, Entropy based = 0.551, Fuzzy Thresholding = 0.319 and Multilevel Thresholding = 0.558. For the Dice measure results, the mean of the proposed method is 0.683, which is greater than the results of; Iterative Thresholding = 0.444, Grey level Histogram = 0.452, Entropy based = 0.620, Fuzzy Thresholding = 0.379 and Multilevel Thresholding = 0.614.
The quality evaluation approach shows better results when compared with the other methods as stated in Table 4.8. The higher value of PSNR measure is the better quality of the thresholded image whereas the lower value of MSE measure is the lower error. For the results of the PSNR measures, the mean of the proposed method is 69.97 while the results of the standard methods are lower; Iterative Thresholding = 56.30, Grey level Histogram = 56.81, Entropy based = 66.38, Fuzzy Thresholding = 53.84 and Multilevel Thresholding =64.56. For the MSE measure results, the mean of the proposed method is 0.011 which is also superior to the results of the standard methods; Iterative Thresholding = 0.070, Grey level Histogram = 0.064, Entropy based = 0.020, Fuzzy Thresholding = 0.080 and Multilevel Thresholding = 0.035.
Moreover, the statistical analysis of ANOVA shows there are statistically significant differences between the methods where p values are less than 0.05. The F statistical values are greater than the F critical value which is 2.253. The p values and F statistic
115
listed in Table 4.9 for the different measures are Jaccard (F = 17.66, p = 0.78 E-14), Dice (F = 22.25, p = 0.31 E-17), PSNR (F = 453.21, p = 0.36 E-117) and MSE (F = 19.19, p = 0.55 E-15). The statistical ANOVA graphs for the results of Jaccard, Dice, PSNR and MSE are shown in Figures 4.7, 4.8, 4.9 and 4.10 respectively. From the graphs, it has proven that the performance of the proposed method is higher than the standard methods when applied on this type of images.
4.4 Results of Tumour Segmentation Phase In this section, the evaluation results of BMRI-MASRG and BMRI-SRGPSOC methods are presented in sub-sections using three evaluation approaches, first approach includes five measures, i.e. TPF, FNF, FPF, TNF and STVF, using the formulas are given in Equations (3.18), (3.19), (3.20), (3.21) and (3.22) respectively. Second approach includes two measures, i.e. Jaccard and MCR. Third approach is the accuracy of selected initial seed automatically compared to the manual selection; the calculations have been made using Equation (3.28).
The results of both proposed methods are compared to the results of previous methods, i.e. KNN, SVM, Bayesian, FCM and IMPST. For statistical analysis ANOVA and ROC are applied to measure the significance of the methods.
4.4.1 Results of Modified Automatic Seeded Region Growing (BMRI-MASRG) The methodology explained earlier in Section 3.6.3 is applied and tested on the RIDER Breast MRI dataset. Figures 4.11 and 4.12 show the different processes steps of BMRI-
116
MASRG method on malignant and benign samples images respectively. The two evaluation methodologies explained in the evaluation criteria in Section 3.7.3 are then applied. The results are tabled in Tables 4.10, 4.11 and 4.12. Table 4.10 illustrates the results of the first evaluation methodology (TPF, FNF, FPF, TNF and STVF).
Table 4.10 Evaluation results of BMRI-MASRG using TPF (Sensitivity), FNF, FPF, TNF (Specificity) and STVF. TPF
FNF
FPF
TNF
STVF
Mean
0.8233
0.1767
0.0976
0.9023
1.7256
Stddev
0.0965
0.0965
0.0970
0.0969
0.1040
Variance
0.0093
0.0093
0.0094
0.0094
0.0108
Max
0.9944
0.4779
0.3523
0.9938
1.8669
Min
0.5221
0.0045
0.0001
0.6477
1.4632
117
(a)
(b)
(c)
(d)
(e)
(f)
Figure 4.11 The proposed (BMRI-MASRG) approach processes on one of malignant RIDER image: (a) after breast skin detection and deletion; (b) after applying the thresholding procces using MMRT algorithm; (c) after applying Morphological Open Operation; (d) initial seed selected using the proposed method; (e) after the SRG method is applied using the proposed method of SRG threshold value; (f) The segmented tumour regions.
118
(a)
(b)
(c)
(d)
(e)
(f)
Figure 4.12 The proposed (BMRI-MASRG) approach processes on one of benign RIDER image: (a) after breast skin detection and deletion; (b) after applying the thresholding procces using MMRT algorithm; (c) after applying Morphological Open Operation; (d) initial seed selected using the proposed method; (e) after the SRG method is applied using the proposed method of SRG threshold value; (f) The segmented tumour regions.
119
Table 4.11 shows the results of the second evaluation methodology (Jaccard and MCR), where the mean of Jaccard for all images is high, which is about 0.7524 and the mean of MCR is low, which is about 0.1767.
Table 4.11
Evaluation results of BMRI-MASRG using Relative Overlap (Jaccard) and MCR Jaccard
MCR
Mean
0.7524
0.1767
Stddev
0.0848
0.0965
Variance
0.0072
0.0093
Max
0.8754
0.4779
Min
0.4931
0.0045
7DEOH VKRZV WKH DXWRPDWLFDOO\ VHOHFWHG LQLWLDO VHHG SL[HO¶V coordinates when FRPSDUHGZLWK WKHPDQXDOO\VHOHFWHGSL[HO¶VFRRUGLQDWHV7KHPDQXDOO\VHOHFWLRQVDUH made according to the position of the segmented tumours in the GT. The calculations in Table 4.12 are made using Equation (3.28).
Table 4.12 Evaluation results of BMRI-MASRG of automatically selected initial VHHGSL[HO¶VFRRUGLQDWHVFRPSDUHGZLWKWKHPDQXDOO\VHOHFWHGSL[HO¶VFRRUGLQDWHV X accuracy %
Y accuracy %
Mean
97.38
97.08
Stddev
1.550
2.890
Variance
2.403
8.352
Max
98.92
99.43
Min
90.80
80.83
120
Results obtained in Table 4.12 indicate that the coordinates of automatic selection for initial seed are precise compared to the coordinates of manual selection, with average of 97.38 % for X coordinate accuracy and average of 97.08 % for X coordinate accuracy.
4.4.2 Results of Integrated Method of SRG and PSO Image Clustering The same approaches which are used to evaluate BMRI-MASRG method (in section 4.4.1) are used here to evaluate BMRI-SRGPSOC method. Figures 4.13 and 4.14 show the different processes steps of BMRI-SRGPSOC method on malignant and benign samples images, respectively.
Table 4.13 shows the summary results of first set of evaluation measures (TPF, FNF, FPF, TNF and STVF). The results are evaluated after the proposed method is applied on the RIDER dataset images.
Table 4.13
Evaluation results of BMRI-SRGPSOC using TPF (Sensitivity), FNF, FPF, TNF (Specificity) and STVF.
TPF
FNF
FPF
TNF
STVF
Mean
0.8345
0.1752
0.0931
0.9069
1.7126
Stddev
0.0900
0.0876
0.0860
0.0860
0.1939
Variance
0.0081
0.0077
0.0074
0.0074
0.0376
Max
0.9821
0.5075
0.5316
1.0000
1.8987
Min
0.4925
0.0179
0.0000
0.4684
0.7262
121
(a)
(b)
(c)
(d)
(e)
(f)
Figure 4.13 The proposed (BMRI-SRGPSOC) approach processes on one of benign RIDER image (a) The resultant image after applying the PSO image clustering. (b) The KLJKHVW362FOXVWHUV¶LQWHQVLW\UHJLRQVDIWHURWKHUUHJLRQVDUHHOLPLQDWHG F Maximum area of the selected PSO regions. (d) Initial seed selected automatically (marked in red) as the centre of the region selected in (c). (e) SRG using the automatic threshold value is applied (marked in blue). (f) The segmented tumour area.
122
(a)
(b)
(c)
(d)
(e)
(f)
Figure 4.14 The proposed (BMRI-SRGPSOC) approach processes on one of malignant RIDER image (a) The resultant image after applying the PSO image FOXVWHULQJ E 7KHKLJKHVW362FOXVWHUV¶LQWHQVLW\UHJLRQVDIWHURWKHUUHJLRQVDUH eliminated. (c) Maximum area of the selected PSO regions. (d) Initial seed selected automatically (marked in red) as the centre of the region selected in (c). (e) SRG using the automatic threshold value is applied (marked in blue). (f) The segmented tumour area.
123
Table 4.14 shows BMRI-SRGPSOC results of the second set of evaluation measures. The mean of Jaccard = 0.7693 which is higher than Jaccard of BMRI-MASRG, the mean of MCR= 0.1655 indicates that the error rate of BMRI-SRGPSOC is lower than BMRIMASRG. 7DEOHVKRZVWKHDXWRPDWLFDOO\VHOHFWHGLQLWLDOVHHGSL[HO¶VFRRUGLQDWHVRI BMRI-SRGPSOC compared to the manually VHOHFWHGSL[HO¶VFRRUGLQDWHV
Table 4.14
Evaluation results of BMRI-SRGPSOC using Relative Overlap (Jaccard) and MCR Jaccard
MCR
Mean
0.7693
0.1655
Stddev
0.1053
0.0900
Variance
0.0111
0.0081
Max
0.8832
0.5075
Min
0.4118
0.0179
Table 4.15 Evaluation results of BMRI-SRGPSOC of automatically selected initial VHHGSL[HO¶VFRRUGLQDWHVFRPSDUHGZLWKWKHPDQXDOO\VHOHFWHGSL[HO¶VFRRUGLQDWHV X accuracy %
Y accuracy %
Mean
96.89
97.11
Stddev
2.68
3.15
Variance
7.18
9.92
Max
98.90
99.43
Min
87.45
78.87
124
From the results obtained as in Table 4.15, it can be noticed that the coordinates of automatic selection for initial seed are precise compared to the coordinates of manual selection, with average of 96.89 % for X coordinate accuracy and average of 97.11 % for X coordinate accuracy. It also can be noticed from comparing the results of seed selection accuracy of both methods BMRI-SRGPSOC and BMRI-MASRG that the difference margin is less than 1%.
4.4.3 Comparison of Proposed Segmentation Approaches (BMRI-MASRG and BMRI-SRGPSOC) and Other Approaches The two proposed methods of tumour segmentation are tested on the same dataset to evaluate which method is more accurate. Figure 4.15 shows the results of final segmented tumour of five RIDER samples (malignant and benign) after applying the two methods, i.e. BMRI-MASRG and BMRI-SRGPSOC. The figure shows also GT of dataset. From the figure, it can be noticed that both methods managed to segment the right regions of tumours compared to the GT of the dataset. The differences of the segmented shapes and the number of voxels are then calculated using the evaluation measures as explained earlier.
The results of the proposed approaches are also compared with the results of the previous ZRUNV LQYROYLQJ ILYH GLIIHUHQW VHJPHQWDWLRQ DSSURDFKHV 7KH SUHYLRXV ZRUN¶V UHVXOWV have been stated in the comparison study of (Azmi et al., 2011a). The approaches are KNN, SVM, Bayesian, FCM and IMPST. The tested data are the same dataset which is (Breast MRI RIDER dataset). The measures used for the evaluation are Relative Overlap
125
(Jaccard), MCR, TPF (Sensitivity), TNF (^ƉĞĐŝĨŝĐŝƚLJ) and STVF. The results of the five approaches and the proposed approaches are stated in Table 4.16.
Figure 4.16 shows the results of applying the proposed approaches (BMRI-MASRG and BMRI-SRGPSOC) in comparison with the five previous approaches on one of the malignant test images of the RIDER dataset. On the other hand, the application of tumour segmentation approaches on a benign test image is shown in Figure 4.17.
126
left side
median filter 3 x 3
left side
BMRI-SRGPSOC
BMRI-MASRG
GT
Original
left side
Figure 4.15 Comparison of segmented tumour using proposed approaches (BMRIMASRG and BMRI-SRGPSOC) by testing five RIDER images with their GT.
127
(a)
(b)
(c)
(d)
(e)
(f)
(g)
(h)
(i)
Figure 4.16 Results of applying BMRI-MASRG and BMRI-SRGPSOC in comparison to previous approaches malignant test image: (a) Original image. (b) GT. (c) KNN. (d) SVM. (e) Bayesian. (f) FCM. (g) IMPST. (h) BMRI-MASRG. (i) BMRI-SRGPSOC
128
(a)
(c)
(d)
(b)
(e)
(f)
(g)
(h)
(i)
Figure 4.17 Results of applying BMRI-MASRG and BMRI-SRGPSOC in comparison to previous approaches benign test image: (a) Original image. (b) GT. (c) KNN. (d) SVM. (e) Bayesian. (f) FCM. (g) IMPST. (h) BMRI-MASRG. (i) BMRI-SRGPSOC
129
Table 4.16 Segmentation results for the proposed approaches (BMRI-MASRG and BMRI-SRGPSOC) and other approaches (KNN, SVM, Bayesian, FCM and IMPST). The method
KNN
SVM
Bayesian
FCM
IMPST
BMRI-MASRG
BMRI-SRGPSOC
Statistic Mean Stddev Variance Max Min Mean Stddev Variance Max Min Mean Stddev Variance Max Min Mean Stddev Variance Max Min Mean Stddev Variance Max Min Mean Stddev Variance Max Min Mean Stddev Variance Max Min
TPF 0.73 0.13 0.02 0.87 0.42 0.75 0.19 0.04 0.94 0.30 0.76 0.19 0.04 0.95 0.30 0.82 0.27 0.07 0.99 0.00 0.79 0.19 0.04 0.96 0.38 0.82 0.10 0.01 0.99 0.52 0.84 0.09 0.01 0.98 0.49
130
TNF 0.75 0.15 0.02 0.97 0.99 0.71 0.21 0.04 0.99 0.17 0.59 0.25 0.06 0.97 0.09 0.42 0.55 0.30 0.96 -0.73 0.83 0.15 0.02 0.99 0.58 0.90 0.10 0.01 0.99 0.65 0.91 0.09 0.01 1.00 0.47
STVF Jaccard MCR 1.49 0.60 0.27 0.19 0.12 0.10 0.04 0.01 0.01 1.75 0.76 0.58 1.85 0.36 0.13 1.46 0.60 0.25 0.31 0.19 0.12 0.10 0.04 0.01 1.85 0.86 0.70 0.67 0.26 0.06 1.34 0.56 0.24 0.35 0.17 0.08 0.12 0.03 0.01 1.76 0.77 0.70 0.69 0.22 0.05 1.24 0.59 0.18 0.71 0.25 0.08 0.50 0.06 0.01 1.83 0.84 1.00 -0.19 0.00 0.01 1.61 0.68 0.21 0.21 0.17 0.12 0.04 0.03 0.01 1.88 0.89 0.62 1.18 0.34 0.04 1.73 0.75 0.18 0.10 0.09 0.10 0.01 0.01 0.01 1.87 0.88 0.48 1.46 0.49 0.01 1.71 0.77 0.17 0.19 0.11 0.09 0.04 0.01 0.01 1.90 0.88 0.51 0.73 0.41 0.02
The results of the evaluation of both methods showed improved performance. In general Table 4.16 shows improved results of BMRI-MASRG and BMRI-SRGPSOC when compared with the other approaches. The Jaccard mean of the proposed approaches (BMRI-MASRG = 0.75 and BMRI-SRGPSOC = 0.77) are higher than the Jaccard mean of IMPST (the best Jaccard mean from the previous approaches) which is 0.68. For MCR results, BMRI-MASRG = 0.18 which is equal to the MCR of FCM and better than the rest of previous methods, which are KNN, SVM, Bayesian and IMPST. MCR results of BMRI-SRGPSOC = 0.17 is the best among the tested methods including BMRIMASRG. Low MCR indicates lower error segmentation rate. The TNF mean of the proposed approaches (BMRI-MASRG = 0.90 and BMRISRGPSOC = 0.91) are higher than the TNF mean of IMPST (the best TNF mean of previous approaches) having a value of 0.83. However, the TPF mean of the proposed approach is 0.82, which is almost equal the TPF mean of FCM whose value is 0.82 but is still higher than the TPF mean of the rest of the approaches (KNN, SVM, Bayesian and IMPST).
Statistical ANOVA test is used on the results of the proposed segmentation methods with comparison to the results of previous methods. In order to find the statistical significant of each proposed method, ANOVA is applied on BMRI-MASRG evaluation results first and the results are summarized in Table 4.17. Then same test is applied on the evaluation results of BMRI-SRGPSOC and the results are summarized in Table 4.18.
131
Table 4.17 Summary of the ANOVA tests analysis for BMRI-MASRG results compared with the results of the other approaches (KNN, SVM, Bayesian, FCM and IMPST). Jaccard
MCR
TPF
TNF
STVF
F statistic
5.5429
2.365
0.9479
9.2330
8.5980
P value
0.0002
0.045
0.4540
0.0001
0.0001
Table 4.18 Summary of the ANOVA tests analysis for BMRI-SRGPSOC results compared with the results of the other approaches (KNN, SVM, Bayesian, FCM and IMPST). Jaccard
MCR
TPF
TNF
STVF
F statistic
5.9790
3.4014
1.2045
9.5610
5.4619
P value
0.0001
0.0072
0.3132
0.0001
0.0002
From Table 4.17, it can be noticed that there are a statistically significant differences between BMRI-MASRG and other methods. Where (p < 0.05) is determined by one-way ANOVA for Jaccard (F = 5.5429, p = 0.0002), MCR (F = 2.365, p = 0.045), TNF (F = 9.2330, p = 0.0001) and STVF (F = 8.598, p = 0.0001), while there is no significant difference for TPF (F = 0.9479, p = 0.4540).
On the other hand, Table 4.18 shows that significant differences of ANOVA test for BMRI-SRGPSOC and previous methods results compared to BMRI-MASRG. The enhancement is slightly for most of the measures such as Jaccard (F = 5.9790, p = 0.0001), MCR (F = 3.4014, p = 0.0072) and TNF (F = 9.5610, p = 0.0001), while the results of
132
STVF is less significance (F = 5.4619, p = 0.0002). Again, there is no significant for TPF (F = 1.2045, p = 0.3132) although it is better. Figures 4.18, 4.19, 4.20, 4.21 and 4.22 show the statistical ANOVA graphs for the average and the variance of the evaluation measures; TPF, TNF, STVF, Jaccard and MCR.
Ϭ͘ϵ
dƌƵĞWŽƐŝƚŝǀĞ&ƌĂĐƚŝŽŶ
Ϭ͘ϴ Ϭ͘ϳ Ϭ͘ϲ Ϭ͘ϱ Ϭ͘ϰ Ϭ͘ϯ
ǀĞƌĂŐĞ
Ϭ͘Ϯ
sĂƌŝĂŶĐĞ
Ϭ͘ϭ Ϭ
dŚĞŵĞƚŚŽĚƐ
Figure 4.18 Statistical ANOVA graphs for BMRI-MASRG and BMRI-SRGPSOC in comparison with the previous methods using results of TPF measure.
133
ϭ Ϭ͘ϵ dƌƵĞEĞŐĂƚŝǀĞ&ƌĂĐƚŝŽŶ
Ϭ͘ϴ Ϭ͘ϳ Ϭ͘ϲ Ϭ͘ϱ Ϭ͘ϰ ǀĞƌĂŐĞ
Ϭ͘ϯ
sĂƌŝĂŶĐĞ
Ϭ͘Ϯ Ϭ͘ϭ Ϭ
dŚĞŵĞƚŚŽĚƐ
Figure 4.19 Statistical ANOVA graphs for BMRI-MASRG and BMRI-SRGPSOC in comparison with the previous methods using results of TNF measure. Ϯ
^ƵŵŽĨdƌƵĞsŽůƵŵĞ&ƌĂĐƚŝŽŶ
ϭ͘ϴ ϭ͘ϲ ϭ͘ϰ ϭ͘Ϯ ϭ Ϭ͘ϴ ǀĞƌĂŐĞ
Ϭ͘ϲ
sĂƌŝĂŶĐĞ
Ϭ͘ϰ Ϭ͘Ϯ Ϭ
dŚĞŵĞƚŚŽĚƐ
Figure 4.20 Statistical ANOVA graphs for BMRI-MASRG and BMRI-SRGPSOC in comparison with the previous methods using results of STVF measure.
134
Ϭ͘ϵ Ϭ͘ϴ Ϭ͘ϳ
:ĂĐĐĂƌĚ
Ϭ͘ϲ Ϭ͘ϱ Ϭ͘ϰ Ϭ͘ϯ
ǀĞƌĂŐĞ
Ϭ͘Ϯ
sĂƌŝĂŶĐĞ
Ϭ͘ϭ Ϭ
dŚĞŵĞƚŚŽĚƐ
Figure 4.21 Statistical ANOVA graphs for BMRI-MASRG and BMRI-SRGPSOC in comparison with the previous methods using results of Jaccard measure. Ϭ͘ϯ
DŝƐĐůĂƐƐŝĨŝĐĂƚŝŽŶZĂƚĞ
Ϭ͘Ϯϱ Ϭ͘Ϯ Ϭ͘ϭϱ
ǀĞƌĂŐĞ
Ϭ͘ϭ
sĂƌŝĂŶĐĞ Ϭ͘Ϭϱ Ϭ
dŚĞŵĞƚŚŽĚƐ
Figure 4.22 Statistical ANOVA graphs for BMRI-MASRG and BMRI-SRGPSOC in comparison with the previous methods using results of MCR measure.
135
For a more statistically evaluation, ROC curve analysis is used to illustrate the TPF (Sensitivity) rate compared with the FPF (1- Specificity) rate for both BMRI-MASRG and BMRI-SRGPSOC in comparison with previous methods as shown in Figure 4.23.
The high AUC indicates improved segmentation performance; a higher value of TPF is achieved at each value of FPF. The AUC results are listed in Table 4.18.
Figure 4.23
The ROC curves for the proposed method and the previous methods.
136
Table 4.19 Area under the Curve for the proposed approaches (BMRI-MASRG and BMRI-SRGPSOC) compared to previous methods (KNN, SVM, Bayesian, FCM and IMPST). The method
AUC
KNN
0.95
SVM
0.95
Bayesian
0.94
FCM
0.94
IMPST
0.97
BMRI-MASRG
0.97
BMRI-SRGPSOC
0.97
From Table 4.19, it can be observed that the values of AUC are the same for the proposed methods (BMRI-MASRG and BMRI-SRGPSOC) which are 0.97. The results show that the AUC values for proposed methods are improved compared to AUC results of previous methods, i.e. KNN, SVM, Bayesian, and FCM, with range of 0.94 to 0.95. However, AUC of proposed methods are equal to the AUC value of IMPST.
4.5 Summary This chapter presented the results of the proposed methods of the segmentation approach. The evaluation criteria explained in Chapter 3 are applied on the different stages of the methodology. Subsequently, statistical analysis tests have been applied in order to find
137
the significance of the methods. All experiments are done on the MRI breast image from RIDER dataset.
In the first stage, breast skin-line exclusion, two evaluation approaches used in the experiments, breast skin-line segmentation process scored high results using two evaluation approaches; first approach (TPF = 0.8128, FNF = 0.0926, FPF = 0.0231, TNF = 0.9401 and STVF = 1.7530), second approach (Jaccard = 0.7708, MCR = 0.1367 and Dice = 0.8404). The breast skin-line removal process also scored high results using the same evaluation approaches; first approach (TPF = 0.8583, FNF = 0.0062, FPF = 0.0484, TNF = 0.9708 and STVF = 1.8291), second approach (Jaccard = 0.8740, MCR = 0.1418 and Dice = 0.9321). ROC curve is applied for both processes and showed improvements, where AUC for the segmentation stage is 0.99 and for the removal stage is 0.95.
In the second stage, image thresholding using the proposed method (MMRT), pixel based evaluation approach and quality evaluation approach are applied the results of the experiments, the results of MMRT are compared to six standard thresholding approaches using the same RIDER dataset images in order to evaluate the improvement of the method. The pixel based evaluation approach demonstrates improved results when compared with the slandered methods (Jaccard = 0.64 and Dice = 0.68). The quality evaluation approach also showed better results when compared with the slandered methods (PSNR = 69.97 and MSE = 0.01). Furthermore, the statistical analysis of ANOVA proved the significant differences between the methods where p values are less than 0.05 for all measures (Jaccard = 0.78 E-14, Dice = 0.31 E-17, PSNR = 0.36 E-117 and MSE = 0.55 E-15).
138
In tumour segmentation stage, both proposed methods (BMRI-MASRG and BMRISRGPSOC) are tested, and the results are evaluated with three evaluation approaches. The results of first evaluation showed comparable outcomes; for BMRI-MASRG, the measures results are TPF = 0.82, FNF = 0.18, FPF = 0.10, TNF = 0.90 and STVF = 1.73. Then, for BMRI-SRGPSOC, the results are TPF = 0.84, FNF = 0.18, FPF = 0.09, TNF = 0.91 and STVF = 1.71. In addition, both proposed methods results are improved comparing to the previous methods, i.e. KNN, SVM, Bayesian, FCM and IMPST. In the second evaluation, BMRI-SRGPSOC showed better results in comparison to BMRIMASRG, where BMRI-MASRG results are (Jaccard = 0.75, MCR = 0.18) and BMRISRGPSOC (Jaccard= 0.77, MCR= 0.17). In the third evaluation approach, the accuracy of seed selection for seeded region growing is calculated. TKH VHHG¶V SL[HO ; DQG
