Dissertation_Mahsa Chitsaz.pdf - DSpace@UM - University of Malaya
Dissertation_Mahsa Chitsaz.pdf - DSpace@UM - University of Malaya
Dissertation_Mahsa Chitsaz.pdf - DSpace@UM - University of Malaya
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
Medical Image Segmentation by Hybridizing Multi-Agent System and<br />
Reinforcement Learning Agent<br />
<strong>Mahsa</strong> <strong>Chitsaz</strong><br />
Supervisor: Dr. Woo Chaw Seng<br />
Faculty <strong>of</strong> Computer Science and Information Technology,<br />
<strong>University</strong> <strong>of</strong> <strong>Malaya</strong><br />
December 2009
UNIVERSITI MALAYA<br />
ORIGINAL LITERARY WORK DECLARATION<br />
Name <strong>of</strong> Candidate: (I.C/Passport No: )<br />
Registration/Matric No:<br />
Name <strong>of</strong> Degree:<br />
Title <strong>of</strong> Project Paper/Research Report/<strong>Dissertation</strong>/Thesis (“this Work”):<br />
Field <strong>of</strong> Study:<br />
I do solemnly and sincerely declare that:<br />
(1) I am the sole author/writer <strong>of</strong> this Work;<br />
(2) This Work is original;<br />
(3) Any use <strong>of</strong> any work in which copyright exists was done by way <strong>of</strong> fair dealing<br />
and for permitted purposes and any excerpt or extract from, or reference to or<br />
reproduction <strong>of</strong> any copyright work has been disclosed expressly and<br />
sufficiently and the title <strong>of</strong> the Work and its authorship have been acknowledged<br />
in this Work;<br />
(4) I do not have any actual knowledge nor do I ought reasonably to know that the<br />
making <strong>of</strong> this work constitutes an infringement <strong>of</strong> any copyright work;<br />
(5) I hereby assign all and every rights in the copyright to this Work to the<br />
<strong>University</strong> <strong>of</strong> <strong>Malaya</strong> (“UM”), who henceforth shall be owner <strong>of</strong> the copyright<br />
in this Work and that any reproduction or use in any form or by any means<br />
whatsoever is prohibited without the written consent <strong>of</strong> UM having been first<br />
had and obtained;<br />
(6) I am fully aware that if in the course <strong>of</strong> making this Work I have infringed any<br />
copyright whether intentionally or otherwise, I may be subject to legal action or<br />
any other action as may be determined by UM.<br />
Candidate’s Signature Date<br />
Subscribed and solemnly declared before,<br />
Name:<br />
Designation:<br />
Witness’s Signature Date<br />
ii
To My Beloved Mother and Father<br />
iii
Abstract<br />
Image segmentation is still a debatable problem although there have been many research<br />
work done in the last few decades. First <strong>of</strong> all, every solution for image segmentation is<br />
problem-based. Secondly, medical image segmentation methods generally have<br />
restrictions because medical images have very similar gray level and texture among the<br />
interested objects.<br />
Therefore, this dissertation presents a framework to extract simultaneity several objects<br />
<strong>of</strong> interest from head Computed Tomography images. The proposed method contains<br />
two phases; training and testing. A Reinforcement-Learning method is proposed for the<br />
training phase, and a new Multi-Agent system is proposed for the testing phase. In the<br />
training phase, a few images are used as a trained image whereas the RL agent will find<br />
the appropriate value <strong>of</strong> each object or region in the input image. The outcome <strong>of</strong> this<br />
training phase is transferred to the next phase, testing phase. In this phase, the images<br />
are segmented by some priori knowledge and the properties <strong>of</strong> local agent.<br />
Proposed reinforcement learning model attains significant result in segmentation<br />
accuracy; the accuracy is more than 95% for each region in the image and the mean<br />
computation time <strong>of</strong> all datasets is less than 13 seconds. Moreover, the number <strong>of</strong><br />
training data set for PRLM can be one or a small number <strong>of</strong> images. Also, PRLM has<br />
the ability to segment simultaneously an image into some distinct regions.<br />
Proposed multi-agent model attains considerable result in segmentation accuracy; the<br />
accuracy is more than 90% for each region in the image and the mean computation time<br />
iv
<strong>of</strong> all datasets is less than 7 seconds. Furthermore, PMAM is capable to segment<br />
simultaneously an image into some distinct regions.<br />
v
Table <strong>of</strong> Contents<br />
DECLARATION………………………………………………………………. ii<br />
DEDICATION…………………………………………………………………. iii<br />
ABSTRACT……………………………………………………………………. iv<br />
TABLE OF CONTENTS………………………………………………………. vi<br />
LIST OF FIGURES ……………………………………………………………. viii<br />
LIST OF TABLES………………………………………………………...…… ix<br />
LIST OF ABRIVATIONS AND SYMBOLS………………………………….. x<br />
LIST OF PUBLICATIONS……………………………………………………. xii<br />
ACKNOWLEDGEMENT……………………………………………………… xiii<br />
1. Introduction………………………………………………………………… 1<br />
1.1 Background………………………………………………………… 1<br />
1.2 Motivation …………………………………………………………. 2<br />
1.3 Problem Description………………………………………………... 4<br />
1.4 Goal and Objectives………………………………………………... 4<br />
1.5 Scope <strong>of</strong> the project………………………………………………… 5<br />
1.6 <strong>Dissertation</strong> Organization…………………………………………... 5<br />
2. Literature Review………………………………………………………….. 7<br />
2.1 Overview……………………………………………………………… 7<br />
2.2 Three-Dimensional Medical Imaging and Skull Anatomy…………… 8<br />
2.2.1 Computed Tomography Images.…………………………………. 9<br />
2.2.2 Magnetic Resonance Imaging……………………………………. 10<br />
2.2.3 Skull Anatomy……………………………………………………. 12<br />
2.2.3 Cranial Bones………….…………………………………. 12<br />
2.2.3 Facial Bones…………...…………………………………. 13<br />
2.2.3 Anatomical Structure that Used as the Case Study....……. 14<br />
2.3 Segmentation Methods………………………………………………... 17<br />
2.3.1 Several Classifications <strong>of</strong> Segmentation Methods……………….. 18<br />
2.3.2 Brief Description <strong>of</strong> Main Segmentation Methods………………. 22<br />
2.3.2.1 Thresholding…………………….……………………… 22<br />
2.3.2.2 Region Growing………………...……………………… 23<br />
2.3.2.3 Edge Detection………………….……………………… 24<br />
2.3.2.4 Classifiers……………………….……………………… 24<br />
2.3.2.5 Clustering……………………….……………………… 25<br />
2.3.2.6 Deformable Models…………….……………………… 26<br />
2.3.2.7 Neural Network…..…………….……………………… 26<br />
2.3.3 Comparison <strong>of</strong> the Segmentation Methods………………………. 27<br />
2.4 Agent and Multi-Agent System………………………………………. 28<br />
2.5 Standard Reinforcement Learning Model…………………………….. 30<br />
2.6 Image Segmentation Methods by Autonomous Agents in Multi-Agent<br />
System………………………………………………………………….. 33<br />
2.6.1 Kagawa et al method…………………………………………...… 33<br />
2.6.2 Wang and Yuan method………………………………………..… 34<br />
2.6.3 Gyohten method………………………………………………….. 35<br />
2.6.4 Guillaud et al. method…………………………………………..... 37<br />
2.6.5 Rodin et al. method……………………………………………..... 38<br />
2.6.6 Melkemi et al. method……………...…………………………….. 39<br />
2.6.7 Spinnu et al. method…………………………………...…………. 40<br />
2.6.8 Boucher et al. method…………………………………..………… 42<br />
vi
2.6.9 Liu and Tang method……………………………………...……... 43<br />
2.6.10 Germond et al. method……………………………………..…… 44<br />
2.6.11 Duchesnay et al. method………………………………………... 45<br />
2.6.12 Khosla and Lai method……………………...………………….. 47<br />
2.6.13 Richard et al. method…………………………………………… 49<br />
2.6.14 Benamrane and Nassane method………………………………... 50<br />
2.6.15 Discussion…………………………………………………......... 51<br />
2.7 Image Segmentation Methods by Reinforcement Learning Model…... 55<br />
2.7.1 Peng and Bhanu method………………………………………….. 55<br />
2.7.2 Shokri method…………………………………………………..... 57<br />
2.7.3 Sahba method…………………………………………………….. 57<br />
2.8 Chapter Summary……………………………………………………... 59<br />
3. Methodology………………………………………………………………... 61<br />
3.1 Image Acquisition..…………………………………………………… 61<br />
3.2 Image Segmentation…………………………………………………... 62<br />
3.2.1 Training Phase……………………………………………………. 66<br />
3.2.1.1 Definition <strong>of</strong> States …………….……………………… 69<br />
3.2.1.2 Definition <strong>of</strong> Actions ….……….……………………… 72<br />
3.2.1.3 Definition <strong>of</strong> Reward…………………………………… 73<br />
3.2.1.4 Graphical User Interface for Training Phase …..……… 73<br />
3.2.2 Testing Phase……………………………………………………... 74<br />
3.2.2.1 Graphical User Interface for Testing Phase …..……….. 78<br />
3.3 Chapter Summary……………………………………………………... 79<br />
4. Experimental Results and Discussion …..………………………………... 81<br />
4.1 Experiment Result <strong>of</strong> Training Procedure……………………….……. 81<br />
4.1.1 Image Data Sets <strong>of</strong> PRLM…………….………..………………... 82<br />
4.1.2 Qualitative Analysis <strong>of</strong> PRLM…………………………………… 83<br />
4.1.3 Quantitative Analysis <strong>of</strong> PRLM………………………….………. 83<br />
4.1.3.1 Accuracy <strong>of</strong> PRLM …..…………………… …..……… 85<br />
4.1.3.2 Efficiency <strong>of</strong> PRLM ..…………………….. …..……… 88<br />
4.2 Experiment Result <strong>of</strong> Testing Procedure……………………………... 88<br />
4.2.1 Image Data Sets <strong>of</strong> PMAM….…………………………………... 91<br />
4.2.2 Qualitative Analysis <strong>of</strong> PMAM …………………………….…... 91<br />
4.2.3 Quantitative Analysis <strong>of</strong> PMAM ……………………………..… 93<br />
4.2.3.1 Accuracy <strong>of</strong> PMAM ……………………….…..……… 93<br />
4.2.3.2 Efficiency <strong>of</strong> PMAM .…………………….. …..……… 95<br />
4.3 Chapter Summary……..…………………………………………..….. 96<br />
5. Conclusions…………………………………………………………………. 97<br />
5.1 The Proposed Reinforcement-Learning model ……………………..... 97<br />
5.2 The Proposed Multi-Agent model………………………………….…. 99<br />
5.3 Achievements………………………………………………………..... 101<br />
5.4 Future work………………………………………………………….... 102<br />
Bibliography…………………………………………………………………... 104<br />
Appendix A: Experimental Results <strong>of</strong> the Training Phase…...……………... 110<br />
Appendix B: TPVF and FPVF <strong>of</strong> the Experimental Results <strong>of</strong> the Training<br />
Phase…………………………………………………………... 114<br />
Appendix C: Experimental Results <strong>of</strong> the Testing Phase.…………………… 116<br />
vii
Appendix D: TPVF and FPVF <strong>of</strong> the Experimental Results <strong>of</strong> the Testing<br />
Phase…………………………………………………………... 119<br />
viii
List <strong>of</strong> Figures<br />
Figure 1.1: (a) A Slice <strong>of</strong> CT Image <strong>of</strong> a Human Head (b) Segmented Image <strong>of</strong> Figure<br />
1.1(a)…………………………………………………………………………... 4<br />
Figure 2.1: CT Scanner (Jeri 2008)………………………………………………………... 10<br />
Figure 2.2: Human Head CT Slices at Axial View (Obaidellah 2006)…………………… 10<br />
Figure 2.3: MRI Scanner (Garrobo 2006)…………………………………………………. 11<br />
Figure 2.4: The Lateral (left) and the Anterior (right) View <strong>of</strong> Skull (Gray 1918)……….. 12<br />
Figure 2.5: Cranial Bones <strong>of</strong> the Skull (Martini 2004)……………………….…………… 13<br />
Figure 2.6: (a): The Maxillae Bone (b): the Mandible Bone (Gray 1918)………………… 14<br />
Figure 2.7: The Top Side <strong>of</strong> Skull (Walter 2007)…………………………………………. 15<br />
Figure 2.8: (a) CT Image <strong>of</strong> 1-3 Section (b) CT Image <strong>of</strong> 1-6 Section (Obaidellah<br />
2006)………………………………………………………………….……….. 15<br />
Figure 2.9: The Middle Part <strong>of</strong> Skull (Walter 2007)………………….………….………... 15<br />
Figure 2.10: (a) CT Image <strong>of</strong> 1-10 Section (b) CT Image <strong>of</strong> 1-12 Section (c) CT Image <strong>of</strong><br />
1-15 Section (d) CT Image <strong>of</strong> 1-20 Section (Obaidellah<br />
2006)…………………………………………………………………………... 16<br />
Figure 2.11: The Inferior Part <strong>of</strong> the Skull (Walter 2007)………………………………….. 16<br />
Figure 2.12: (a) CT Image <strong>of</strong> 1-22 Section (b) CT Image <strong>of</strong> 1-23 Section (Obaidellah<br />
2006)…………………………………………………….…….......................... 17<br />
Figure 2.13: Uniformity Predicate <strong>of</strong> the Segmentation (Awcock 1995)……..……………. 17<br />
Figure 2.14: A Clustering Approach (Jain 1989)………………………………..……….…. 25<br />
Figure 2.15: The Internal Structure <strong>of</strong> a Typical Agent (<strong>Chitsaz</strong> 2008)………………..…... 29<br />
Figure 2.16: The Standard RL Model (Kaelbling 1996)…………..………………………... 31<br />
Figure 2.17: Pseudocode for Q-Learning Algorithm (Watkins, 1989)……………..…… 32<br />
Figure 2.18: The Proposed Method <strong>of</strong> Kagawa (Kagawa 1999) …………………………... 34<br />
Figure 2.19: The Proposed Method <strong>of</strong> Gyothen (Gyohten 2000).......................…………… 36<br />
Figure 2.20: The Procedure <strong>of</strong> Proposed Method by Guillaud et al. (Guillaud 2000) ……... 37<br />
Figure 2.21: The MAS which Proposed by Rodin et al.(Rodin 2004) …………………...… 38<br />
Figure 2.22: The Finite State Machine that is describing an Agent’s Behavior (Rodin<br />
2004)…………………………………………………………………………... 39<br />
Figure 2.23: The Architecture <strong>of</strong> the Proposed Method <strong>of</strong> Melkemi (Melkemi 2006) ……. 40<br />
Figure 2.24: The Proposed Method using MAS by Spinu (Spinu 1996) …….…………….. 41<br />
Figure 2.25: The Proposed MAS by Boucher et al.(Boucher 1998) …………….…………. 43<br />
Figure 2.26: The Local Neighboring Region <strong>of</strong> an Agent at Location (i,j) (Liu 1999)…..… 44<br />
Figure 2.27: A Global View <strong>of</strong> the Framework and Information Flow <strong>of</strong> the Proposed<br />
Method by Germond (Germond 2000)……..……………………………….… 45<br />
Figure 2.28: The Conceptual Framework <strong>of</strong> Duchesnay et al. (Duchesnay 2001) ………… 46<br />
Figure 2.29: The Graphical Representations <strong>of</strong> the Seven Behaviors (Duchesnay 2003) …. 47<br />
Figure 2.30: The Multi-Agent Optimization Model for Unstained Cell Images by Khosla<br />
et al. (Khosla 2003) ……...……………………………………........................ 48<br />
Figure 2.31: The Multi-Agent S<strong>of</strong>t Computing Model for Unstained Cell Images by Lai et<br />
al. (Lai 2003) ……….……………………………………………………...…. 48<br />
Figure 2.32: The Proposed Multi-Agent Framework by Richard et al.(Richard 2004) ……. 50<br />
Figure 2.33: The Proposed Method by Benamrane et al. (Benamrane 2007) ……………… 51<br />
Figure 2.34: The Conceptual Diagram <strong>of</strong> the Phoenix Segmentation Algorithm (Peng<br />
1998 b)………………………………………………………………………… 56<br />
Figure 2.35: The Segmentation Evaluation by RL (Bhanu 2000) ……................................. 56<br />
Figure 2.36: The Standard Model <strong>of</strong> RL (Shokri, 2003) …………………………………... 57<br />
Figure 2.37: The RL Model used in the Proposed Method by Sahba (Sahba 2006 b) …….. 58<br />
Figure 2.38: The General Model used in the Proposed Method (Sahba 2008) ………..…… 58<br />
Figure 3.1: The Global View <strong>of</strong> our Proposed Model...…………….…………………….. 63<br />
Figure 3.2: An Example for Calculating the Number <strong>of</strong> the Agents within a Window Size<br />
<strong>of</strong> 7�7 over an Image Size <strong>of</strong> 16�16…...……..…..………………………… 65<br />
Figure 3.3: (a) The Original CT Image, (b) The Manually Segmented Image……………. 66<br />
Figure 3.4: The Global View <strong>of</strong> our Proposed Method in Training Phase...……………… 67<br />
Figure 3.5: The RL Agent’s Behavior.…………………………………………….…….... 68<br />
Figure 3.6: The Example <strong>of</strong> the number <strong>of</strong> States for an Image with two Regions……….. 70<br />
Figure 3.7: The Example <strong>of</strong> the number <strong>of</strong> States for an Image with three Regions; each<br />
window shows a typical sub-image…………………………………………... 71<br />
ix
Figure 3.8: An Example <strong>of</strong> Defining Action using the Maximum and Minimum<br />
Thresholding Gray-scale Value <strong>of</strong> a Typical Sub-image………………...…… 72<br />
Figure 3.9: GUI <strong>of</strong> the Training Phase..……………………..…………….………………. 74<br />
Figure 3.10: The Global View <strong>of</strong> our Proposed Method in Testing Phase.......……………. 75<br />
Figure 3.11: The Agents’ Behavior in Testing Phase….…..……………...………………... 77<br />
Figure 3.12: GUI <strong>of</strong> the Testing Phase..……………………………………………………. 79<br />
Figure 4.1: The Segmentation Example from Two Experiments and Four Different Slices<br />
<strong>of</strong> 3D Images, (a) Input Image (b) Result from Proposed Method and (c)<br />
Ground Truth Image…………………………………………………………... 84<br />
Figure 4.2: ROC Curve for the First Data set…………………….……………………….. 87<br />
Figure 4.3: ROC Curve for the Second Data set……….………………………………….. 87<br />
Figure 4.4: GUI <strong>of</strong> the Training Phase to Suggest the User Thresholding Range <strong>of</strong> each<br />
Region...……………………………………………………………………….. 90<br />
Figure 4.5: GUI <strong>of</strong> the Testing Phase…………….……………………………………….. 90<br />
Figure 4.6: The Segmentation Example from Two Experiments and Four Different Slices<br />
<strong>of</strong> Data sets, (a) Result from our Method, (b) Input Image.…………………... 92<br />
Figure 4.7: ROC Curve <strong>of</strong> the First Data set…………………………….………………… 94<br />
Figure 4.8: ROC Curve <strong>of</strong> the Second Data set …………………………………………... 94<br />
Figure 4.9: The Computation Time <strong>of</strong> all Data sets; X axis shows the image number<br />
(identity) and Y axis shows the computation time …...…………………….… 96<br />
x
List <strong>of</strong> Tables<br />
Table 2.1 The Comparison among the Segmentation Methods (<strong>Chitsaz</strong> 2008) …………… 27<br />
Table 2.2: The Comparison <strong>of</strong> the Segmentation Methods using Non-medical Images by<br />
Agent Properties (<strong>Chitsaz</strong> 2008) ………………………………………………. 52<br />
Table 2.3: The Comparison <strong>of</strong> the Segmentation Methods using Medical Images by Agent<br />
Properties (<strong>Chitsaz</strong> 2008) ….……………………………………………………. 53<br />
Table 2.4: The Comparison between Multi-Agent and Non-Agent Segmentation Methods<br />
(<strong>Chitsaz</strong> 2008)…………………………………………………………………… 54<br />
Table 4.1: Details <strong>of</strong> the Image Data set used in PRLM……………..…………………… 82<br />
Table 4.2: TPVF and FPVF <strong>of</strong> PRLM…………………………..……………….……….. 86<br />
Table 4.3: Efficiency <strong>of</strong> the PRLM……….………………………………….…..………. 88<br />
Table 4.4: Details <strong>of</strong> the Image Data set used in PMAM…………...…………………… 91<br />
Table 4.5: TPVF and FPVF for the Testing Phase <strong>of</strong> PMAM…………..……………….. 93<br />
Table 4.6: Mean User-interaction Time and Computation Time <strong>of</strong> PMAM………………. 95<br />
Table 5.1: Efficiency Comparison <strong>of</strong> Image Segmentation Methods………………………. 98<br />
xi
A. List <strong>of</strong> Abbreviations<br />
List <strong>of</strong> Abbreviations and Symbols<br />
2D Two-Dimensional<br />
3D Three-Dimensional<br />
CT Computed Tomography<br />
DICOM Digital Imaging and Communications in Medicine<br />
EM Expectation-Maximization<br />
FPS Frames per Second<br />
FPVF False Positive Volume Fraction<br />
GA Genetic Algorithm<br />
GUI Graphical User Interface<br />
HSV Hue, Saturation, and Value<br />
ICA Intelligent Control Agent<br />
IPA Image Processing Agents<br />
KP Knowledge Processors<br />
KS Knowledge Servers<br />
MAS Multi-Agent System<br />
MRI Magnetic Resonance Imaging<br />
PMAM Proposed Multi-Agent Model<br />
PRLM Proposed Reinforcement-Learning Model<br />
RF Radi<strong>of</strong>requency<br />
ROC Receiver Operating Characteristic<br />
RL Reinforcement Learning<br />
SPRT Sequential Probability Ratio Test<br />
SRLM Standard Reinforcement Learning Model<br />
TPVF True Positive Volume Fraction<br />
xii
B. List <strong>of</strong> Symbols<br />
� Complete array <strong>of</strong> image pixels<br />
f(x, y) Intensity <strong>of</strong> pixel at (x, y)<br />
Q(s,a) Sum <strong>of</strong> future pay<strong>of</strong>fs r obtained by taking action a from state s.<br />
s State<br />
a Action<br />
r Reward<br />
� Learning Rate<br />
� Discount Factor<br />
� Probability Factor <strong>of</strong> � -Greedy Algorithm<br />
C M<br />
d<br />
Segmented Image<br />
A� ( C,<br />
f ) Image<br />
C 2D rectangular array <strong>of</strong> pixels<br />
f (c)<br />
Intensity <strong>of</strong> any pixel c in C<br />
U d<br />
Binary image representation <strong>of</strong> a reference superset <strong>of</strong> pixels<br />
xiii
List <strong>of</strong> Publications<br />
<strong>Chitsaz</strong>, M., Woo, C.S. (2008). The Rise <strong>of</strong> Multi-Agent and R.L. Segmentation Methods<br />
for Biomedical Images. The 4th Malaysian S<strong>of</strong>tware Engineering Conference<br />
(MySEC’08), Kuala Terengganu, Malaysia.<br />
Abstract-Image segmentation is an important operation in image analysis. We present a critical<br />
assessment <strong>of</strong> conventional methods for image segmentation. Current segmentation approaches<br />
have some common disadvantages. They are sensitive to noise and require manual fine-tuning. We<br />
found that multi-agent and reinforcement learning (RL) methods are suitable for biomedical image<br />
segmentation. They have shown very outstanding results and high adaptability in most <strong>of</strong> the<br />
segmentation scenarios.<br />
<strong>Chitsaz</strong>, M., Woo, C.S. (2009). Medical Image Segmentation using Reinforcement<br />
Learning Agent. International Conference on Digital Image Processing (ICDIP'09),<br />
Bangkok, Thailand, IEEE Computer Society Press.<br />
Abstract—the principal goal <strong>of</strong> this work is to design a framework to extract one or more objects <strong>of</strong><br />
interest from Computed Tomography (CT) images. The learning phase is based on reinforcement<br />
learning (RL). The input image is divided into several sub-images, and each RL agent works on it<br />
to find the suitable value for each object in the image. For each state in the environment has been<br />
defined some actions; also a reward function compute reward for each action <strong>of</strong> RL agent. Finally<br />
the valuable information is stored in Q-Matrix, and the final result can be used to segment new<br />
similar input images by applying it. The experimental results for cranial CT image show accuracy<br />
<strong>of</strong> segmented image is more that 95%.<br />
<strong>Chitsaz</strong>, M., Woo, C.S. (2009). A Multi-Agent System Approach for Medical Image<br />
Segmentation. International Conference on Future Computer and Communication<br />
(ICFCC'09), Kuala Lumpur, Malaysia, IEEE Computer Society Press.<br />
Abstract— Image segmentation still requires improvements although there have been research<br />
works since the last few decades. This is coming due to some issues. Firstly, most image<br />
segmentation solutions are problem-based. Secondly, medical image segmentation methods<br />
generally have restrictions because medical images have very similar gray level and texture among<br />
the interested objects. The goal <strong>of</strong> this work is to design a framework to extract simultaneously<br />
several objects <strong>of</strong> interest from Computed Tomography (CT) images by using some prioriknowledge.<br />
Our method used properties <strong>of</strong> agent in a multi-agent environment. The input image is<br />
divided into several sub-images, and each agent works on a sub-image and tries to mark each pixel<br />
as a specific region by means <strong>of</strong> given priori-knowledge. During this time the local agent marks<br />
each cell <strong>of</strong> sub-image individually. Moderator agent checks the outcome <strong>of</strong> all agents’ work to<br />
produce final segmented image. The experimental results for cranial CT images demonstrated<br />
segmentation accuracy around 90%.<br />
xiv
Acknowledgement<br />
I would like to thank Dr. Woo Chaw Seng, my supervisor, for patiently giving me good<br />
advice over the course <strong>of</strong> this program, and for spending a lot <strong>of</strong> time talking to me<br />
about ideas that led up to this works.<br />
Financial support from Dr Woo along with an ample supply <strong>of</strong> awards and<br />
Fellowship/research assistantship from, or through, Institute <strong>of</strong> Graduate Studies,<br />
Institute <strong>of</strong> Research Management and Monitoring, Faculty <strong>of</strong> Computer Science and<br />
Information Technology, and Department <strong>of</strong> Artificial Intelligence contributed greatly<br />
toward the completion <strong>of</strong> this research.<br />
I would like to thank all my pr<strong>of</strong>essors and colleagues from whom I learned so much.<br />
Special thanks go to my parents, my brothers and my friends, for their continual love,<br />
support, and encouragement throughout my time in graduate school. A particular debt<br />
<strong>of</strong> thanks is due to Hadi for his efforts in preparing many <strong>of</strong> the line drawings in my<br />
dissertation.<br />
Certain studies described in this thesis would not have been possible without images<br />
from <strong>University</strong> <strong>of</strong> <strong>Malaya</strong> Hospital.<br />
xv
Chapter 1<br />
Introduction<br />
1.1 Background<br />
Medical images can represent in a two-dimensional array <strong>of</strong> picture element (pixel) or<br />
in a 3D array (voxel). Medical images are normally stored accordingly to Digital<br />
Imaging and Communications in Medicine (DICOM) standard for distributing and<br />
viewing <strong>of</strong> any medical image regardless <strong>of</strong> its origin. Digital image processing is<br />
concerned with the analysis and manipulation <strong>of</strong> images by computer. Enhancement,<br />
segmentation, quantification, registration, visualization, and compression are the most<br />
famous tasks in image processing and computer vision (Bankman, 2008). Image<br />
segmentation is a pre-processing task for separating an image to several components. In<br />
other words, segmentation is a process <strong>of</strong> identifying the objects in an image. For<br />
example in a computed tomography (CT) image <strong>of</strong> the head, the components may<br />
consist <strong>of</strong> bone, tooth, fat, etc. In another definition, image segmentation is categorizing<br />
<strong>of</strong> the image to some disjoint partitions whereas the whole <strong>of</strong> partitions reconstruct the<br />
primarily image again (Pham, 2000).
Image segmentation techniques have been an invaluable task in many domains such as<br />
quantification <strong>of</strong> tissue volumes, diagnosis, localization <strong>of</strong> pathology, study <strong>of</strong><br />
anatomical structure, treatment planning, partial volume correction <strong>of</strong> functional<br />
imaging data, and computer-integrated surgery (Pham, 2000). Segmentation algorithms<br />
play a unique role in machine vision systems as it creates a bridge between the low-<br />
level and high-level processing operations. Low-level operation is carried out on image<br />
array <strong>of</strong> rows data thus adopts a bottom-up approach to image analysis. High-level<br />
processing is important with the manipulation <strong>of</strong> high-level data, abstract data<br />
presentation and thus favors a top-down approach. Segmentation can employ either or<br />
both <strong>of</strong> these approaches (Awcock, 1995).<br />
The image segmentation can perform manually but this is a time-consuming task.<br />
Therefore, automatic segmentation is preferable. Although there have been many<br />
researches carried out in last two decades, image segmentation is still a debatable<br />
problem (Liu, 2006). Image segmentation suffers from two main problems, which these<br />
generally caused the segmentation an unsolved problem. Firstly, every solution for<br />
image segmentation is problem-based. Secondly, the noisy nature <strong>of</strong> medical images<br />
makes it difficult to relate each pixel within an image to different texture classes<br />
(Withey, 2006).<br />
1.2 Motivation<br />
There is a broad range <strong>of</strong> facial diseases that could cause face deformation. Facial<br />
diseases involve facial area injuries and tumors, which inherit by syndromes such as<br />
Crouzon-Syndrom, Apert-Syndrom and developmental diseases like distributed growth<br />
<strong>of</strong> the jaw (Koch, 2002). Face is the first organ <strong>of</strong> body that has seen at the first sight.<br />
Everybody solicits that own face appear gracefully in a first sight so these general<br />
2
elieve his/her importance <strong>of</strong> facial diseases. Thus, the importance <strong>of</strong> facial surgery is<br />
very tangible, and the risk <strong>of</strong> this kind <strong>of</strong> treatment operation is very high. The specialist<br />
and the patient are very eligible to imagine the face after the surgery, however, the<br />
specialist would be predicting the post-operation <strong>of</strong> face but the outcome is not very<br />
clear for the patients. Consequently, simulating facial surgery and predicting the effect<br />
<strong>of</strong> an operation is very pivotal. One <strong>of</strong> the primary preprocessing levels in every kind <strong>of</strong><br />
surgical simulation is image segmentation. Therefore, the accuracy <strong>of</strong> this level is<br />
essential because the result will affect the overall outcome.<br />
Our motivation caused that we propose a method that can segment the head CT image<br />
<strong>of</strong> patient by the way that the result will be practical for the remainder <strong>of</strong> planning.<br />
Because the accuracy <strong>of</strong> the image segmentation can be impressed in the reconstruction<br />
<strong>of</strong> 3D image <strong>of</strong> patient face, the significant <strong>of</strong> the segmentation is very evident.<br />
The other motivation for doing this research is the current methods in image<br />
segmentation field have problems in some areas. The Multi-Agent System (MAS) has<br />
used in several image segmentations that investigate in chapter 2. However, in practice<br />
they still encounter the three main limitations:<br />
a. Since the Reinforcement Learning (RL) system is an effective to find the<br />
optimal result in the system. A few researchers have investigated this system<br />
for image segmentation. Furthermore, no research has attempted in the field<br />
<strong>of</strong> CT image segmentation.<br />
b. Other researchers do still not do combination <strong>of</strong> the characteristics <strong>of</strong> pure<br />
agent with the RL agent. We expect this can be resulted in a better<br />
segmentation outcome.<br />
3
c. Due to the lack <strong>of</strong> practical framework for segmentation, we endeavor to<br />
present a structure for image segmentation, which it can use for some<br />
different image modalities.<br />
1.3 Problem Description<br />
The problem in this dissertation is to segment CT image <strong>of</strong> the head, as shown in figure<br />
1.1(a), which the outcome <strong>of</strong> this process will be used for reconstruction <strong>of</strong> 3D image <strong>of</strong><br />
the face, or other post-processing steps. The aim <strong>of</strong> our segmentation method is to<br />
separate the skin tissue from the bone and the background. Figure 1.1(b) shows a<br />
segmented image <strong>of</strong> Figure 1.1(a). The red color shows the background or air, green is<br />
skin, and the blue color depicts the bone. The represented colors are optional. A new<br />
framework for this kind <strong>of</strong> image segmentation proposes using the MAS and a special<br />
type <strong>of</strong> agent, which calls RL agent.<br />
Figure 1.1 (a): A Slice <strong>of</strong> CT Image <strong>of</strong> a Human Head, (b): Segmented Image <strong>of</strong> (a)<br />
1.4 Goal and Objectives<br />
(a) (b)<br />
The goal <strong>of</strong> this dissertation is to propose a method to segment the CT image <strong>of</strong> a head.<br />
For the image segmentation <strong>of</strong> head, we have the following objectives:<br />
4
� Review <strong>of</strong> existing experimental studies and investigate the Agents Technology<br />
for image segmentation;<br />
� Propose a novel approach by RL agent in the multi-agent framework which will<br />
be quicker, more accurate, and more robust;<br />
� Evaluate the outcome <strong>of</strong> segmented image by the proposed method with an<br />
appropriate estimation approach.<br />
1.5 Project Scope<br />
The scope in this thesis is the segmentation <strong>of</strong> head CT image for identification <strong>of</strong><br />
targeted tissues, structure and bone, this problem is carried out by RL agent in the MAS,<br />
which they have used both the social ability <strong>of</strong> pure agent and the properties <strong>of</strong> RL<br />
agent. Therefore, the scope <strong>of</strong> problem is just concerned with CT image <strong>of</strong> head and the<br />
methodology which is proposed, is implemented by collaborating, cooperating and<br />
negotiation <strong>of</strong> RL agent. The data sets, which used in this thesis, consist <strong>of</strong> two different<br />
resources. The first data set includes 33 images from Hospital <strong>of</strong> <strong>University</strong> <strong>of</strong> <strong>Malaya</strong>.<br />
The second data set contains 28 images that downloaded from<br />
http://pubimage.hcuge.ch:8080/.<br />
1.6 <strong>Dissertation</strong> Organization<br />
The remainder <strong>of</strong> this thesis has the following components. The next chapter, literature<br />
review, is a concise perusal <strong>of</strong> the past and recent methods in medical image<br />
segmentation. Before summarizing and analysis those methods with the proposed<br />
framework, the history <strong>of</strong> agent and MAS, and the face anatomy present. There are brief<br />
descriptions <strong>of</strong> recent methods, and there are comparison between conventional<br />
5
methods and agent-based methods. In addition, advantages and disadvantages <strong>of</strong> the<br />
other methods mention and compare.<br />
Chapter 3 brings about our proposed methodology which consists two phases. The first<br />
phase bases on the RL approach. Another phase implements by multi-agent properties.<br />
In discussion section, some contributions <strong>of</strong> our proposed framework discuss.<br />
Chapter 4 discusses how we evaluate our proposed method, both qualitatively, through<br />
image display, and quantitatively, through evaluation experiments. In addition, the<br />
efficiency <strong>of</strong> proposed method compared to the other methods.<br />
The last chapter provides discussion for each phase <strong>of</strong> the proposed method. It mentions<br />
all contributions and weaknesses <strong>of</strong> proposed method. At the end, the achievement, and<br />
ideas for future works present.<br />
Finally, more results <strong>of</strong> our method provide in appendices.<br />
6
Chapter 2<br />
Literature Review<br />
2.1 Overview<br />
Based on the problems mentioned in chapter 1, this chapter presents review <strong>of</strong> recent<br />
works in image segmentation. In addition, the backgrounds <strong>of</strong> research described here.<br />
The Three-dimensional Medical Imaging section is about two commonly used<br />
modalities for medical imaging; Computed Tomography (CT) and Magnetic Resonance<br />
Imaging (MRI). The Skull Anatomy section described anatomy <strong>of</strong> the face, which uses<br />
in our research to segment the CT image <strong>of</strong> head. Moreover, different CT images are<br />
shown to clarify each image is related to different parts <strong>of</strong> head and its details.<br />
The Segmentation Methods section represents general methods, which are widely used<br />
in medical image segmentation. Moreover, those methods categorize based on the<br />
characteristics <strong>of</strong> the proposed model. This classification takes from a number <strong>of</strong><br />
publications regarding medical image segmentation. Since the publications are quite<br />
large the information intend to be representative rather than exhaustive.<br />
7
The Agent and MAS section defines agent and multi-agent system. In the end <strong>of</strong> the<br />
section, the advantages <strong>of</strong> using agent in a typical system are listed, on the other words,<br />
the main reason which we used the agent for segmenting medical image is mentioned.<br />
The Standard-Reinforcement Learning Model (SRLM) section explains a thorough<br />
standard model <strong>of</strong> RL agent, and some basic characteristics <strong>of</strong> RL agent.<br />
The Image Segmentation by Autonomous Agents in MAS section introduces the recent<br />
segmentation methods, which employed agents to segment images. The Image<br />
Segmentation by Reinforcement Learning Model discusses about the existing<br />
segmentation literature, which employed RL model.<br />
Finally, the summarizing <strong>of</strong> this chapter mentions in the last section with a discussion<br />
about image segmentation methods using agent or RL agent and methods that do not<br />
use agent.<br />
2.2 Three-Dimensional Medical Imaging and Skull Anatomy<br />
In this section, a brief explanation <strong>of</strong> two commonly used modalities for medical<br />
imaging provides. Furthermore, the skull anatomies will explain to facilitate the<br />
correlation with the anatomical images.<br />
In general, CT is the modality <strong>of</strong> choice for bony details, and MRI is superior to CT for<br />
s<strong>of</strong>t tissue details. CT is superior to MRI for detecting classification, and it is the study<br />
<strong>of</strong> choice for evaluation <strong>of</strong> foreign bodies. Moreover, MRI has not known biological<br />
side effect. Each <strong>of</strong> the modalities images the body through using a form <strong>of</strong> energy to<br />
8
map the internal structures. X-rays and radio waves are the energies used for CT, and<br />
MRI, respectively (Valvassori, 1995).<br />
2.2.1 Computed Tomography Images<br />
CT is one <strong>of</strong> the most powerful diagnostic tools available in medicine nowadays.<br />
Advantages <strong>of</strong> the CT system are in its ability to distinguish smaller differences in x-ray<br />
attenuation between tissues, and to provide extremely high spatial resolution data. These<br />
advantages have resulted in considerable improvement in image quality. One <strong>of</strong> the<br />
major advances <strong>of</strong> the CT has been the ability to generate a comprehensive scan <strong>of</strong> a<br />
region or even entire body (Walter, 2007).<br />
The CT scanner, as shown in Figure 2.1, consists <strong>of</strong> a gantry that rotates around the<br />
patient. The x-ray tube and detectors mount on the gantry, along with a host <strong>of</strong><br />
additional electronics and equipment. A table for the patient moves through a<br />
cylindrical opening in the middle <strong>of</strong> the gantry. The gantry rotates at a high rate <strong>of</strong><br />
speed around the patient, who positions within the bore <strong>of</strong> the CT scanner. Data record<br />
as the patient moves through the x-ray beam, creating what call ‘projections’. For each<br />
individual image, multiple projections at various angles acquire. The data collected<br />
from these multiple projections then transfer to a computer, which uses a mathematical<br />
algorithm to reconstruct the CT image and store it in digital format (Walter, 2007).<br />
The CT image is a digital file consisting <strong>of</strong> the pixels. The CT system calculates the<br />
amount <strong>of</strong> the x-ray attenuation for each pixel. These attenuation values are<br />
standardized and called ‘Hounsfield’ or ‘CT numbers’. Once the CT numbers calculate<br />
they typically map to a shape <strong>of</strong> gray to create an image. The black areas represent<br />
9
egions with lower CT attenuation (like air) while the white areas represent regions with<br />
higher CT numbers (like bone), a sample <strong>of</strong> the CT images has shown in Figure 2.2.<br />
Figure 2.1: CT scanner (Jeri, 2008)<br />
Figure 2.2: Human Head CT Slices at Axial View (Obaidellah, 2006)<br />
2.2.2 Magnetic Resonance Imaging<br />
MRI provides an extremely high level <strong>of</strong> the detail and concerning the anatomy and<br />
pathology in vivo with the radio waves and a strong magnetic field. The Magnetic<br />
10
Resonance Image Scanner is a medical device used to generate images <strong>of</strong> the s<strong>of</strong>t<br />
tissues for the diagnosis <strong>of</strong> illnesses.<br />
In MRI, the patient is placed within the bore <strong>of</strong> a large magnet. This magnet creates the<br />
external magnetic field; a MRI scanner is shown in Figure 2.3. The signal used to create<br />
an image is created by moving the group <strong>of</strong> spins out <strong>of</strong> alignment with external<br />
magnetic field (the z-axis). An image is created by measuring the signal or echo <strong>of</strong> the<br />
protons processing about external magnetic field after the application <strong>of</strong> a<br />
radi<strong>of</strong>requency (RF) pulse sequence. A specific RF pulse is broadcast into the body and<br />
then moves the net magnetization vector so that it is processing in x-y plane. Motion<br />
and flowing blood can detect on MR images. Thus, flowing blood can map as bright or<br />
dark depending upon the pulse sequence used to obtain the images.<br />
Figure 2.3: MRI Scanner (Garrobo, 2006)<br />
The spatial information in MR images obtain by applying smaller external magnetic<br />
gradients across the patient. These gradients determine what level or slice <strong>of</strong> tissue is to<br />
be imaged. Through the appropriate application <strong>of</strong> these magnetic gradients, MR images<br />
11
can be obtained in any plane throughout the body: coronal, sagittal, transaxial, or<br />
oblique (Walter, 2007).<br />
2.2.1 Skull Anatomy<br />
The bones <strong>of</strong> the skull protect the brain and guard the entrances to the digestive and<br />
respiratory systems. The skull contains the twenty-two bones; 8 from the cranium, or<br />
braincase, and 14 are associated with the face (Marieb, 2000). Figure 2.4 shows the<br />
anterior and lateral view <strong>of</strong> the skull.<br />
Figure 2.4: The Lateral (left) and the Anterior (right) View <strong>of</strong> the Skull (Gray, 1918)<br />
2.2.1.1 Cranial Bones<br />
The cranium consists <strong>of</strong> the 8 cranial bones: the occipital bone, frontal bone, sphenoid,<br />
ethmoid, and the paired parietal and temporal bones. Together, the cranial bones<br />
enclose the cranial cavity, a chamber that supports the brain. The outer surface <strong>of</strong> these<br />
bones provides an extensive area for the attachment <strong>of</strong> the muscles that move the eyes,<br />
jaws, and head. Figure 2.5 represents the cranial bones.<br />
12
2.2.1.2 Facial Bones<br />
Figure 2.5: Cranial Bones <strong>of</strong> the Skull (Martini, 2004)<br />
Facial bones protect and support the entrances to the digestive and respiratory tracts.<br />
The superficial facial bones (the paired maxillary, lacrimal, nasal, and zygomatic bones<br />
and the single mandible) provide areas for the attachment <strong>of</strong> the muscles that control<br />
facial expressions and assist in manipulating food. The deeper facial bones (the palatine<br />
bone, inferior nasal conchae, and vomer) help separate the oral and nasal cavities,<br />
increase the surface area <strong>of</strong> the nasal cavities (Martini, 2004). The two maxillae fuse to<br />
form the upper jaw. All facial bones except the mandible join the maxillae; thus the<br />
maxillae are the main, or ‘keystone’, bone <strong>of</strong> the face (Marieb, 2000). The mandible,<br />
forming the skeleton <strong>of</strong> the chin, is one <strong>of</strong> the largest bones <strong>of</strong> the skull and the only<br />
moveable one (Hiatt, 1982). Figure 2.6 (a) and (b) show the maxillae and the mandible<br />
bone <strong>of</strong> facial face respectively.<br />
13
(a)<br />
(b)<br />
Figure 2.6: (a) The Maxillae Bone (b) The Mandible Bone (Gray, 1918)<br />
2.2.1.3 Anatomical Structure that Used as the Case Study<br />
In this section, we present some sketches <strong>of</strong> the human skull in order to show the case<br />
studies <strong>of</strong> this research. Figure 2.7 shows the top side <strong>of</strong> the skull, through this sketch<br />
the brain is visible, but in CT images the s<strong>of</strong>t tissue is not very clear meanwhile the<br />
Figure 2.8 (a) and (b) show one slice <strong>of</strong> this part <strong>of</strong> the skull, 1-3 and 1-6 section<br />
respectively. Figure 2.9 depicts a wide spectrum <strong>of</strong> the skull, where the Figures 2.10(a-<br />
d) shows the CT images <strong>of</strong> the specific section <strong>of</strong> Figure 2.9. In the Figure 2.11, the<br />
14
inferior part <strong>of</strong> the skull is shown; also Figure 2.12(a-b) shows the CT image <strong>of</strong> the<br />
particular section <strong>of</strong> Figure 2.11.<br />
Figure 2.7: The Top Side <strong>of</strong> the Skull (Walter, 2007)<br />
(a) (b)<br />
Figure 2.8: (a) CT Image <strong>of</strong> 1-3 section (b) CT image <strong>of</strong> 1-6 section (Obaidellah, 2006)<br />
Figure 2.9: The Middle Part <strong>of</strong> the Skull (Walter, 2007)<br />
15
(a) (b)<br />
(c) (d)<br />
Figure 2.10: (a) CT Image <strong>of</strong> 1-10 Section (b) CT Image <strong>of</strong> 1-12 Section (c) CT Image<br />
<strong>of</strong> 1-15 Section (d) CT Image <strong>of</strong> 1-20 Section (Obaidellah, 2006)<br />
Figure 2.11: The Inferior Part <strong>of</strong> the Skull (Walter, 2007)<br />
16
Figure 2.12: (a) CT Image <strong>of</strong> 1-22 Section (b) CT Image <strong>of</strong> 1-23 Section (Obaidellah,<br />
2006)<br />
2.3 Segmentation Methods<br />
The objective <strong>of</strong> the segmentation method is to identify the disjoint objects into image,<br />
which have certain uniformity. A formal definition <strong>of</strong> the segmentation in (Awcock,<br />
1995) is presented in following.<br />
Consider an image array <strong>of</strong> m columns by n rows, Figure 2.13(a). Let � denote this<br />
complete array <strong>of</strong> the image pixels, the set <strong>of</strong> the pairs {i,j} where i=0,1,2,…,(m-1) and<br />
j=0,1,2,…,(n-1).<br />
(n-1)<br />
(a) (b)<br />
Rb<br />
�<br />
(m-1)<br />
(a) (b)<br />
Figure 2.13: Uniformity Predicate <strong>of</strong> the Segmentation (Awcock 1995)<br />
R1<br />
R2<br />
Rg<br />
R3<br />
R4<br />
Rh<br />
Rt<br />
17
Let Ra be a non-empty subset <strong>of</strong> R consisting <strong>of</strong> the sequential image pixels. A<br />
uniformity predict, P(Ra), is a logical statement which assigns the value True or False to<br />
Ra, depending on the properties related to the intensity matrix f(i,j) for the point <strong>of</strong> Ra.<br />
A segmentation <strong>of</strong> the array <strong>of</strong> R, Figure 2.13(b), is a partition <strong>of</strong> R into disjoint non-<br />
empty subsets R1, R2, R3… Rt and can be defined mathematically as:<br />
I. � Rg � R for g=1,2,3,…,t.<br />
II. R g<br />
is a connected region; g=1,2,3,…,t.<br />
III. Rg� Rh��<br />
for all g and h; g � h.<br />
IV. P ( Rg)<br />
= True for g=1,2,3,…,t.<br />
V. P( Rg�<br />
Rh)<br />
= False for g � h.<br />
There are a number <strong>of</strong> the image features, which can be used in image segmentation<br />
methods, such as gray-level values, color parameters, boundary and range information,<br />
texture and motion. These features can be determined how much the segmentation<br />
methods have achieved to the uniformity criteria.<br />
2.3.1 Several Classifications <strong>of</strong> Segmentation Methods<br />
In this section, it is presented some main approaches based on the opinion <strong>of</strong><br />
(Bovenkamp, 2004; Withey, 2006; Awcock, 1995; Umbaugh, 1998; Kagawa, 1999;<br />
Pham, 2000) other researchers. The next section, a brief description <strong>of</strong> main<br />
segmentation methods that are classified by the mentioned researchers brings out.<br />
18
Awcock and Thomas (Awcock, 1995) have dichotomized the segmentation methods<br />
into:<br />
(a) Pixel-based or local or discontinuity methods;<br />
(b) Region-based or global or similarity approaches.<br />
These approaches are complementary; in practice the results are not the same in each<br />
case. The first method (a) detects and enhances the edges within element and links these<br />
to construct an object. They only used point-wise or nearest-neighbor local information<br />
and no information is taken <strong>of</strong> the general properties <strong>of</strong> the whole region. The second<br />
one seeks to create regions directly by collecting the common features <strong>of</strong> a group <strong>of</strong> the<br />
pixels into areas or region <strong>of</strong> the uniformity.<br />
Some methods mention for the approach (a) such as Edge Detection, and Boundary<br />
Detection. The examples <strong>of</strong> the approach (b) are Region merging and splitting, and<br />
Thresholding.<br />
Umbaugh in his book (Umbaugh, 1998) mentioned the segmentation techniques into<br />
three classification; region growing and shrinking, clustering methods, and boundary<br />
detection. The one sample <strong>of</strong> region growing is, the segmentation starts with smallest<br />
level and only merges, with no region splitting. In region shrinking, the entire image<br />
consider as initial region, and then follows an algorithm, the image is only spitted. The<br />
clustering techniques are image segmentation methods; based on some measure <strong>of</strong> the<br />
similarity in image, the image is cluster to some groups. The boundary detection is<br />
carried out by seeking the boundary between objects, this method is usually commenced<br />
by marking pixel may be a part <strong>of</strong> an edge.<br />
19
Pham et al., in their survey (Pham, 2000) divided the common segmentation methods<br />
into eight categories:<br />
(a) Thresholding approaches;<br />
(b) Region growing approaches;<br />
(c) Classifiers;<br />
(d) Clustering approaches;<br />
(e) Markov random field models;<br />
(f) Artificial neural networks;<br />
(g) Deformable models, and<br />
(h) Atlas-guided approaches.<br />
However, the thresholding, classifier, clustering, and Markov random field approaches<br />
can consider as independent methods <strong>of</strong> the pixel classification.<br />
Kagawa et al. in (Kagawa, 1999), considered the segmentation methods to the four main<br />
approaches:<br />
(a) Edge detection,<br />
(b) Region growing,<br />
(c) Method <strong>of</strong> clustering,<br />
(d) Statistical methods.<br />
Edge detection is a method, which detects the location <strong>of</strong> the feature pixel, is changed<br />
precipitously. Region growing is an approach that it makes use <strong>of</strong> share common feature<br />
that the regions have. Methods <strong>of</strong> clustering are the approach that uses the feature space<br />
<strong>of</strong> the color information and so on. Statistical methods are the approach that it makes<br />
use <strong>of</strong> the statistical and/or structural texture that the image has.<br />
20
Bovenkamp et al., in (Bovenkamp, 2004) discussed the basic image interpretation<br />
strategy. They distinguished them to three basic strategies:<br />
(a) Bottom-up,<br />
(b) Top-down,<br />
(c) Hybrid.<br />
The bottom-up strategy does not have any information <strong>of</strong> the object within image and<br />
based on pixels in the image can achieve the segmentation goal. The top-down strategy<br />
is to assume an object to be in the image and then to seek for it, for example using a<br />
deformable model such as snakes and active shape model. For conveying limitation<br />
which exists in the above two strategies, the hybrid strategy is posed. This strategy<br />
combined both bottom-up and top-down process, a common implementation <strong>of</strong> the<br />
hybrid strategy is to create a feedback loop from the symbols back to image, for<br />
example locally re-segment the image given evidence from the image data and<br />
reasoning process.<br />
Withey in his thesis (Withey, 2006) classified the segmentation methods into three<br />
generations; the first generation has carried out the most primary and lowest level <strong>of</strong> the<br />
processing, the image models, optimization methods, and uncertainty models are used<br />
in the second generation, and the third generation algorithms have capability <strong>of</strong> the<br />
incorporating knowledge.<br />
In the first generation, thresholding, region growing, region split/merge, edge detection,<br />
and edge tracing methods are noticeable. The statistical methods, C-mean clustering,<br />
Fuzzy connectedness, deformable models, watershed algorithm, neural networks, and<br />
multi-resolution methods are citable for the second-generation algorithms. The method<br />
combinations and knowledge-based segmentation is generated the third.<br />
21
2.3.2 Brief Description <strong>of</strong> Main Segmentation Methods<br />
2.3.2.1 Thresholding<br />
The thresholding technique is the primitive technique in image segmentation. This<br />
method produces regions <strong>of</strong> the uniformity within an image based on some threshold<br />
criterion, T. the function T defines:<br />
T = T{x, y, A(x, y), f(x, y)},<br />
Where f(x, y) is the intensity <strong>of</strong> the pixel at (x, y), and A(x, y) denotes some local<br />
property in the neighborhood <strong>of</strong> this pixel.<br />
A threshold image g(x, y) defines:<br />
g(x, y) =<br />
�1<br />
�<br />
�0<br />
if f ( x,<br />
y)<br />
� T �<br />
� .<br />
if f ( x,<br />
y)<br />
� T �<br />
The thresholding technique can identify as:<br />
(a) Global threshold: T = T{f(x, y)},<br />
Where T is depended only on the intensity <strong>of</strong> the pixel at x, y.<br />
(b) Local threshold: T = T{A(x, y), f(x, y)},<br />
Where T is depended on a neighborhood property <strong>of</strong> the pixel as well as its<br />
intensity.<br />
(c) Dynamic threshold: T = T{x, y, A(x, y), f(x, y)},<br />
Where T is depended on the pixel coordinates, and the other two criteria<br />
Selection <strong>of</strong> the value <strong>of</strong> the threshold, T, is critical issue. It is common to study<br />
histogram in order to find the appropriate threshold. One variation on the simple<br />
threshold is interval threshold operation. A binary image is produced where all gray-<br />
level values falling between two threshold values T1 and T2. However, a complex image<br />
exhibits the more gray-level threshold in its histogram, for this kind, multiple<br />
22
thresholding can be used to reduce the number <strong>of</strong> the gray-level values in the image<br />
(Awcock, 1995).<br />
2.3.2.2 Region Growing<br />
Region growing algorithms based on the growth <strong>of</strong> a region whenever its interior is<br />
homogeneous according to certain features as intensity, color or texture. The<br />
implemented algorithm follows the strategy <strong>of</strong> a typical Region Growing: it is based on<br />
the growth <strong>of</strong> a region by adding similar neighbors. Region Growing is one <strong>of</strong> the<br />
simplest and most popular algorithms for region-based segmentation. The most<br />
traditional implementation starts by choosing a starting point called seed pixel. Then,<br />
the region grows by adding similar neighboring pixels according to a certain<br />
homogeneity criterion, increasing gradually the size <strong>of</strong> the region. Therefore, the<br />
homogeneity criterion has the function <strong>of</strong> deciding whether a pixel belongs to the<br />
growing region or not. The decision <strong>of</strong> merging generally based only on the contrast<br />
between the evaluated pixel and the region. However, it is not easy to decide when this<br />
difference is small (or large) enough to take a decision.<br />
The Split-and-Merge algorithm is related to the region growing; a typical split and<br />
merge techniques consist <strong>of</strong> two basic steps. First, the whole image is considered as one<br />
region. If this region does not satisfy a homogeneity criterion the region is split into four<br />
quadrants (sub-regions) and each quadrant is tested in the same way; this process is<br />
recursively repeated until every square region created in this way contains<br />
homogeneous pixels. Next, in the second step, all adjacent regions with similar<br />
attributes may be merged following other (or the same) criteria. The criterion <strong>of</strong> the<br />
homogeneity is generally based on the analysis <strong>of</strong> the chromatic characteristics <strong>of</strong> the<br />
region. A region with small standard deviation in the color <strong>of</strong> its members (pixels) is<br />
23
considered homogeneous. The integration <strong>of</strong> the edge information allows adding to this<br />
criterion another term to take into account. So, a region is considered homogeneous<br />
when is very free <strong>of</strong> the contours (Chen, 1980).<br />
2.3.2.3 Edge Detection<br />
The edge detection methods attempt to sketch the object in image by boundary instead<br />
<strong>of</strong> volume <strong>of</strong> it. The edge detection is not a pure segmentation methods, it is used by the<br />
other methods for supplementing.<br />
2.3.2.4 Classifiers<br />
The classifier methods are pattern recognition techniques, which classify object into one<br />
<strong>of</strong> several categories based on feature space. A feature space is the range space <strong>of</strong> any<br />
function <strong>of</strong> the image, with the most common feature space is the image intensities.<br />
The classifiers are known as supervised approach; those divide to distribution free or<br />
statistical. Distribution free methods do not require knowledge <strong>of</strong> any priori probability<br />
distribution functions and reasoning also heuristics are the basis <strong>of</strong> these. Statistical<br />
techniques are based on probability distribution models.<br />
Suppose there are K different objects or pattern classes S1, S2, …, Sk, …, SK. each class is<br />
characterized by Mk prototype, which have N � 1 feature vectors<br />
(k )<br />
y m , m=1, …, Mk. A<br />
simple classifier is k-nearest neighbor classifier, for classifying the image to Si, if<br />
among a total <strong>of</strong> k nearest prototype neighbors, the maximum number <strong>of</strong> the neighbors<br />
belong to class Si. In statistical classifiers techniques it is assumed the different object<br />
classes and feature vector have a probability density. Let P(Sk) be a priori probability <strong>of</strong><br />
24
the occurrence <strong>of</strong> the class Sk and p(x) be the probability density function <strong>of</strong> the random<br />
feature vector observed as x. The Bayes’ minimum-risk classifier is a kind <strong>of</strong> the<br />
statistical classifiers that its objective is to minimize the average loss or risk in assigning<br />
x to a wrong class.<br />
There are also some sequential classification techniques such as sequential probability<br />
ratio test (SPRT), where decision can be made initially using fewer than N features and<br />
refined as more features are acquired sequentially (Jain, 1989; Pham, 2000).<br />
2.3.2.5 Clustering<br />
Clustering methods carry out the same function as classifier methods without the use <strong>of</strong><br />
the training data. In addition, they are termed unsupervised methods. A cluster is a set<br />
<strong>of</strong> the feature space, which their local density is large in comparison with the density <strong>of</strong><br />
the feature point in the neighbors. In order to tackle the lack <strong>of</strong> the training data, as<br />
shown in Figure 2.14, clustering methods iterate between segmenting the image and<br />
characterizing the properties <strong>of</strong> the each class. In a sense, clustering methods train<br />
themselves using the available data.<br />
Input<br />
data<br />
Partition<br />
Test and<br />
Merge<br />
Split<br />
Conve<br />
rgence<br />
Figure 2.14: A Clustering Approach (Jain, 1989)<br />
25
Three commonly used clustering algorithms are the K-means, the fuzzy c-means<br />
algorithm, and the expectation-maximization (EM) algorithm. The K-means clustering<br />
algorithm clusters data by iteratively computing a mean intensity for each class and<br />
segmenting the image by classifying each pixel in the class with the closest mean. The<br />
fuzzy c-means algorithm generalizes <strong>of</strong> K-means allowing for s<strong>of</strong>t segmentations based<br />
on fuzzy set theory. The EM algorithm applies the same clustering principles with the<br />
underlying assumption that the data follows a Gaussian mixture model. It iterates<br />
between computing the posterior probabilities and computing maximum likelihood<br />
estimates <strong>of</strong> the means, co-variances, and mixing coefficients <strong>of</strong> the mixture model<br />
(Jain, 1989; Pham, 2000).<br />
2.3.2.5 Deformable Models<br />
Deformable models are known techniques for boundary extraction and segmentation <strong>of</strong><br />
the medical images. One <strong>of</strong> the earlier active contours is snake. It formulated as a<br />
parametric model, they consist <strong>of</strong> a curve, which can dynamically match to object<br />
shapes in response to internal and external forces. To describe an object boundary in an<br />
image, a closed curve or surface must first place near the desired boundary and then<br />
allowed to undergo an iterative relaxation process. Internal forces are computed from<br />
within the curve or surface to keep it smooth throughout the deformation. External<br />
forces are usually derived from the image to drive the curve or surface towards the<br />
desired feature <strong>of</strong> the interest (Giraldi, 2006; Pham, 2000).<br />
2.3.2.6 Neural Network<br />
An artificial neural network is a set <strong>of</strong> the parallel elements called neurons that emulate<br />
a biological neural learning system. Each neuron can perform elementary computation<br />
26
such that weights assign to the connections is achieved to learning (Pham, 2000). The<br />
neural network acts as classifiers where a set <strong>of</strong> the features is determined for each<br />
image pixel and presented as input to the neural networks (Withey, 2006).<br />
2.3.3 Comparison <strong>of</strong> the Segmentation Methods<br />
It is useful to compare all segmentation methods in the previous section. In Table 2.1,<br />
the advantages and disadvantages <strong>of</strong> all methods list by reviewing <strong>of</strong> these literatures<br />
(Pham, 2000; Freixenet, 2002; Kirbas, 2003; Withey, 2006).<br />
Table 2.1: The Comparison among the Segmentation Methods (<strong>Chitsaz</strong>, 2008)<br />
Methods Advantages Disadvantages<br />
Thresholding Simple implementation.<br />
Region Growing<br />
Good for small and simple<br />
structure, easy to detect the<br />
global structure <strong>of</strong> the image.<br />
Edge detection Useful for boundary detection<br />
Classifiers<br />
Clustering<br />
Deformable models<br />
Neural networks<br />
Can apply to multiple-channel<br />
image.<br />
Do not need training data, fast<br />
computation, robustness to<br />
intensity inhomogeneities<br />
Robustness to noise and spurious<br />
edges,<br />
Parallel, easily incorporate<br />
spatial information to<br />
classification procedures, ability<br />
to learn<br />
Sensitive to noisy image, can<br />
not apply to multiple-channel<br />
image.<br />
Manual interaction is needed,<br />
Over-segmentation, sensitive to<br />
noisy image<br />
It is difficult to detect edge in<br />
complex image, sensitive to<br />
noisy image.<br />
Requirement to manual<br />
interaction, sensitive to<br />
intensity inhomogeneities<br />
Dependency to the number <strong>of</strong><br />
the clusters and features,<br />
dependency to initial<br />
segmentation<br />
Requirement to manual<br />
interaction, dependency to<br />
parameter value<br />
require to train every time a<br />
new feature is introduced the<br />
network, difficult to debug the<br />
performance <strong>of</strong> the network<br />
27
2.4 Agent and Multi-Agent System<br />
Although many people use the term <strong>of</strong> the agent and multi-agent who are working in<br />
closely related areas. There are no widely accepted definitions <strong>of</strong> these terms, and the<br />
definitions are still open challenge. However, in many literatures, some attributes <strong>of</strong> the<br />
agent are similar.<br />
The following properties are common for a hardware or s<strong>of</strong>tware-based computer<br />
system agent in a weak notation (Wooldridge, 1997; Kagawa, 1999):<br />
� autonomy: agents accomplish without the direct interposition <strong>of</strong> the humans or<br />
others, and having control over their actions and internal state;<br />
� social ability: agents cooperate with other agents (and may be humans)<br />
� reactivity: realizing their environment, and responding to changes that occur in it;<br />
� Pro-activeness: having ability to exhibit goal-directed behavior by taking the<br />
initiative;<br />
� Robustness: should be prepared to learn and to recover from failure.<br />
The other properties <strong>of</strong> the agent, which related to its context discussed. For example,<br />
mobility is the ability <strong>of</strong> an agent to move around an electronic network, such as<br />
moving agent from a computer to another. Veracity is the assumption that an agent will<br />
not knowingly communicate false information. Benevolence is the assumption that<br />
agents always attempt to do what is asked <strong>of</strong> it. Finally, rationality is the assumption<br />
that an agent will strongly act in order to achieve its goals (Wooldridge, 1995).<br />
28
The internal structure <strong>of</strong> an agent may consist <strong>of</strong> several units as shown in Figure 2.15.<br />
In general, agents have the following units (Rares, 1999):<br />
� Input units, for receiving incoming data;<br />
� Output units, for delivering agent’s results;<br />
� Planning units, for determining the processing strategy;<br />
� Control units, which put into practice the plan elaborated by the planning units,<br />
and coordinate the execution;<br />
� Evaluation units, for checking the quality <strong>of</strong> the processing operations;<br />
� Learning units, for knowledge acquisition and adaptive behavior.<br />
Agent<br />
Planning Unit<br />
Figure 2.15: The Internal Structure <strong>of</strong> a Typical Agent (<strong>Chitsaz</strong>, 2008)<br />
These units are varied by nature <strong>of</strong> each problem, and probability in some cases the<br />
other units has been added or decreased.<br />
Control Unit<br />
Input Evaluation<br />
Unit<br />
Output<br />
Learning Unit<br />
MAS is overall a system with several entities which they have some mutual behavior<br />
like cooperation, coordination and negotiation. In (Jennings, 1998) MAS is defined as<br />
they are ideally suited to representing problems that have multiple problem solving<br />
29
methods, multiple perspectives and/or multiple problem solving entities. Such systems<br />
have the traditional advantages <strong>of</strong> the distributed and concurrent problem solving, but<br />
have the additional advantage <strong>of</strong> the sophisticated patterns <strong>of</strong> interactions”. Therefore,<br />
in MAS every agent is an ingredient in a massive system, and they just have knowledge<br />
about their environs so by cooperating, coordinating and negotiating they can able to<br />
achieve goal quickly.<br />
The MAS is broadly used in variety fields, such as robotic, etc. we intend to use this<br />
system as the skeleton <strong>of</strong> our framework, because <strong>of</strong> following reasons (Crevier, 1997):<br />
� Ease <strong>of</strong> construction and maintenance. It is easier to set up and repair a<br />
collection <strong>of</strong> independent modules than a single huge program<br />
� Ability to benefit from parallel architectures.<br />
� Focusing ability. Not all knowledge requires for all tasks. Modularizing<br />
provides the ability to focus the system’s efforts in the most productive manner.<br />
� Heterogeneous problem solving. The methods best appropriate to one part <strong>of</strong> a<br />
problem may not be best for working on another part.<br />
� Reliability. If one agent provides a wrong answer or clue, the consensus <strong>of</strong> other<br />
agents may yet provide the true answer.<br />
2.5 Standard Reinforcement Learning Model<br />
Learning to act in ways that are rewarded is a sign <strong>of</strong> intelligence (Watkins, 1989). For<br />
example, it is natural to train elephant in circus by rewarding it when the elephant acts<br />
correctly in reaction <strong>of</strong> a command. That animal can learn to obtain more rewards than<br />
punishment, and this aspect <strong>of</strong> animal intelligence has been studied extensively in<br />
experimental psychology (Watkins, 1989).<br />
30
In the standard RL model, an agent is interacted to its environment via perception and<br />
action, as shown in Figure 2.16. On each step <strong>of</strong> interaction the agent receives as input,<br />
i, the current state, s, <strong>of</strong> the environment; the agent then chooses an action, a, to<br />
generate an output. The action changes the state <strong>of</strong> the environment and the value <strong>of</strong> this<br />
state transition which are undertaken to the agent through a reinforcement signal<br />
(reward/punishment), r. The agent's behavior, B, should choose actions that tend to<br />
increase the overall sum <strong>of</strong> the rewards values. Agent can learn to do this over time by<br />
systematic trial and error (Kaelbling, 1996). The RL agent does not have any knowledge<br />
about environment; it is just trained by obtaining rewards or punishment based on its<br />
action from environment. It is important that the agent gather useful experience about<br />
the possible system states, actions, rewards and punishment actively to behave<br />
optimally.<br />
Figure 2.16: The Standard RL Model (Kaelbling, 1996)<br />
As a whole, the model consists <strong>of</strong>:<br />
� a discrete set <strong>of</strong> the environment states, S;<br />
� a discrete set <strong>of</strong> the agent actions in turn <strong>of</strong> the states, A;<br />
� a set <strong>of</strong> the scalar reward/punishment for each associated action or a sequence <strong>of</strong><br />
the actions ; typically [0,1], or the real numbers.<br />
31
Q-learning (Watkins, 1989) is a recent form <strong>of</strong> the RL algorithm. Q-learning algorithm<br />
works by estimating the values <strong>of</strong> the state-action pairs. The value Q(s,a) is defined to<br />
be the expected sum <strong>of</strong> the future pay<strong>of</strong>fs r obtained by taking action a from state s.<br />
Once these values have learned, the optimal action from any state is the one with the<br />
highest Q-value. After being initialized to arbitrary numbers, Q-values are estimated<br />
based on the experience as shown in Figure 2.17:<br />
1. From the current state s, select an action a. This will cause a receipt <strong>of</strong> an<br />
immediate pay<strong>of</strong>f r, and arrival at a next state s'.<br />
2. Update Q(s,a) based upon this experience as follows:<br />
Q(<br />
s,<br />
a)<br />
� ( 1��<br />
) Q(<br />
s,<br />
a)<br />
��[<br />
r ��<br />
maxQ(<br />
s�,<br />
a�)]<br />
e 2.17: Pseudocode for Q-Learning Algorithm (Watkins, 1989)<br />
Figur<br />
This algorithm is guaranteed to converge to the correct Q-values with the probability<br />
one if the environment is stationary and depends on the current state and the action<br />
taken in it; a lookup table (Q-Matrix) is used to store the Q-values, every state-action<br />
pair continues to be visited, and the learning rate is decreased appropriately over time.<br />
This exploration strategy does not specify which action to select at each step. In<br />
practice, a method for choosing action is usually chosen that will ensure sufficient<br />
exploration while still actions with higher value estimates.<br />
a�<br />
where � is the learning rate and 0 < � < 1 is the discount factor<br />
3. Go to step 1.<br />
32
2.6 Image Segmentation by Autonomous Agents in Multi-Agent System<br />
There have been many researches to segment an image using agent approaches. In this<br />
section, the previous works in image segmentation using the MAS will be considered.<br />
The following will first look at previous attempt to image segmentation via MAS. The<br />
literatures mentioned in this section are not limited to medical image segmentation. The<br />
objective is to present all models, which are related to image segmentation, and multi-<br />
agent models.<br />
Different multi-agent approaches have presented lately for the segmentation <strong>of</strong> the<br />
image or edge detection on image. The first category is suitable for non-medical images.<br />
In addition, the second category is suitable for medical images.<br />
2.6.1 Kagawa et al method<br />
The method <strong>of</strong> Kagawa et al. (Kagawa, 1999) has been presented two basic phases;<br />
region segmentation phase and region integrating phase, as shown Figure 2.18. The<br />
agents are distributed in the image and calculate several features <strong>of</strong> every pixel in an<br />
image. Subsequently, they move onto the pixel, which has the most similar features.<br />
Following the inactive agents are modified in order to activate them again. When an<br />
agent cannot find any pixel whose similarities are higher than a defined threshold, it has<br />
been vanished. After the phase <strong>of</strong> the region segmentation, the segmented regions are<br />
integrated into larger region, which are parts <strong>of</strong> the objects on the given image. The<br />
result for this proposed approach is based on landscape images and the color feature on<br />
the image.<br />
33
Figure 2.18: The Proposed Method <strong>of</strong> Kagawa (Kagawa, 1999)<br />
The proposed method by Kagawa et al. requires less amount <strong>of</strong> the calculation. In<br />
addition, it could be applied to a wide variety <strong>of</strong> the natural images. Nevertheless, this<br />
method can improve by utilizing not only the color plane, but also the frequency space<br />
as the features agents.<br />
2.6.2 Wang and Yuan method<br />
Wang and Yuan (Wang Y., 2000; Wang Y., 2002a; Wang Y., 2002b) proposed face<br />
detection model by evolutionary agents. Several agents are uniformly anchored in the<br />
each pixel on 2D image environment to seek the skin-like pixel. The evolutionary agent<br />
was defined as Agent= < p, d, a, f, fml, Diff, Rep, Die >. p denotes the position <strong>of</strong> an<br />
agent in image. d represents its current diffusion direction. a stands for the age <strong>of</strong> an<br />
agent. f symbolizes its fitness, which indicates the adaptability <strong>of</strong> an agent and can be<br />
computed using the number <strong>of</strong> steps the agent takes to find a skin-like point. fml<br />
represents the family index. The five states, which defined, are representing the internal<br />
34
state <strong>of</strong> an agent while the Diff (diffusion), Rep (reproduction) and Die describe the<br />
behavior <strong>of</strong> the agent. The agents have different behaviors; such as self-production,<br />
diffusion and death. The researchers first investigate the color skin <strong>of</strong> 50 images and<br />
they found the HSV (hue, saturation, and value) <strong>of</strong> the skin is in the following range:<br />
0
The system assesses its state based on all agents’ intention. The intention is represented<br />
at the knowledge that each agent experience. The information that should be included in<br />
the agent knowledge is such as kind, type, plausibility, relationship, and search area.<br />
Kind is a class <strong>of</strong> the objects to be extracted. The more small groups that object <strong>of</strong> the<br />
same kind are classified into type. The measurement that evaluates how best the agent<br />
satisfied the constraints is plausibility. The behavior <strong>of</strong> the agents based on their state is<br />
different; such as producing son agents, decaying son agents, constructing parent agent,<br />
changing agent knowledge and resolving overlap between agents. The proposed method<br />
was applied to line drawing recognition and character segmentation.<br />
The proposed method is a hierarchical multi-agent based method to extract object from<br />
a given image. It can obtain the desired letter only with the knowledge on them. Then,<br />
this method does not need direct control on agents. However, the computation time<br />
should improve in future.<br />
Figure 2.19: The Proposed Method <strong>of</strong> Gyothen (Gyohten, 2000)<br />
36
2.6.4 Guillaud et al. method<br />
Guillaud et al. (Guillaud, 2000) presented MAS for ring detection on fish otoliths. They<br />
used two types <strong>of</strong> agent, the dark and light agent. Each agent should check pixels<br />
around it in a circular neighborhood. If it finds that its neighborhood satisfies the<br />
condition to be a region, the central pixel will be marked and new agents will generate<br />
to grow the region. The agents can move in the gray scale image <strong>of</strong> the otolith, every<br />
dark (light) agent try to find darker (lighter) pixel. They save the path. When the agent<br />
has run over a loop, it can validate the path as a ring. In addition, the researchers have<br />
added high-level information about the shape <strong>of</strong> the contour to improve the detection;<br />
the procedure is shown in Figure 2.20.<br />
The proposed method has an acceptable result to find the continuous ring <strong>of</strong> the otolith<br />
image. Besides, detecting nucleus position is automated. However, the tuning <strong>of</strong> the<br />
agents parameters is not easy.<br />
Figure 2.20: The Procedure <strong>of</strong> the Proposed Method by Guillaud et al. (Guillaud ,<br />
2000).<br />
37
2.6.5 Rodin et al. method<br />
Rodin et al. (Rodin, 2004) proposed MAS for biological image segmentation. There are<br />
two types <strong>of</strong> the agents; lightening agent and darkening agent. In the system, each agent<br />
can sense the environment. Then, based on its type, it marks the current located pixel.<br />
After that, each agent records its path until it has rotated again. In following lifetime <strong>of</strong><br />
the agent, when an agent recognizes a path, the first discoverer agent kills the other<br />
agents that located in the path; this procedure take place to avoid validation <strong>of</strong> the same<br />
ring. At the end, the agent draws a polygon corresponding to the path it has just passed.<br />
The finite state machine <strong>of</strong> describing the behavior <strong>of</strong> the agent is shown in the Figure<br />
2.21. The color <strong>of</strong> this polygon is depended on type <strong>of</strong> the discoverer agent. In Figure<br />
2.22, it can been seen the proposed model which the input data is brought from<br />
environment and the behavior consist <strong>of</strong> decreasing or increasing brightness based on<br />
type <strong>of</strong> the agent (darken or lighter), rotating movement, or go forward (Go to).<br />
Figure 2.21: The MAS which Proposed by Rodin et al. (Rodin, 2004)<br />
38
Figure 2.22: The Finite State Machine that is describing an Agent’s Behavior (Rodin,<br />
2004)<br />
As a conclusion, the proposed method was automated, and can be used to different type<br />
<strong>of</strong> the images.<br />
2.6.6 Melkemi et al. method<br />
Melkemi et al. (Melkemi, 2004; Melkemi, 2005; Melkemi, 2006) proposed a model,<br />
which is a hybridization <strong>of</strong> MAS and Markov Random Field and Genetic Algorithm, the<br />
architecture is shown on Figure 2.23. The model has two types <strong>of</strong> agents, the<br />
coordinator agent and the segmentation agent. The first step is that each segmentation<br />
agent segments a part <strong>of</strong> the image by Iterated Conditional Modes procedure and the<br />
initial sub-optimal configuration, however the initial configuration is created by agents<br />
which they use K-means and a chaotic mapping. The second step is that the result <strong>of</strong><br />
every segmented agent transfers to the coordinator agent. It decides which result is<br />
better and based on the initial configuration, Genetic Algorithm, the coordinator<br />
produces some new initial configurations, and these configurations transmit to the<br />
segmentation agents. This procedure iterates until the stable situation is achieved.<br />
39
Consequently, experimental results <strong>of</strong> the proposed method are very encouraging which<br />
show the feasibility, the convergence and the robustness <strong>of</strong> the method. In addition, the<br />
method found to be much faster than traditional methods.<br />
Figure 2.23: The Architecture <strong>of</strong> the Proposed Method <strong>of</strong> Melkemi (Melkemi, 2006)<br />
From beginning <strong>of</strong> the section to this point, many literatures reviewed which uses agent<br />
for segmenting non-medical images. Following the other approaches will summarize<br />
regarding medical images.<br />
2.6.7 Spinnu et al. method<br />
Spinnu et al. (Spinu, 1996) proposed a multi-agent approach to edge detection in<br />
medical images. They have defined two basic agent types; knowledge servers (KS), and<br />
knowledge processors (KP). KS agents manage the problem elements that are<br />
represented by objects and attributes. KP agents manage the processing and reasoning<br />
methods. Any agent may get or set attribute values, create or delete object instances and<br />
modify system configuration as well as dynamically creating new agents. The major<br />
40
types <strong>of</strong> KP agent are KP noise, KP texture, KP config, KP operator, KP evaluation, and<br />
KP split. KP noise and KP texture generate maps <strong>of</strong> estimated noise and texture<br />
characteristics <strong>of</strong> the given image. KP config activates the created processing groups<br />
that will start running concurrently and cooperatively. KP operator selects the<br />
appropriate operator, for example, the Deriche operator may select for a zone affected<br />
by additive noise. KP evaluation minimizes the estimated error and inconsistency error.<br />
KP split evaluates the current result and proposes a partition <strong>of</strong> the region into sub-<br />
region. The proposed MAS is shown in Figure 2.24.<br />
Figure 2.24: The Proposed Method using MAS by Spinu (Spinu, 1996)<br />
The proposed method achieves to its goals defined. However, there are some other<br />
improvements to reach the optimal solution. For example, localization error can be<br />
41
taken in formulation error. Besides, the contrast characteristic could be used in addition<br />
<strong>of</strong> noise and texture characteristics.<br />
2.6.8 Boucher et al. method<br />
Boucher et al. (Boucher, 1996; Boucher, 1998) proposed MAS to segment image <strong>of</strong> the<br />
living cells; the model is shown in Figure 2.25. The living cell image has four different<br />
regions; nucleus, pseudopod, white halo and background. Therefore, these components<br />
determine the type <strong>of</strong> the agents. Also, the internal manager agent is used which<br />
manages the execution <strong>of</strong> the agents. The segmentation is based on region-growing<br />
approach. Every agent assesses the four neighbor pixels. Then, each pixel value that has<br />
highest evaluation pixel is labeled by a region. An evaluation function used for deciding<br />
the highest evaluation pixel. This function uses six criteria like as variance similarity,<br />
compact, gray level similarity, gradient direction similarity, cell and nucleus image<br />
thresholding. If two types <strong>of</strong> the agent label a pixel then this pixel is added to event list<br />
<strong>of</strong> the manager. The behavior <strong>of</strong> the agents is categorized to three kinds; Merging,<br />
Negotiation and Reproducing. Merging occurs when two discovered regions are the<br />
same so these regions merge and life <strong>of</strong> one <strong>of</strong> the agent is terminated. Negotiation<br />
carries out when two agents are in conflict with the one region. For example, a region<br />
has two different labels by two kind <strong>of</strong> the agent, in a situation agents negotiate about<br />
the correct labeling. Reproduction means an agent can produce a new agent. This occurs<br />
when an agent finds a new region in its environment for another kind agent so it can<br />
produce an agent, which related to new region. In another example, when segmentation<br />
sufficiency is mature, an agent can produce another agent so the new agent can access to<br />
information <strong>of</strong> the previous frame, which the terminated agent have done.<br />
42
Figure 2.25: The Proposed MAS by Boucher et al. (Boucher A., 1998)<br />
Therefore, the proposed method is adaptable to the shape and size <strong>of</strong> the living cells to<br />
distinguish them from image. In addition, this method provides richness information<br />
from images. This richness comes from the outcome <strong>of</strong> each agent in duration <strong>of</strong> its<br />
adaptability.<br />
2.6.9 Liu and Tang method<br />
Liu and Tang (Liu, 1999) proposed MAS to segment a MRI <strong>of</strong> brain. The brain has the<br />
four basic elements; like as outline, branching region, enclosing region and tumor<br />
region. For detecting each four regions, they assign some threshold range. The agent<br />
behavior is one <strong>of</strong> these four types: breeding, pixel labeling, diffusion, and decay.<br />
Breeding means when an agent is in a homogeneous segment it should be produce some<br />
new agents in neighborhood pixel. The significant difference in this paper is that the<br />
neighborhood region is determined by a sector <strong>of</strong> a circle with specified radius like<br />
Figure 2.26. Diffusion is finding new homogeneous-segment pixels by moving to<br />
43
neighborhood pixel. When an agent encounters with a new pixel from an existed<br />
homogeneous segment, this agent labels this pixel and it will become inactivated. After<br />
each agent passes its life span, it must be vanished or decayed.<br />
The proposed method has less computation time in comparison with conventional<br />
method. However, there is a problem to distribute agent over an image optimally.<br />
Figure 2.26: The Local Neighboring Region <strong>of</strong> an Agent at Location (i,j) (Liu, 1999)<br />
2.6.10 Germond et al. method<br />
Germond et al. (Germond L., 2000) proposed a framework, which composed <strong>of</strong> MAS, a<br />
deformable model, and an edge detector. The framework is shown in Figure 2.27; the<br />
image is brain MRI. There are three different types <strong>of</strong> the agent; region agent, edge<br />
agent, and scheduler. The region agents specialize for gray matter or for white matter<br />
segmentation. Edge agents specialize for the brain boundary detection. The agents are<br />
autonomous and concurrent. A shared memory is used for communicating <strong>of</strong> the agents.<br />
The MAS carries out segmentation <strong>of</strong> MRI scans. The proposed method uses seeded-<br />
region-growing method, a priori domain knowledge, and a statistical method whose<br />
parameters are acquired at run time. The aim <strong>of</strong> the deformable model is to detect the<br />
44
general boundary <strong>of</strong> the brain. The edge detector module is used for its ability to detect<br />
a precise and robust localization <strong>of</strong> the boundaries for the all edges in a given image.<br />
Figure 2.27: A Global View <strong>of</strong> the Framework and Information Flow <strong>of</strong> the Proposed<br />
Method by Germond (Germond, 2000)<br />
As a result, the proposed method has the mean quality percentage <strong>of</strong> 96%. However, the<br />
method needs considerable user interaction.<br />
2.6.11 Duchesnay et al. method<br />
Duchesnay et al. (Duchesnay, 2001; Duchesnay, 2003) proposed MAS to organize and<br />
structure the knowledge according to irregular pyramid as shown Figure 2.28, the used<br />
image is mammography. The pyramid is a stack <strong>of</strong> the graphs recursively built from<br />
base to the apex and it provides removing geometrical constraint due to the fixed<br />
structure <strong>of</strong> the neighborhood. This method has two different types <strong>of</strong> the agent; region<br />
agent, and an edge agent.<br />
45
Figure 2.28: The Conceptual Framework by Duchesnay et al. (Duchesnay, 2001)<br />
The agents can use seven behaviors; territory marking and feature extraction,<br />
exploration, merging planning, cooperation and negotiation, which are consisted<br />
decimation, reproduction and attachment. The procedure <strong>of</strong> this framework is as follow.<br />
First, the image is separated into two partitions and several agents are stayed at different<br />
parts <strong>of</strong> the image. After that, every agent seeks features around it and decides to merge<br />
with other agents based on similarity in features. In some cases the agents cannot decide<br />
due to the specified threshold is not fixed. Therefore, the agents cooperate and negotiate<br />
with the other agent <strong>of</strong> the same type in order to decide how to merge. All behaviors <strong>of</strong><br />
the agents are presented in Figure 2.29.<br />
Accordingly, the proposed method does not require substantial tuning effort. In<br />
addition, it is completely autonomous. Furthermore, it does not require priori<br />
46
information to segment images. Another interesting result is that this method can use to<br />
segment some different images as well.<br />
Figure 2.29: The Graphical Representations <strong>of</strong> the Agent’s Behaviors (Duchesnay,<br />
2003)<br />
2.6.12 Khosla and Lai method<br />
Khosla and Lai (Khosla, 2003) would like to segment a Chinese Hamster Ovarian<br />
image which recognizes the number <strong>of</strong> the cells on image. Manually the technicians<br />
after inserting some chemical thing to the cells count the cells.<br />
They proposed a framework with 3 components, as shown in Figure 2.30. In their own<br />
framework, there are two types <strong>of</strong> the agent; Intelligent Control Agent (ICA) and Image<br />
Processing Agents (IPA). IPA consists <strong>of</strong> the segmentation agent <strong>of</strong> water immersion<br />
and the mathematical morphology segmentation agent. IPA segments the image with its<br />
own algorithm. Then, the result will transfer to ICA. ICA is like human operator, it<br />
collects the accepted segmentation and unaccepted one. Then, ICA decides which <strong>of</strong><br />
them is better so the accepted segmentation puts on the result. ICA can decide by means<br />
<strong>of</strong> the neural network agent and moment invariant transformation agent.<br />
47
Water Immersion<br />
Segmentation IPA<br />
(ICA)<br />
Neural<br />
Network<br />
Agent<br />
Intelligent Control Agent<br />
Moment<br />
Invariant<br />
Transformation<br />
Agent<br />
Mathematical Morphology<br />
Segmentation Agent<br />
1……..N<br />
Figure 2.30: The Multi-Agent Optimization Model for Unstained Cell Images by<br />
Khosla et al. (Khosla, 2003)<br />
Lai et al. in (Lai, 2003) carried out the same function as previous research <strong>of</strong> Khosla by<br />
a little difference. They added a GA component to own model for initializing the IPA<br />
and for deciding which result is accepted for segmentation, as shown in Figure 2.31.<br />
The proposed method can achieve to segment cell images with accuracy <strong>of</strong> 100 percent.<br />
Figure 2.31: The Multi-Agent S<strong>of</strong>t Computing Model for Unstained Cell Images by Lai<br />
et al. (Lai, 2003)<br />
48
2.6.13 Richard et al. method<br />
Richard et al. (Richard, 2004) proposed MAS in which the aim is to segment the brain<br />
MR images. The framework, as shown in Figure 2.32, based on parallel execution <strong>of</strong> the<br />
agents. System manager launches agent executions in a sequential way. The agents are<br />
autonomous and have ability <strong>of</strong> the cooperation. In their framework, three types <strong>of</strong> the<br />
agents coexist such as global agent, local agent, and tissue-dedicated agents. The global<br />
agent specifies particular task, which performs to the whole image then to create local<br />
agents distributed over the image. The local agents create the tissue-dedicated agents, to<br />
estimate model parameters and to encounter tissue models for final labeling decisions.<br />
The tissue-dedicated agents execute tasks distributed by tissue type (gray matter, white<br />
matter, and cerebro-spinal fluid). They acquire the tissue models from the neighborhood<br />
and label the voxels using a region-growing process.<br />
The proposed method shows the correct estimation <strong>of</strong> the tissue-intensity distribution in<br />
different locations in the image, despite large intensity variations inside the same tissue.<br />
In addition, in comparison with the other methods, the proposed method has the<br />
significant performance in spite <strong>of</strong> the increasing non-uniformity <strong>of</strong> intensity.<br />
49
Figure 2.32: The Proposed Multi-Agent Framework by Richard et al. (Richard, 2004)<br />
2.6.14 Benamrane and Nassane method<br />
Benamrane et al. (Benamrane, 2007) proposed a multi-agent approach permitting<br />
segmenting brain MRI. They used two main types <strong>of</strong> the agent; global agent, and<br />
region agent, as shown in Figure 2.33. Global agent has three basic behaviors; initial<br />
segmentation, creating and launching the region agents, and coordinating <strong>of</strong> the region<br />
agents. Region agent can behave one <strong>of</strong> these six types; discovering the neighborhood<br />
agents, selection <strong>of</strong> the best fusion criterion from neighbor verifying, finding with<br />
which agent merges, growing, and disappearing.<br />
The proposed method is based on three steps. Firstly, the global agent segments image<br />
by region growing approach. Secondly, iterative merging <strong>of</strong> the initial regions from the<br />
previous step will merge the intermediate segmentation <strong>of</strong> the initial image. Finally,<br />
segmentation <strong>of</strong> the intermediate segmentation by iterative merging <strong>of</strong> the intermediate<br />
regions is obtained using a fusion criterion.<br />
50
Figure 2.33: The Proposed Method by Benamrane et al. (Benamrane, 2007)<br />
The proposed method has had acceptable results; each region presents clear-cut limits,<br />
particularly the tumor regions, which correctly detected. However, the execution time is<br />
exceedingly high.<br />
2.6.15 Discussion<br />
In Table 2.2 and Table 2.3, all researches related to segmentation using agent<br />
approaches are compared based on the properties <strong>of</strong> the agent using non-medical and<br />
medical images respectively. They are divided to class <strong>of</strong> the non-medical and medical<br />
image segmentations. Furthermore, this table is based on our opinion. Subsequently, the<br />
comparison results for each agent properties are concluded from the researchers work<br />
and the definition <strong>of</strong> the agent properties.<br />
Besides, Table 2.4 lists the advantages or disadvantages <strong>of</strong> some mentioned methods in<br />
comparison with non-agent method. These advantages and disadvantages came from the<br />
researchers beliefs.<br />
51
Non Medical image segmentation<br />
Table 2.2: The Comparison <strong>of</strong> the Segmentation Methods using Non-medical<br />
Kagawa et<br />
al.<br />
Wang et<br />
al.<br />
Gyohten<br />
Guillaud<br />
et al.<br />
Rodin et<br />
al<br />
Melkemi<br />
et al<br />
Images by Agent Properties (<strong>Chitsaz</strong>, 2008)<br />
Landscape<br />
photo<br />
Face<br />
photo<br />
Document<br />
image<br />
fish<br />
otoliths<br />
fish<br />
otoliths<br />
Landscape<br />
photo<br />
Image<br />
Modality<br />
Yes<br />
Yes<br />
Yes<br />
Yes<br />
Yes<br />
Yes<br />
Autonomy<br />
No<br />
No<br />
Yes<br />
Yes<br />
Yes<br />
Yes<br />
Social<br />
ability<br />
Agent<br />
Properties<br />
N/A<br />
Yes<br />
N/A<br />
No<br />
Yes<br />
Yes<br />
Reactivity<br />
No<br />
No<br />
Yes<br />
No<br />
No<br />
Yes<br />
Pro-activity<br />
N/A<br />
N/A<br />
5<br />
2<br />
2<br />
2<br />
Number <strong>of</strong><br />
Agent types<br />
52
Table 2.3: The Comparison <strong>of</strong> the Segmentation Methods using Medical Images by<br />
Medical image segmentation<br />
Spinnu<br />
et al.<br />
Boucher<br />
et al.<br />
Liu and<br />
Tang<br />
Germond<br />
et al.<br />
Duchesnay et<br />
al.<br />
Khosla<br />
and Lai<br />
Richard<br />
et al.<br />
Benamrane<br />
and<br />
Nassane<br />
Muscle<br />
cell and<br />
MRI<br />
Living<br />
cell<br />
Brain<br />
MR<br />
images<br />
Brain MR<br />
images<br />
Mammography<br />
Chinese<br />
Hamster<br />
Ovarian<br />
Cells<br />
Brain<br />
MR<br />
images<br />
Brain MR<br />
images<br />
Image<br />
Modality<br />
Agent Properties (<strong>Chitsaz</strong>, 2008)<br />
Yes<br />
Yes<br />
Yes<br />
Yes<br />
Yes<br />
Yes<br />
Yes<br />
Yes<br />
Autonomy<br />
No<br />
Yes<br />
No<br />
Yes<br />
Yes<br />
No<br />
Yes<br />
Yes<br />
Social<br />
ability<br />
Agent<br />
Properties<br />
Yes<br />
Yes<br />
Yes<br />
No<br />
Yes<br />
Yes<br />
Yes<br />
Yes<br />
Reactivity<br />
No<br />
Yes<br />
No<br />
No<br />
Yes<br />
No<br />
No<br />
Yes<br />
Proactivity<br />
8<br />
5<br />
4<br />
3<br />
2<br />
3<br />
3<br />
2<br />
Number<br />
<strong>of</strong> Agent<br />
types<br />
53
Table 2.4: The Comparison between Multi-Agent and Non-Agent Segmentation<br />
Methods<br />
Researchers<br />
Kagawa et al.<br />
Wang et al.<br />
Image size and<br />
modality<br />
Landscape image<br />
with size <strong>of</strong><br />
160×240<br />
30 fps for 160×120<br />
video size<br />
12 fps for 320×240<br />
video size<br />
Gyohten Document image<br />
Guillaud et al.<br />
Rodin et al.<br />
Melkemi et al.<br />
Spinnu et al.<br />
Otolith image with<br />
the size <strong>of</strong> 512×512<br />
or 1024×1024<br />
Otolith image with<br />
the size <strong>of</strong> 150×150<br />
Synthetic noisy<br />
image<br />
Muscle cell and<br />
MRI<br />
Boucher et al. Living cells<br />
Liu an Tang<br />
MRI <strong>of</strong> brain<br />
612×792<br />
Germond L. et al. MRI <strong>of</strong> brain<br />
Duchesnay et al.<br />
Khosla and Lai<br />
192×192 images<br />
including both<br />
medical and non-<br />
medical one<br />
Chinese Hamster<br />
Ovarian Cells<br />
Richard et al. MRI <strong>of</strong> brain<br />
Benamrane and<br />
Nassane<br />
MRI <strong>of</strong> brain that<br />
contains tumors<br />
Comparison with non-agent<br />
methods<br />
Less computation and can be applied<br />
to different images variety.<br />
No restriction to the face pose, face<br />
moving direction and speed. Also, it is<br />
much faster than template-based and<br />
neural network-based methods.<br />
Line recognition only with the<br />
knowledge on image. More<br />
computation time.<br />
An acceptable result to find the<br />
continuous ring <strong>of</strong> the otolith image.<br />
Automated detecting <strong>of</strong> the nucleus<br />
position. Difficult to tune the agents’<br />
parameters.<br />
Automated, and can be used to<br />
different type <strong>of</strong> the images.<br />
Experiment result shows the<br />
feasibility, convergence and<br />
robustness <strong>of</strong> this method. Faster than<br />
traditional method.<br />
Can find the optimal solution properly.<br />
The method is adaptable and the result<br />
has rich information.<br />
Less computation time. Agent<br />
distribution is not optimal.<br />
The mean quality percentage is equal<br />
to 96%. Considerable user interaction.<br />
The approach does not require<br />
substantial tuning effort and it is<br />
completely autonomous. Not required<br />
priori information.<br />
The accuracy is 100%.<br />
Adaptation to intensity non-uniformity<br />
and noise.<br />
Good success in image includes<br />
heterogeneous, local and repartee<br />
information.<br />
54
2.7 Image Segmentation by Reinforcement Learning Model<br />
The following describes the basic ideas <strong>of</strong> the researchers who contributed to the field<br />
<strong>of</strong> the image segmentation regarding RL system. Some <strong>of</strong> the following methods also<br />
used the MAS to implement a RL system in order to segment an image.<br />
2.7.1 Peng and Bhanu method<br />
Peng and Bhanu (Peng, 1998 a; Peng, 1998 b; Bhanu, 2000) proposed a framework for<br />
object recognition using RL approach. Some pre-processing steps need to achieve<br />
successful object recognition, like segmentation and feature extraction. The algorithm<br />
was used for segmentation is Phoenix Segmentation Algorithm. This algorithm works<br />
based on a recursive region splitting. It uses information from histogram <strong>of</strong> red, green<br />
and blue image components to split the region based on a peak/valley analysis <strong>of</strong> each<br />
histogram. The conceptual diagram is shown in Figure 2.34.<br />
The evaluation framework for object recognition is implemented by RL. In every loop,<br />
the best threshold for peak/valley selects and uses in segmentation algorithm; the<br />
diagram is shown in Figure 2.35.<br />
Consequently, the proposed method is capable <strong>of</strong> exploring a significant portion <strong>of</strong> the<br />
search space, resulting in the discovery <strong>of</strong> the good solutions due to the stochastic<br />
nature <strong>of</strong> RL. In general, this result cannot achieve by any deterministic or simple<br />
supervised learning methods.<br />
55
Figure 2.34: The Conceptual Diagram <strong>of</strong> the Phoenix Segmentation Algorithm (Peng,<br />
1998b)<br />
Figure 2.35: The Segmentation Evaluation by RL (Bhanu, 2000)<br />
56
2.7.2 Shokri method<br />
Shokri (Shokri, 2003) used concept <strong>of</strong> RL for finding best thresholding <strong>of</strong> an image.<br />
Their approach was based on standard RL framework, as shown in Figure 2.36. The<br />
model has states, actions and a matrix that saved the reward or punishment. The matrix<br />
update on each iteration. In addition, for achieving global goal, states should examine<br />
several times.<br />
Figure 2.36: The Standard Model <strong>of</strong> RL (Shokri, 2003)<br />
The model <strong>of</strong> the reward/punishment has two types; subjective and objective.<br />
Subjective case means an experienced user will assign a reward/punishment to the<br />
outcome image. Objective case is defined based on the black pixel ratio, the area <strong>of</strong> the<br />
object, the tolerance for area deviation, and the number <strong>of</strong> the objects.<br />
The proposed method achieves performance 87% for subjective method, and 60% for<br />
objective method. Additionally, this method needs considerable user interaction to<br />
achieve a better performance.<br />
2.7.3 Sahba method<br />
Sahba et al. (Sahba, 2008; Sahba, 2006 b) proposed a RL model to segment an ultra<br />
sound image <strong>of</strong> the prostate. First, the image is divided to some sub-images. Then,<br />
agents find the optimal threshold <strong>of</strong> all sub-images. After completing the segmentation<br />
57
<strong>of</strong> all sub-images, the result has compared by a manual segmented image (gold image).<br />
Subsequently, the reward or punishment is assigned to every agent. After training, the<br />
agent finds the best threshold for the image and possible to segment another image <strong>of</strong><br />
the same type. In addition, researchers have used a simple deformable model for<br />
extracting prostate from image; prostate has elliptical shape. The modified RL model is<br />
shown in Figure 2.37. Figure 2.38 shows the proposed reinforcement model.<br />
Figure 2.37: The RL Model used in the Proposed Method by Sahba (Sahba, 2006 b)<br />
Figure 2.38: The General Model used in the Proposed Method (Sahba, 2008)<br />
58
2.8 Chapter Summary<br />
In this chapter, some background knowledge <strong>of</strong> our research mentions. Brief<br />
explanations <strong>of</strong> two commonly used modalities for medical imaging are provided.<br />
Furthermore, the skull anatomy is explained to facilitate the correlation with the<br />
anatomical images. Subsequently, the basic image segmentation methods are elaborated<br />
to compare the agent-based method with conventional methods.<br />
In addition, researches, this related to our work, presents. The review covers most <strong>of</strong> the<br />
image segmentation methods that uses MAS. These methods are employed to segment<br />
medical images or non-medical images. These methods have acceptable results in<br />
comparison to the conventional methods (<strong>Chitsaz</strong>, 2008). In addition, reviews <strong>of</strong> the<br />
methods that use the RL system presents. Some <strong>of</strong> these works used solely the RL<br />
method to segment an image. The results are satisfactory for these proposed methods<br />
but are not very useful. This is because <strong>of</strong> the modified model <strong>of</strong> RL that was used in<br />
image segmentation needs the manually segmented image. Therefore, if we want to<br />
segment each image, it requires the manually segmented image too. Nevertheless, the<br />
best threshold for an image can be found using the RL method and then we can segment<br />
another image with the same characteristic using the result. Therefore, the RL method is<br />
a pre-processing method employed by the other methods proposals.<br />
As the result <strong>of</strong> reviewing the researches <strong>of</strong> other people, it is concluded that they<br />
employed the MAS with the automatic agent, or used reinforcement agent without any<br />
social ability <strong>of</strong> the agent. These disadvantages are our motivation to hybridize the<br />
MAS with RL agents, which this framework uses both the properties <strong>of</strong> the automatic<br />
agent and RL agent. Therefore, we want to propose a RL method to find the best<br />
59
threshold value <strong>of</strong> the images, and then propose another method to use the agent<br />
properties and the result <strong>of</strong> the RL method to segment another image.<br />
60
Chapter 3<br />
Methodology<br />
This chapter contains studies on the proposed medical image segmentation algorithm<br />
and represents a progression in its development. In conjunction with previous chapter,<br />
which studied some <strong>of</strong> the current image segmentation methods by means <strong>of</strong> MAS and<br />
RL model, this chapter discusses about developing and implementing our method based<br />
on the properties <strong>of</strong> the local agent and the RL agent.<br />
This chapter includes few sections. The next section describes the images, which used<br />
in this project. After that, the methodology <strong>of</strong> this research globally describes. The<br />
following section talks about the development <strong>of</strong> the system and how will initialize. The<br />
last section summarized the overall description <strong>of</strong> our method; it also describes how the<br />
research is concluded and why this conclusion can show the result is satisfactory.<br />
3.1 Image Acquisition<br />
The aim <strong>of</strong> this research is to segment the CT image <strong>of</strong> the head using RL agent<br />
properties into MAS. These images have been collected from the <strong>University</strong> <strong>of</strong> <strong>Malaya</strong><br />
Hospital (Obaidellah, 2006), and collected from internet (DICOMsample).<br />
61
The experiments are from two different data sets. In the first experimental data set, head<br />
CT images from UM Hospital, are acquired on a CT scanner with an image size<br />
512�512, and a pixel size <strong>of</strong> 0.5mm�0.5mm. Upper human body CT images for the<br />
second experiment (DICOMsample) are acquired on the same machine. The imaging<br />
protocol used is image size <strong>of</strong> 512�512, and a pixel size <strong>of</strong> 0.55mm � 0.55 mm.<br />
3.2 Image Segmentation<br />
Image segmentation has to employ important amount <strong>of</strong> the information, the importance<br />
<strong>of</strong> this task will be more when segmentation has to be processed on a sequence <strong>of</strong> the<br />
images. Complicated segmentation problems require sophisticated algorithm with more<br />
priori-knowledge from image. Moreover, algorithm with learning ability needs more<br />
training sets.<br />
To solve these problems, a trainable and parallel processing approach has developed.<br />
The proposed method consists <strong>of</strong> two disjoin phases; training phase, and testing phase<br />
as shown in Figure 3.1. In the following paragraphs, each phase will elaborate in detail.<br />
Figure 3.1 shows the relation between these two phases, training and testing. In the<br />
training phase, a little image is used as a trained image whereas the RL agent will find<br />
the appropriate value <strong>of</strong> each object or region in the input image. The outcome <strong>of</strong> this<br />
training phase is transferred to the next phase, testing phase. In this phase, the images<br />
are segmented by some priori knowledge and the properties <strong>of</strong> the local agent.<br />
62
Figure 3.1: The Global View <strong>of</strong> our Proposed Model<br />
In the training phase, new algorithm is proposed based on RL agent in order to segment<br />
the CT images. The RL agent can learn to segment the images over time by systematic<br />
trial and error. The RL agent is trained by obtaining rewards from its environment<br />
whereas these signals are based on its actions on its environment. Because <strong>of</strong> the<br />
dynamic nature <strong>of</strong> RL agent, it is appropriate to use this type <strong>of</strong> the agent to segment the<br />
complex textured images. The goal <strong>of</strong> RL agent is to find out an optimal way to reach<br />
the best answer with some signals, which obtained after each action. The best answer is<br />
the most accurate segmented image.<br />
For using the RL method in medical image segmentation, it is defined by some actions<br />
to identify regions in an image. In addition, there are states based on the number <strong>of</strong> the<br />
interest objects in image. Firstly, the agent takes the image and applies some values.<br />
The input image is divided into several sub-images, and each RL agent works on it to<br />
find the suitable value for each object in the image. Various actions are defined for each<br />
state in the environment. Besides, a reward function computes reward for each action <strong>of</strong><br />
63
RL agent. Therefore, the agent tries to learn which actions can gain the highest reward.<br />
Finally, the valuable gained information can be used to segment new similar images in<br />
the next phase. The valuable information is the thresholding value <strong>of</strong> each object in the<br />
image.<br />
In the testing phase, new algorithm based on local agent in MAS is proposed to segment<br />
the similar images <strong>of</strong> trained image. This method uses some priori knowledge, which<br />
can obtain from training phase. Local agent has some important properties <strong>of</strong> the agent,<br />
which described in Section 2.4, such as autonomy, reactivity, and social ability, less<br />
robustness and pro-activity.<br />
In the beginning <strong>of</strong> our proposed algorithm, the user should provide the number <strong>of</strong> the<br />
extracted objects from an input image. Additionally, a near estimation <strong>of</strong> the extreme<br />
thresholding <strong>of</strong> each object must be provided. The input image is divided into several<br />
sub-images, and each local agent works on it to mark pixels for each object in the image<br />
by means <strong>of</strong> the input data. If the thresholding range <strong>of</strong> the objects in the image has<br />
overlapped, then the local agent cannot individually decide to mark the pixels.<br />
Therefore, the local agent uses its social ability to make decision. In addition, this<br />
algorithm works in parallel, so each local agent simultaneously works with other local<br />
agents. Whenever agents have problem to mark a particular pixel, they would get some<br />
help from their neighbors.<br />
The number <strong>of</strong> the agent is directly depended on the window size and image size in both<br />
phases. If we choose a window size <strong>of</strong> 7�7, then the number <strong>of</strong> the agent can be<br />
acquired by dividing the image size to 49. Because the RL agent should stay on the<br />
center <strong>of</strong> the window and the quotient is a real number, the number <strong>of</strong> the agents is less<br />
64
than the quotient. In our framework, the image has 512�512=262144 pixels; the result<br />
for division <strong>of</strong> 262144 by 49 is 5349.87. As a result, some agents, which locate in<br />
rightmost and bottom side <strong>of</strong> the image, extend their areas to cover all pixels over<br />
image, so the number <strong>of</strong> the agent is 5329.<br />
For an example, Figure 3.2 shows an image with size <strong>of</strong> 16�16 pixels. We want to<br />
calculate the number <strong>of</strong> the agents that should be appointed into image when the<br />
window size is 7�7. As shown in the Figure 3.2, the agent 1, 2, 3, and 4 are placed in<br />
center <strong>of</strong> yellow, dark blue, dark green and dark orange window respectively. However,<br />
there are some uncovered pixels in rightmost and bottom <strong>of</strong> the image. Then, the<br />
window size <strong>of</strong> agent 2, 3, and 4 should extend to cover all pixels in the image.<br />
Therefore, the result <strong>of</strong> diving 16�16=256 to 49 is 5.22, but it only needs four agents to<br />
cover all pixels <strong>of</strong> the exemplified image. Because <strong>of</strong> this, the number <strong>of</strong> the agent for<br />
our framework, 5329, is less than the division result, 5349.<br />
1 2<br />
3 4<br />
Figure 3.2: An Example for Calculating the Number <strong>of</strong> the Agents within a Window<br />
Size <strong>of</strong> 7�7 over an Image Size <strong>of</strong> 16�16<br />
65
3.2.1 Training Phase<br />
The RL agent has already used in image processing task where it discussed in Chapter<br />
2. In this section, it will show that the RL agent is suitable for segmenting medical<br />
images in parallel. This method is specifically useful for medical images where there are<br />
several images from a patient having most similar characteristics. In such a case, some<br />
<strong>of</strong> the images can use as training images. Then, the appropriate parameter can find for<br />
segmenting the other similar images.<br />
In this phase, the image is divided into several sub-images. First, the user should<br />
provide the number <strong>of</strong> the interest regions. Besides, the manually segmented image <strong>of</strong><br />
the input image has to provide. Figure 3.3 (a) and (b) show a cranial CT image and its<br />
manually segmented version. They are applied as an input for the RL agent to obtain the<br />
knowledge from image. The RL agent determines the local thresholding value for each<br />
individual sub-image via dividing the maximum gray-scale <strong>of</strong> the input image by the<br />
given number <strong>of</strong> objects within image. The maximum gray-scale <strong>of</strong> our experimental<br />
result is 256. The Q-matrix will be constructed regards to states and actions.<br />
(a) (b)<br />
Figure 3.3: (a) The Original CT Image, (b) Manually Segmented Image<br />
The RL agent needs three components to learn from its environment; states, actions and<br />
reward. The RL agent starts its work using an input image and the manually segmented<br />
66
input image. The RL agent tries to find the appropriate state <strong>of</strong> an image, after that it<br />
chooses one <strong>of</strong> the defined actions. During this time, the RL agent changes the local<br />
thresholding values for each sub-image individually. By taking each action, the agent<br />
receives the corresponding reward for that state-action pair and updates the<br />
corresponding value in Q-matrix. After this process, the agent has explored many<br />
actions and has tried to exploit the most rewarding ones. The global view <strong>of</strong> this<br />
procedure is shown in Figure 3.4. The state is perceived from images. Then, the RL<br />
agent chooses one action to alter the image, after that it received a reward as evaluation<br />
<strong>of</strong> its work.<br />
Images<br />
State<br />
Action<br />
Agents<br />
Reward<br />
Image<br />
Processing<br />
Tasks<br />
Moderator<br />
RL Agents<br />
Figure 3.4: The Global View <strong>of</strong> our Proposed Method in Training Phase<br />
Figure 3.5 shows the flowchart <strong>of</strong> RL agent behavior. At the beginning, the RL agent<br />
discovers all pixels in the sub-image, and marks those pixels based on some fixed<br />
thresholding range. This fixed thresholding range could be acquired by dividing the<br />
maximum gray-scale <strong>of</strong> the image by the number <strong>of</strong> the acquired regions in image.<br />
67
Begin<br />
Discover every pixel in a 7x7 window and mark each pixel<br />
State selecting, the number <strong>of</strong> different type in<br />
window determines state<br />
Action selecting, changing the thresholding <strong>of</strong> each region<br />
No<br />
Receiving Reward<br />
Evaluation<br />
Is accuracy in<br />
window<br />
satisfactory?<br />
Yes<br />
Death<br />
End<br />
Figure 3.5: The RL Agent’s Behavior<br />
68
As an example, if the image has three object for segmenting, the first thresholding range<br />
will be [0, 85] for the first region, [85,170] for the second region, and [170,256] for<br />
third region. After marking all pixels in the window, RL agent would find its state.<br />
There are some actions for each state; RL agent should select one <strong>of</strong> them. � -Greedy is<br />
a method, which helps the RL agent to choose better possible action. Where � is a<br />
probability to choose action with most Q-value. A lookup table <strong>of</strong> Q-Matrix is used to<br />
store the Q-values. If � is less than a predefined parameter, the RL agent selects action<br />
with most Q-value, otherwise it will select action randomly.<br />
After choosing the action, the RL agent alters the primary thresholding value <strong>of</strong> each<br />
region by means <strong>of</strong> the maximum and the minimum thresholding in the current sub-<br />
image. A reward function calculates the number <strong>of</strong> the true-segmented pixels <strong>of</strong> the<br />
image. Then, information <strong>of</strong> this state-action is saved in the Q-matrix by following<br />
formula:<br />
Q(<br />
s,<br />
a)<br />
�(1�α)<br />
Q(<br />
s,a)<br />
�α[<br />
r�γ<br />
maxQ(<br />
s�,<br />
a�)]<br />
a�<br />
(3.1)<br />
After evaluating the RL agent work, if the result is satisfactory, the RL agent lifetime is<br />
finished. The satisfactory result is obtained when the accuracy <strong>of</strong> the segmented sub-<br />
image is more than 95%.<br />
In the following paragraphs, the definition <strong>of</strong> RL components presents. State, action and<br />
reward are described in depth.<br />
3.2.1.1 Definition <strong>of</strong> States<br />
The number <strong>of</strong> regions in the sub-image identifies the number <strong>of</strong> states. There are<br />
pixels, which related to a particular region in each sub-image. For example, there are<br />
three different regions such as bone, skin, and air in Figure 3.3(a).<br />
69
Refer to Figure 3.6; if the number <strong>of</strong> acquired regions in an image identified as two,<br />
then the states will be three as follows:<br />
� the sub-image that consists <strong>of</strong> the pixels <strong>of</strong> the first region type, as shown in<br />
Figure 3.6(a);<br />
� the sub-image that consists <strong>of</strong> the pixels <strong>of</strong> the second region type, as shown in<br />
Figure 3.6(b);<br />
� The sub-image consists <strong>of</strong> the pixels <strong>of</strong> both region types, as shown in<br />
Figure 3.6(c).<br />
Sub-image 1: Sub-image 2:<br />
(a) (b) (c)<br />
Figure 3.6: The Example <strong>of</strong> the number <strong>of</strong> States for an Image with two Regions<br />
Refer to Figure 3.7; if the number <strong>of</strong> the acquired region in image identified as three,<br />
then the states is seven as follows:<br />
Region I:<br />
Region II:<br />
� the sub-image that includes the pixels <strong>of</strong> the first region type, as shown in<br />
Figure 3.7(a);<br />
� the sub-image that includes the pixels <strong>of</strong> the second region type, as shown in<br />
Figure 3.7(b);<br />
� the sub-image that includes the pixels <strong>of</strong> the third region type, as shown in<br />
Figure 3.7(c);<br />
� the sub-image that includes the pixels <strong>of</strong> both first and second region type, as<br />
shown in Figure 3.7(d);<br />
70
� the sub-image that includes the pixels <strong>of</strong> both first and third region type, as<br />
shown in Figure 3.7(e);<br />
� the sub-image that includes the pixels <strong>of</strong> both second and third region type, as<br />
shown in Figure 3.7(f);<br />
� the sub-image that includes the pixels <strong>of</strong> all region types, as shown in<br />
Figure 3.7(g);<br />
Figure 3.7: The Example <strong>of</strong> the number <strong>of</strong> States for an Image with three Regions; each<br />
window shows a typical sub-image<br />
Therefore, the number <strong>of</strong> states can find by the following formula:<br />
n �<br />
�n<br />
N � � � �<br />
�<br />
� i<br />
i �1<br />
�<br />
�<br />
�<br />
�<br />
�<br />
�<br />
(3.2)<br />
Where N is the total number <strong>of</strong> states, and n is the number <strong>of</strong> acquired objects in the<br />
image.<br />
(a) (b) (c)<br />
(d)<br />
(e) (f)<br />
(g)<br />
Region I:<br />
Region II:<br />
Region III:<br />
71
3.2.1.2 Definition <strong>of</strong> Actions<br />
Actions change the local thresholding value <strong>of</strong> each sub-image. There are some actions<br />
for each state. An action employs the maximum and minimum gray-scale value in the<br />
image. This distance can be divided into several intervals with a predefined parameter.<br />
Each action would be defined as one <strong>of</strong> the intervals <strong>of</strong> this distance to threshold the<br />
sub-image. The number <strong>of</strong> actions depends on the predefined parameter, and the state.<br />
For example, if the predefined parameter is 2. The minimum and maximum the gray-<br />
scale value <strong>of</strong> the sub-image are 25 and 27 respectively, as shown in Figure 3.8. The<br />
actions for the state that shown in Figure 3.6(c) is choosing one <strong>of</strong> these sets {[25,25],<br />
[25,27]}, {[25,26], [26,27]}, or {[25,27], [27,27]}. One <strong>of</strong> these sets would choose as an<br />
action to segment the sub-image. Hence, the total number <strong>of</strong> actions is three. These sets<br />
can find via dividing the distance between maximum and minimum gray-scale value <strong>of</strong><br />
image by the predefined parameter. In this example, (27-25)/2 is one, so there are six<br />
different intervals; [25,25] , [25,26] , [25,27] , [26,26] , [26,27] , and [27,27]. Because<br />
the numbers <strong>of</strong> interest regions are two, it should select two intervals from those.<br />
Therefore, there are three sets.<br />
25 26 27<br />
Figure 3.8: An Example <strong>of</strong> Defining Action using the Maximum and Minimum<br />
Thresholding Gay-sale Value <strong>of</strong> a Typical Sub-image<br />
If the minimum and maximum values are similar for an action interval <strong>of</strong> n, it means<br />
pixels in the sub-image do not included the region number <strong>of</strong> n. For example, the first<br />
interval in the first set <strong>of</strong> the previous example is [25,25]. The minimum and maximum<br />
values are similar. It means there are no pixels, which marked by region type I in the<br />
72
sub-image, therefore all the pixels are from region type II. For another example, in this<br />
set {[25,27], [27,27]}, there is an interval with the same value at the first and the last<br />
interval, so it means there are no pixels with the type <strong>of</strong> region II in the sub-image.<br />
3.2.1.3 Definition <strong>of</strong> Reward<br />
The reward shows how well the image was segmented by RL agent. As a result an<br />
appropriate segmented <strong>of</strong> truth is needed for evaluation in place <strong>of</strong> true delineation. The<br />
reward function is defined as the number <strong>of</strong> pixels, which are segmented correctly. This<br />
function is like True-Positive, which discuss in Chapter 4.<br />
3.2.1.4 Graphical User Interface for Training Phase<br />
A GUI has developed for the training phase. There are some defined images used in the<br />
application. Therefore, all defined images are shown in a drop-down list to select one <strong>of</strong><br />
them. Then, after pushing the ‘segment’ button, the true positive volume fraction<br />
(TPVF) and false positive volume fraction (FPVF) <strong>of</strong> the segmented image are written<br />
in the box. Additionally, the estimation <strong>of</strong> the thresholding value for each region has<br />
shown in another textbox. Figure 3.9 shows GUI for training phase, whereas the input<br />
image and the segmented one are shown in different box. In addition, the TPVF and<br />
FPVF are written in different box.<br />
73
3.2.2 Testing Phase<br />
Figure 3.9: GUI <strong>of</strong> the Training Phase<br />
Autonomous agents have already been used in image processing tasks were discussed in<br />
Chapter 2. In this section, different MAS are proposed for segmenting medical images<br />
in parallel way using priori-knowledge from the training phase.<br />
Figure 3.10 shows a global view <strong>of</strong> this framework where there are three basic<br />
components; input materials, agent and its environment, and image processing task. The<br />
input materials are similar to the input images <strong>of</strong> the training phase. Additionally, the<br />
user gives an estimation <strong>of</strong> the thresholding for each region within an input image. This<br />
estimation came from the training phase. There are two main agent types, Moderator<br />
74
agent and Local Agent. These agents have the responsibility to segment the input image<br />
using the priori-knowledge and image processing tasks. Image processing tasks are<br />
procedures that related to segmentation <strong>of</strong> the images. We use these terms<br />
interchangeably in this thesis: local agent, autonomous agent, or agent.<br />
Input Image,<br />
Thresholding<br />
for each region<br />
<strong>of</strong> input image<br />
Local Agents<br />
Figure 3.10: The Global View <strong>of</strong> our Proposed Method in Testing Phase<br />
In this phase, the image was divided into several sub-images like the training phase. As<br />
shown in Figure 3.10, there are two different agents here. The moderator agent has a<br />
managing role in the framework, and the local agents are like labor.<br />
There is a moderator agent to create and initialize the local agents, after that each local<br />
agent commences its life. The moderator agent decides to create local agents in different<br />
parts <strong>of</strong> an image. Moreover, the moderator agent terminates the lifetime <strong>of</strong> a particular<br />
local agent if there is no progress for that local agent. Besides, after terminating the<br />
lifetime <strong>of</strong> all agents, moderator agent forces that local agents marked all pixels in the<br />
image. If there are some undiscovered pixels, the moderator agent will create a second<br />
generation <strong>of</strong> the local agents in the unmarked area.<br />
There are many local agents to segment the image. First, an estimation <strong>of</strong> the<br />
thresholding for each region within the image should enter. This priori knowledge can<br />
be derived from the training phase. At the final state <strong>of</strong> the training phase, RL agent<br />
Moderator<br />
Image Processing<br />
Tasks<br />
75
finds the thresholding for each region in sub-images. Therefore, there is little user-<br />
interaction to obtain the overall thresholding for each region within the image. For this<br />
attempt, our designed interface would help the user to find better thresholding for each<br />
region, by changing the thresholding and viewing the result.<br />
After the local agent had obtained the input materials, the local agent tried to mark each<br />
pixel in sub-images by means <strong>of</strong> the thresholding input. In the duration <strong>of</strong> the marking<br />
procedure, each agent should make a decision about label <strong>of</strong> each pixel in the sub-<br />
image. A local agent can do it by the given priori knowledge, but there are overlapped<br />
or gapped areas between given thresholding ranges <strong>of</strong> each region. In this situation, an<br />
agent uses its properties to negotiate with the other agents. It means that if an agent<br />
cannot find the type <strong>of</strong> a pixel, it negotiates with its neighbor agent to find an<br />
appropriate region type. Nevertheless, if no agent exists in the neighborhood <strong>of</strong> the<br />
current agent, or a neighbor agent has not known proper information yet, the agent will<br />
leave that pixel as unmarked for further processing <strong>of</strong> the other agents. Figure 3.11<br />
shows all the behavior <strong>of</strong> agents in the testing phase.<br />
76
Yes<br />
Death <strong>of</strong><br />
Local agent<br />
Start<br />
Moderator agent creates many local agents in different part <strong>of</strong> image.<br />
No<br />
Are all<br />
pixels<br />
marked?<br />
Local agent marks the pixel in its sub-image.<br />
Can mark<br />
pixel?<br />
Local agent uses one <strong>of</strong> the<br />
negotiation approaches to mark pixel.<br />
Can mark<br />
pixel?<br />
Mark the pixel to<br />
unable-to-mark<br />
Moderator agent checks all the pixels in the image. If there<br />
are undiscovered or unable-to-mark pixels, then it generates<br />
another generation <strong>of</strong> local agents.<br />
No<br />
Yes<br />
Yes<br />
Are all<br />
pixels<br />
discovered?<br />
End<br />
No<br />
No<br />
Yes<br />
Figure 3.11: The Agents’ Behavior in Testing Phase<br />
77
A local agent tries to find the appropriate label for the current pixel. If it cannot find the<br />
meaningful label, it goes to negotiate. The negotiation term is to calculate the mean<br />
value <strong>of</strong> the 3�3 window <strong>of</strong> the negotiable pixel. If this mean value is in the range <strong>of</strong><br />
discovered thresholding, then the negotiable pixel can mark by this mean value. The<br />
discovered thresholding means the thresholding range is not in the overlapped or<br />
gapped distance. The other approach for negotiation is to count the number <strong>of</strong><br />
discovered neighboring pixels. The major type in the counting approach specifies the<br />
label <strong>of</strong> the negotiable pixel. For example, there are 8 pixels around the negotiable<br />
pixel. If there are 3 pixels labeled as Region I, 4 pixels labeled as Region II, and 1 pixel<br />
labeled as Region III. Then, the outcome from this approach is to mark the negotiable<br />
pixel as Region II because majority <strong>of</strong> its neighbors were marked as Region II.<br />
3.2.2.1Graphical User Interface for Testing Phase<br />
A GUI has developed for the testing phase too. There are some pre-defined images for<br />
use in the application. Therefore, all defined images are shown in a drop-down list to<br />
allow a user select one <strong>of</strong> them. Then, the user should give the estimation <strong>of</strong> the<br />
thresholding value. This priori knowledge can find from the training phase. By pushing<br />
the ‘segment’ button, TPVF and FPVF <strong>of</strong> the segmented image are written in a box.<br />
Additionally, the input image and the segmented one <strong>of</strong> it will be shown in another<br />
window. Figure 3.12 shows the GUI for the testing phase, the input image and the<br />
segmented one are shown in different windows. In addition, the TPVF and FPVF are<br />
written in the box.<br />
78
3.3 Chapter Summary<br />
Figure 3.12: GUI <strong>of</strong> the Testing Phase<br />
The main purpose <strong>of</strong> this work is to segment medical images simultaneously with some<br />
different interest regions. Bearing in mind the obstacles <strong>of</strong> the medical image<br />
segmentation, the training methods need a huge training sample. Agent can learn to<br />
perform segmentation over time by systematic trial and error. The RL agent is trained<br />
by obtaining rewards or punishment based on its action in an environment. The goal <strong>of</strong><br />
the RL agent is to find out an optimal way to reach the best answer given some signals<br />
obtained after each action. The state and action should be defined for using RL method<br />
in medical image segmentation. Firstly, the agent takes an image and applies some<br />
values. The input image is divided into several sub-images, and each RL agent works on<br />
it to find the suitable value for each object in the image. Each state in the environment is<br />
associated with some actions. A reward function computes the reward for each action <strong>of</strong><br />
79
RL agent. Therefore, the agent tries to learn which actions can gain the highest reward.<br />
Finally, the gained valuable information will use to segment new similar images. Due to<br />
the dynamic nature <strong>of</strong> RL agent, it is suitable for segmenting images with high<br />
complexity.<br />
Another algorithm based on MAS is proposed to use the valuable information from the<br />
previous method. Local agent can use its properties to perform segmentation over time.<br />
The goal <strong>of</strong> the agent is to find out appropriate label for each pixel in an image. Firstly,<br />
a moderator agent creates and initializes the local agents within the image. The local<br />
agent takes a sub-image and applies some values. The input image is divided into<br />
several sub-images, and each agent works on it and tries to mark each pixel in sub-<br />
images by means <strong>of</strong> the given priori knowledge. During this time, the local agent marks<br />
each cell <strong>of</strong> a sub-image individually. Finally, the moderator agent checks the outcome<br />
<strong>of</strong> all agents’ work to produce a final segmented image.<br />
Finally, it is required to show how well our algorithms are. Therefore, the outcome<br />
evaluation <strong>of</strong> these two proposed methods will discuss in the next chapter in order to<br />
show how well the algorithms are proposed.<br />
80
Chapter 4<br />
Experimental Results and Discussion<br />
Based on the methods described in Chapter 3, we carried out experiments and analyzed<br />
the results. This chapter has two major parts, i.e., training phase results and testing<br />
phase results. Two separate groups <strong>of</strong> images are used for training and testing; in each<br />
group there are two data sets, one is from UM and another is from public domain on the<br />
internet. The experimental results will discuss, analyze qualitatively, and quantitatively.<br />
4.1 Experimental Results <strong>of</strong> Training Procedure<br />
In this section, we consider the experiments and results for the training phase, both<br />
qualitatively, and quantitatively. Image display uses for qualitative evaluation, and<br />
statistical analysis uses for quantitative evaluation. We name the proposed training<br />
method Proposed Reinforcement Learning Model (PRLM). For evaluating this phase,<br />
we used 33 different images <strong>of</strong> the same modality, size, and tissues in image.<br />
RL agent has a remarkable role in PRLM. As mentioned in the methodology chapter,<br />
every RL agent should work on a 7�7 window over the image. After marking every<br />
81
pixels <strong>of</strong> the window, it should receive a reward based on the training data set, and<br />
manually segmented image <strong>of</strong> the corresponded image.<br />
The heart <strong>of</strong> RL in the system uses the eq. 4.1 to learn:<br />
Q(<br />
s,<br />
a)<br />
� ( 1 � � ) Q(<br />
s,<br />
a)<br />
� �[<br />
r � � max Q(<br />
s�,<br />
a�)]<br />
(4.1)<br />
a�<br />
Where s is a state, a is an action, r is a reward, � is a learning rate that was initialized to<br />
0.9, and � is the discount factor that was set to 0.1.<br />
We have used the � -Greedy method for choosing an action, � set to 0.7. The number<br />
<strong>of</strong> iterations, to examine every action <strong>of</strong> a specified state, is another fixed parameter<br />
employs. We found in our experiment by trial and error, the reasonable number is 200.<br />
This number <strong>of</strong> the iterations should not be bigger because the processing time would<br />
increase. In addition, if the number <strong>of</strong> iterations becomes smaller, the RL agent cannot<br />
examine more actions to reach steady state. In other words, the Q-matrix would not be<br />
full for all state-action pairs<br />
4.1.1 Image Data Sets <strong>of</strong> PRLM<br />
The images used in the experiments list in Table 4.1. In the first experiment, head CT<br />
images are acquired on a CT scanner with an image size 512�512, and a pixel size <strong>of</strong><br />
0.5mm�0.5mm. Upper human body CT images for the second experiment<br />
(DICOMsample) are acquired on the same machine. The imaging protocol used is<br />
image size <strong>of</strong> 512�512, and a pixel size <strong>of</strong> 0.55mm � 0.55mm.<br />
Table 4.1: Details <strong>of</strong> the Image Data set used in the Experiments<br />
Data Modality Object Image size Pixel size No. <strong>of</strong><br />
set<br />
images<br />
First CT Head 512�512 0.55mm�0.55mm 18<br />
Second CT Upper human<br />
body<br />
512�512 0.55mm�0.55mm 15<br />
82
In overall, 33 slices has selected from full 3D CT images; these slices are at the same<br />
location in the body and the same orientation with respect to the body. Each image set<br />
has used for the training phase and testing phase. The numbers <strong>of</strong> training data are 33<br />
images, and all slices are used as a test set.<br />
4.1.2 Qualitative Analysis <strong>of</strong> PRLM<br />
A subjective inspection discovered that in all experiments and in all data, the results are<br />
close to their manually segmented images. Some examples are displayed in Figure 4.1.<br />
In addition, initialization was done in identical manner for all experiments images to<br />
evaluate the results <strong>of</strong> the input images.<br />
Figure 4.1 shows the segmented image by our proposed algorithm, the input image, and<br />
the ground truth image. Each row shows different slice <strong>of</strong> the 3D images from data set.<br />
The first two rows are from the first experiment data set, and the two remaining rows<br />
are from the second data set. More results are shown in Appendix A.<br />
4.1.3 Quantitative Analysis <strong>of</strong> PRLM<br />
The objective <strong>of</strong> the quantitative segmentation evaluation is to assess segmentation<br />
methods, or compare some segmentation methods with each other. The inter-technique<br />
and intra-technique are two classes <strong>of</strong> the quantitative analysis. The inter-technique<br />
reveals the performance <strong>of</strong> different techniques in segmenting the same type <strong>of</strong> images.<br />
The intra-technique recognizes the behavior <strong>of</strong> the considered technique in segmenting<br />
various kinds <strong>of</strong> the images (Zhang, 2001). In this research, we assess the proposed<br />
methods through various kinds <strong>of</strong> the images.<br />
83
(a) (b) (c)<br />
(a)<br />
(a)<br />
(a)<br />
(b)<br />
(b)<br />
Figure 4.1: The Segmentation Example from Two Experiments and Four Different<br />
Slices <strong>of</strong> 3D CT Images, (a) Input Image (b) Result from Proposed method and (c)<br />
Ground Truth Image<br />
(c)<br />
(c)<br />
(b) (c)<br />
84
Another classification <strong>of</strong> the evaluation methods consist <strong>of</strong> analytical methods,<br />
goodness methods, and discrepancy methods. For assessing the complexity <strong>of</strong><br />
algorithm, the analytical methods employ the algorithm for segmenting image by<br />
considering the requirement, variants, etc. The goodness methods evaluate the<br />
segmented image by measuring the intra-region uniformity, inter-region contrast, and<br />
region shape. The discrepancy methods compute the difference between segmented<br />
image and ground truth <strong>of</strong> the considered image. In other words, these methods try to<br />
determine the number <strong>of</strong> missed-segmented pixels, position <strong>of</strong> missed-segmented<br />
pixels, feature value <strong>of</strong> segmented objects, and miscellaneous quantities (Zhang, 2001).<br />
Choosing an appropriate evaluation method is an important task. However, it is obvious<br />
now that two types <strong>of</strong> metrics must measure for evaluation: accuracy, and efficiency<br />
(Fenster, 2005). Accuracy <strong>of</strong> a segmentation technique is to show how far actually<br />
segmented image differs from the manually segmented one. Efficiency <strong>of</strong> a segmented<br />
method is the segmentation time that all aspects <strong>of</strong> the user interaction should consider.<br />
4.1.3.1 Accuracy <strong>of</strong> PRLM<br />
The difference between the actually segmented image and the manually segmented<br />
image determine as accuracy <strong>of</strong> a segmentation technique. As a result, an appropriate<br />
segmented ground truth requires for evaluation in place <strong>of</strong> the true delineation. In all<br />
experiments, all data sets have manually segmented in the domain. For any<br />
image A� ( C,<br />
f ) , where C is a 2D (or higher-dimensional) rectangular pixels array, and<br />
f (c)<br />
denotes the intensity <strong>of</strong> any pixel c in C, letC M<br />
be the segmentation result <strong>of</strong><br />
d<br />
method M which obtained from C, and Ctd is the true delineation. U is a binary<br />
d<br />
image representation <strong>of</strong> a reference superset <strong>of</strong> pixels that is used to express the two<br />
85
measures as a fraction. We have used true positive volume fraction (TPVF) and false<br />
positive volume fraction (FPVF) from (Udupa, 2006). Eq. 4.2 and 4.3 are sufficient to<br />
describe the accuracy <strong>of</strong> the method:<br />
TPVF<br />
M<br />
d<br />
�<br />
C<br />
M<br />
d<br />
C<br />
U<br />
�<br />
C<br />
M<br />
td<br />
C<br />
C<br />
C<br />
td<br />
�100<br />
(4.2)<br />
�<br />
M d td<br />
FPVF � �100<br />
, (4.3)<br />
d<br />
�<br />
d<br />
td<br />
The sample cases were 33 slices <strong>of</strong> the CT images; therefore, Table 4.2 listed the mean<br />
and standard deviation <strong>of</strong> TPVF and FPVF achieved in our experiments by the proposed<br />
method. The TPVF <strong>of</strong> all data sets are above 96%, and their FPVF do not exceed 0.9%.<br />
The TPVF and FPVF obtain for each slice; the results <strong>of</strong> first and second data set are<br />
shown in Appendix B.<br />
The Receiver Operating Characteristic curve (ROC) is a plot <strong>of</strong> the true positive fraction<br />
against the false positive fraction. The closer the curve follows the left-hand border and<br />
then the top border <strong>of</strong> the ROC space, the more accurate the test. The closer the curve<br />
comes to the 45-degree diagonal <strong>of</strong> the ROC space, the less accurate the test. In<br />
addition, the area under the curve is a measure <strong>of</strong> the accuracy. Figure 4.2 and Figure<br />
4.3 demonstrate the ROC curve for head CT and upper human body data sets<br />
respectively. As the result <strong>of</strong> this curve, the result is accurate; also, the skin result is<br />
more accurate than the segmented result from bone.<br />
Data Set<br />
Head CT<br />
Images<br />
Upper human<br />
body<br />
Table 4.2: TPVF and FPVF <strong>of</strong> PRLM<br />
TPVF (%) FPVF (%)<br />
BG Skin Bone BG Skin Bone<br />
97.63<br />
99.94 �<br />
� 0.9<br />
0.07<br />
6<br />
96.66 � 1.26 0.34 � 0.16 0.38 � 0.20 0.60 � 0.25<br />
99.96 � 0.03 99.55 � 1.5 96.10 � 1.41 0.17 � 0.05 0.17 � 0.07 0.07 � 0.03<br />
86
Figure 4.2: ROC Curve for the First Data set<br />
Figure 4.3: ROC Curve for the Second Data set<br />
87
4.1.3.2 Efficiency <strong>of</strong> PRLM<br />
In determining the efficiency <strong>of</strong> a segmentation method time, all aspects <strong>of</strong> the user<br />
interaction should consider (Fenster, 2005). PRLM implements on 2.00 GHz Intel Core<br />
2 Duo and 2.00 GB RAM. The efficiency <strong>of</strong> our segmentation method provides<br />
information on the sensible use <strong>of</strong> the algorithm. In the proposed algorithm, a user does<br />
not interact with the program in the training phase. Therefore, the time for user<br />
interaction has ignored. The computation time directly relates to the image size and the<br />
number <strong>of</strong> iteration for filling the Q-matrix. However, in the training phase, the reward<br />
function needs the manually segmented image <strong>of</strong> current image, therefore there is in<br />
average 10 minutes to segment an image by means <strong>of</strong> the imaging s<strong>of</strong>tware like<br />
Photoshop.<br />
Table 4.3 depicts the mean computation time for all data sets. For every set, the<br />
program ran 15 times for each slide and the mean computation time is measured.<br />
Table 4.3: Efficiency <strong>of</strong> the PRLM<br />
Data set Computation time<br />
Head CT Image 13 Seconds<br />
Upper human body 7 Seconds<br />
4.2 Experimental Results <strong>of</strong> Testing Procedure<br />
In this section, we demonstrate both qualitatively, through image display, and<br />
quantitatively, through evaluation experiments for the next phase <strong>of</strong> our algorithm. In<br />
training phase, the RL agent achieved to discover the thresholding range for each<br />
region, so the result can use in testing phase to segment the similar images from the<br />
88
input images <strong>of</strong> the training phase. We name our proposed testing method Proposed<br />
Multi-Agent Model (PMAM).<br />
As mentioned in Chapter 3, local agent starts to segment the image using RL agent<br />
result. In this phase, each local agent through marked-pixel table can negotiate with<br />
other local agent. Therefore, the social ability, reactivity, and autonomy <strong>of</strong> the agent<br />
properties have been satisfied. However, it is required to adjust the result <strong>of</strong> the training<br />
phase a few times, so some user interactions are necessary in this phase. In the training<br />
phase, the exact thresholding range obtains for each window in an image. Meanwhile<br />
some user interaction is required to find the exact thresholding range for each region<br />
globally.<br />
Predetermined parameter does not use through this phase. We have a global marked-<br />
pixel table, which can refer in experiment <strong>of</strong> other local agent. Figure 4.4 shows the<br />
GUI, which brings out the thresholding value for each window and the mean value <strong>of</strong><br />
each row. After consideration <strong>of</strong> those numbers, a user should suggest the thresholding<br />
for each region. Figure 4.5 depicts the GUI <strong>of</strong> the testing phase, which after each<br />
suggestion the TPVF, FPVF, and segmented image have shown, then the user can<br />
suggest to get better results.<br />
89
Figure 4.4: GUI <strong>of</strong> the Training Phase to Suggest the User Thresholding Range <strong>of</strong> each<br />
Region<br />
Figure 4.5: GUI <strong>of</strong> the Testing Phase<br />
90
4.2.1 Image Data Sets <strong>of</strong> PMAM<br />
The image data sets used in the experiments are exactly the same properties that briefly<br />
described in section 4.1.1. In this phase, 28 images as test sets are used. Table 4.4 shows<br />
the details <strong>of</strong> data sets. For testing each image <strong>of</strong> data set, only one image from training<br />
phase is used.<br />
Table 4.4: Details <strong>of</strong> the Image Data set used in PMAM<br />
Data Modality Object Image size Pixel size No. <strong>of</strong><br />
set<br />
images<br />
First set CT Head 512�512 0.55mm�0.55mm 15<br />
Second CT Upper 512�512 0.55mm�0.55mm 13<br />
set<br />
human body<br />
4.2.2 Qualitative Analysis <strong>of</strong> PMAM<br />
A subjective inspection discovered that in all experiments and in all data, the results are<br />
satisfactory. Figure 4.6 shows the segmented image using our proposed algorithm in<br />
the second phase, the input image, and the ground truth. Each row shows different slice<br />
<strong>of</strong> the 3D images from the CT image <strong>of</strong> the head. The first two rows are from the first<br />
experiment data set, and the two remaining rows are from the second data set. For the<br />
first two rows, the result <strong>of</strong> RL agent from the first slice, Figure 4.1 (the first row), have<br />
used. Meanwhile for the second two rows <strong>of</strong> Figure 4.6, the result <strong>of</strong> RL agent from the<br />
253 rd slice, Figure 4.1 (the third row), was employed. More results <strong>of</strong> first and second<br />
data set from testing phase are shown in Appendix C.<br />
91
(a) (b)<br />
(a) (b)<br />
(a) (b)<br />
(a) (b)<br />
Figure 4.6: The Segmentation Example from Two Experiments and Four Different<br />
Slice <strong>of</strong> the Data set, (a) Result from our Method, (b) Input Image<br />
92
4.2.3 Quantitative Analysis <strong>of</strong> PMAM<br />
For evaluating PMAM, we used those two types <strong>of</strong> metrics: accuracy, and efficiency<br />
(Fenster, 2005). The following are comprehensive explanations for each <strong>of</strong> these<br />
metrics.<br />
4.2.3.1 Accuracy <strong>of</strong> PMAM<br />
TPVF and FPVF were defined in Eq. (4.2) and (4.3) at Section 4.1.3. Table 4.5 lists the<br />
mean and standard deviation values <strong>of</strong> TPVF and FPVF achieved in the two<br />
experiments. All data sets have TPVF above 90% except bone tissue in the second data<br />
set and FPVF less than 1.5%. The TPVF and FPVF slice obtains for each slice. The<br />
results <strong>of</strong> first and second data set are shown in Appendix D.<br />
The ROC curve is a plot <strong>of</strong> the true positive fraction against the false positive fraction.<br />
Figure 4.7 demonstrates the ROC curve for the head CT data set, and Figure 4.8 depicts<br />
the ROC curve for upper human body CT data set.<br />
Table 4.5: TPVF and FPVF for the Testing Phase <strong>of</strong> PMAM<br />
D a t a S e t<br />
TPVF (%) FPVF (%)<br />
BG Skin Bone BG Skin Bone<br />
Head CT<br />
Images<br />
95.55<br />
99.13 �<br />
� 2.5<br />
0.53<br />
8<br />
91.63 � 2.33 0.88 � 0.82 1.41 � 0.70 1.28 � 0.76<br />
Upper<br />
human body<br />
98.82<br />
99.61 �<br />
� 0.6<br />
0.54<br />
6<br />
85.53 � 3.53 0.26 � 0.21 1.02 � 0.52 0.21 � 0.17<br />
93
Figure 4.7: ROC Curve <strong>of</strong> the First Data set<br />
Figure 4.8: ROC Curve <strong>of</strong> the Second Data set<br />
94
4.2.3.2 Efficiency <strong>of</strong> PMAM<br />
PMAM implements on a 2.00 GHz Intel Core 2 Duo CPU, and 2.00 GB RAM personal<br />
computer. The efficiency <strong>of</strong> our segmentation method provides information on the<br />
sensible use <strong>of</strong> the algorithm. In the testing phase, a user only needs minimal interaction<br />
with the program. The time for user interaction has been denoted TU. In addition, the<br />
computation time, TC, which directly related to the image size and the number <strong>of</strong> agents,<br />
has considered.<br />
Table 4.6 depicts the mean user interaction time and the mean computation time for<br />
each image. For every slice in each data set, 15 times the program ran, so the mean<br />
computation time is measured. In addition, Figure 4.9 shows the bar chart <strong>of</strong> the<br />
computation time for all data sets.<br />
Table 4.6: Mean User-interaction Time and Computation Time <strong>of</strong> PMAM<br />
Data set TU TC<br />
Head CT Images 60 Seconds 7 Seconds<br />
Upper human body 60 Seconds 7 Seconds<br />
95
Figure 4.9: The Computation Time <strong>of</strong> all Data sets; X axis shows the slice number<br />
(identity) and Y axis shows the computation time (second).<br />
4.3 Chapter Summary<br />
In this chapter, evaluation <strong>of</strong> our proposed method is discussed both qualitatively<br />
through image display, and quantitatively through evaluation experiments.<br />
In the training phase, PRLM is based on RL model has demonstrated impressive results.<br />
The accuracy is more than 95% for each region and the computation time is less than 13<br />
seconds<br />
In the testing phase, PMAM is implemented. The result is satisfactory; the accuracy<br />
found is more than 85% for each region and the computation time recorded in less than<br />
6 seconds. However, the achieved accuracy is less than PRLM, but the efficiency is<br />
much better. Therefore, the result is satisfactory because there is a balance between<br />
accuracy and efficiency.<br />
96
Chapter 5<br />
Conclusions<br />
5.1 The Proposed Reinforcement-Learning Model<br />
Section 3.2 contains a description <strong>of</strong> the proposed RL model (PRLM), it is also identify<br />
a progression <strong>of</strong> the development. Section 4.1 presents the evaluation <strong>of</strong> PRLM.<br />
PRLM utilizes a standard RL model to segment images. State, action, and reward<br />
function also defined. The RL agents have used them to learn from the image. Every<br />
state has one or more actions. RL agent in each state decides to choose one action.<br />
Therefore, RL agent marked the image. It means each pixel in the image labels as a<br />
specific tissue such as bone, skin, etc. Finally, a reward function evaluates the accuracy<br />
<strong>of</strong> the segmented image, and gives a reward to the RL agent.<br />
Quantitative comparison <strong>of</strong> PRLM results in the training phase does not anticipate;<br />
however, having the manually segmented image for side-by-side comparison gives an<br />
opportunity to consider the advantages <strong>of</strong> our proposed method. Furthermore, the<br />
97
qualitative comparison shows encouraging results. The accuracy <strong>of</strong> PRLM is more than<br />
95% for each region in the image. PRLM is almost automatic. It only requires the<br />
manually segmented image for the reward function. The most significant advantage <strong>of</strong><br />
our proposed method is segmentation <strong>of</strong> an image into more than two regions in a<br />
parallel way. It means the regions <strong>of</strong> the interest can be more than one and with<br />
complex characteristics. For example, the CT image <strong>of</strong> the head consists <strong>of</strong> three<br />
different regions, such as the air (background), bone, and skin. PRLM segments the<br />
image into these three different objects simultaneously. In addition, the number <strong>of</strong><br />
training data set decreases in comparison with the neural networks approaches, as well<br />
as, the other learning methods. The efficiency illustrates this method is quick in<br />
comparison to the other method. The achieve computation time <strong>of</strong> PRLM is less than 13<br />
seconds. In Table 5.1, the efficiency <strong>of</strong> some segmentation methods lists. Some <strong>of</strong> them<br />
have a better efficiency; it is because <strong>of</strong> the smaller image sizes that used in their<br />
framework.<br />
Table 5.1: Efficiency Comparison <strong>of</strong> the Image Segmentation Methods<br />
Researcher Method Data Set PC Specification Efficiency<br />
Liang and<br />
Rodriguez<br />
(Liang, 1996)<br />
Pan and Lu<br />
(Pan, 2007)<br />
Lu and Bao<br />
(Lu, 2006)<br />
<strong>Chitsaz</strong> and<br />
woo (<strong>Chitsaz</strong>,<br />
2009 a)<br />
<strong>Chitsaz</strong> and<br />
woo (<strong>Chitsaz</strong>,<br />
2009 b)<br />
Fuzzy C-Mean<br />
Region-Growing<br />
Extended Image<br />
Force Model <strong>of</strong><br />
Snakes<br />
Reinforcement<br />
Learning<br />
Multi-Agent<br />
System<br />
MR images <strong>of</strong> a<br />
patient’s head<br />
(256�256)<br />
Skull CT image<br />
(256�256) and<br />
liver (384�384)<br />
Heart (160�169)<br />
Lung (225�211)<br />
2 different data<br />
sets <strong>of</strong> the CT<br />
image <strong>of</strong> head<br />
(512�512)<br />
2 different data<br />
sets <strong>of</strong> the CT<br />
image <strong>of</strong> head<br />
(512�512)<br />
Sun SPARCstation<br />
10/50<br />
Windows 2000 and<br />
VC++ 6.0 platforms<br />
Not available<br />
2.00 GHz Intel Core<br />
2 Duo and Java<br />
platform<br />
2.00 GHz Intel Core<br />
2 Duo and Java<br />
platform<br />
Pixel-based:<br />
9.7 (min)<br />
Region-based:<br />
0.73 (min)<br />
Skull: more than<br />
10 second<br />
Liver: less than<br />
5 second<br />
Heart: 8.3 s<br />
Lung: 30.2 s<br />
First Data set: 13s<br />
Second Dataset: 7s<br />
First Data set: 7s<br />
Second Dataset: 7s<br />
Although the qualitative result shows accuracy <strong>of</strong> PRLM, the method does not work for<br />
all images in the data set because <strong>of</strong> some reasons. First <strong>of</strong> all, at the beginning <strong>of</strong> the<br />
98
algorithm, some predetermined conditions, such as the number <strong>of</strong> iterations,� ,� ,<br />
and� , need to be set. These conditions cannot change in the middle <strong>of</strong> program<br />
execution. Therefore, these predetermined conditions may not suit some specific images<br />
in the duration <strong>of</strong> running the program. For example, the number <strong>of</strong> iteration has been<br />
set to 200 cycles, for images with a much narrower histogram, this number is not<br />
sufficient to fill all the cells in Q-Matrix.<br />
Moreover, states define based on gray-scale value. This should improve to cover more<br />
image features like texture, or shape in the future. In addition, the numbers <strong>of</strong> states<br />
depend on the number <strong>of</strong> objects to recognize in the image.<br />
Finally, the number <strong>of</strong> actions for each state is rigid. For each image window, which<br />
covers a bigger range <strong>of</strong> gray-scale, the actions are not satisfactory. Since the maximum<br />
and minimum thresholding <strong>of</strong> the gray-scale in the window is too large, finding the<br />
appropriate gray-scale would not be achievable for each region in window because that<br />
involve huge computation cycles. Meanwhile, the choosing method <strong>of</strong> the action is � -<br />
Greedy in our framework. It can change to a comprehensive method to obtain better<br />
performance in the accuracy.<br />
5.2 The Proposed Multi-Agent model<br />
In section 3.3, we have proposed a multi-agent model (PMAM) to segment an image, by<br />
input <strong>of</strong> the maximum and the minimum gray-scale value <strong>of</strong> each region in image.<br />
Section 4.2 evaluated PMAM both qualitatively and quantitatively.<br />
Quantitative assessment <strong>of</strong> PMAM shows the results are aggravated in comparison with<br />
PRLM result. The achieved accuracy from PMAM is more than 85% in each region <strong>of</strong><br />
99
the image. Furthermore, the efficiency time is less than 7 seconds for all data sets.<br />
Nevertheless, it is conceivable to improve result by some morphological operations.<br />
Furthermore, the qualitative comparison shows interesting result, and better<br />
computation time. The proposed method is almost automatic. It requires a little<br />
adjustment on the result from PRLM outcome. The most significant advantage is<br />
segmenting image to more than two regions in a parallel way. PRLM has this property<br />
too. It means the interest regions can be more than one and with different<br />
characteristics. For example, the CT image <strong>of</strong> the head consists <strong>of</strong> three different<br />
regions, such as air, bone, and skin. PMAM segments the image into three different<br />
objects simultaneously. Moreover, the efficiency illustrates PMAM is very fast in<br />
comparison to PRLM. It is possible to compare the result from testing phase with the<br />
other methods’ result in Table 5.1.<br />
However, the qualitative result shows high accuracy <strong>of</strong> our proposed method; but the<br />
method has a few defects because <strong>of</strong> some reasons. First, PMAM is such a simple<br />
approach that used for images with less noise. When there is some noise or foreign<br />
object such tooth filling material in the CT image, the noise would label as bone tissue<br />
instead <strong>of</strong> background or air, as shown in Figure 4.6. Of course, it is possible to improve<br />
the method by putting more constraint in the model. For example, we can improve the<br />
social ability <strong>of</strong> each agent. In PMAM, the local agent uses the gray-scale value <strong>of</strong><br />
neighborhood pixels and the mean value <strong>of</strong> neighborhood pixels for making decision to<br />
mark each pixel. This straightforward manner can modify to a comprehensive one in the<br />
future. Finally, the adjustment <strong>of</strong> gray-scale value from training phase is a tedious work;<br />
the expert should consider the mean value <strong>of</strong> each low or high extreme for every region<br />
to conclude an appropriate gray-scale range. However, the GUI can help the user to see<br />
100
each segmented image after each adjustment <strong>of</strong> the thresholding, as shown in Figure 4.4<br />
and 4.5.<br />
5.3 Achievements<br />
In summary, we have shown that the proposed methods can be used to segment<br />
different anatomic structure in medical images. Thus, the methods fulfilled our<br />
objectives mentioned in the first chapter. For the image segmentation <strong>of</strong> the head, we<br />
have the some objectives. The first objective is to review the existing experimental<br />
studies were investigating the Agents Technology for image segmentation. We have<br />
done this objective in Chapter 2 that we have presented some important methods that<br />
related to RL system and MAS. Another objective is to propose a novel approach by RL<br />
agent in the multi-agent framework that will be quicker, more accurate, and more<br />
robust. In Chapter 3, the two proposed method explain in detail. The final objective is to<br />
evaluate the outcome <strong>of</strong> segmented image by the proposed method with an appropriate<br />
estimation approach. In Chapter 4, there are a comprehensive evaluation <strong>of</strong> our method<br />
to show the accuracy and efficiency <strong>of</strong> it.<br />
In the following are the main results <strong>of</strong> PRLM:<br />
� PRLM attains significant result in segmentation accuracy; the accuracy is<br />
more than 95% for each region in the image.<br />
� PRLM achieves satisfactory result in computation time; the mean<br />
computation time <strong>of</strong> all datasets is less than 13 seconds.<br />
� The number <strong>of</strong> training data set for PRLM can be one or a small number <strong>of</strong><br />
images.<br />
� PRLM has the ability to segment simultaneously an image into some<br />
distinct regions.<br />
101
Furthermore, we have shown that PMAM can be used to segment different anatomic<br />
structure in medical images, which the input data obtained from PRLM. The main<br />
results <strong>of</strong> this method summarizes below:<br />
� PMAM attains significant result in segmentation accuracy; the accuracy is<br />
more than 85% for each region in the image.<br />
� PMAM achieves satisfactory result in computation time; the mean<br />
computation time <strong>of</strong> all datasets is less than 7 seconds.<br />
� PMAM is capable to segment simultaneously an image into some distinct<br />
regions.<br />
Simulating facial surgery and predicting the effect <strong>of</strong> an operation is very pivotal. One<br />
<strong>of</strong> the primary preprocessing levels in every kind <strong>of</strong> surgical simulation is image<br />
segmentation. Therefore, the accuracy <strong>of</strong> this level is essential because the result will<br />
affect the overall outcome. This work shows the achieved accuracy is such impressive<br />
that it can use in simulating facial surgery as a primary preprocessing level.<br />
5.4 Future work<br />
To tackle the remaining problems in PRLM that mentioned in Section 5.1, three ideas<br />
stand out for future works. The first would be an investigation to use comprehensive<br />
state model, which can cover more features <strong>of</strong> an image such as tissue, shape, etc. This<br />
would increase the scope <strong>of</strong> images that can be used to segment. For reducing the<br />
problem <strong>of</strong> action choosing, it is conceivable to use some existing method, which is not<br />
simple. This would convey the problem <strong>of</strong> choosing the optimal action <strong>of</strong> each state.<br />
Finally, the action model can modify to involve more neighboring pixels in the<br />
processing window, not only making decision by the maximum and minimum gray-<br />
scale value <strong>of</strong> each region in the window but also getting help from neighboring pixels.<br />
102
Another future work is an improvement <strong>of</strong> PMAM; consequently, two ideas were<br />
inspired. First, the Multi-Agent model should modify to a comprehensive one; the agent<br />
properties mentioned in chapter 2.5 should employ to improve the model. In addition,<br />
the finding <strong>of</strong> optimal input data is a time-consuming work; therefore, the model should<br />
be less dependent to the input data.<br />
103
Bibliography<br />
Alan, L., T. Frank, et al. (2003). A Survey <strong>of</strong> Surgical Simulation: Applications,<br />
Technology, and Education. Presence: Teleoperators and Virtual Environments,<br />
12(6), 599-614.<br />
Awcock, G. J., Thomas, R. (1995). Applied Image Processing: McGraw-Hill.<br />
Bankman, I. (2008). Handbook <strong>of</strong> Medical Image Processing and Analysis: Academic<br />
Press.<br />
Bathe, K. J. (1996). Finite Element Procedures: Prentice-Hall.<br />
Benamrane, N., Nassane, S. (2007). Medical Image Segmentation by a Multi-Agent<br />
System Approach. Proceedings <strong>of</strong> the 5th German Conference on Multiagent<br />
System Technologies, Leipzig, Germany, Springer-Verlag.<br />
Bettega, G., Y. Payan, et al. (2000). A Simulator for Maxill<strong>of</strong>acial Surgery Integrating<br />
3D Cephalometry and Orthodontia. Journal <strong>of</strong> Image Guided Surgery, 5(3), 156-<br />
165.<br />
Bhanu, B., Peng, J. (2000). Adaptive Integrated Image Segmentation and Object<br />
Recognition. IEEE Transaction on Systems, Man, and Cybernetics, 30(4), 427-<br />
441.<br />
Boucher, A., Garbay, C. (1996). A Multi-Agent System to Segment Living Cells.<br />
Proceedings <strong>of</strong> the 13th International Conference on Pattern Recognition, 3,<br />
558-562.<br />
Boucher, A., A. Doisy, et al. (1998). A Society <strong>of</strong> Goal-Oriented Agents for the<br />
Analysis <strong>of</strong> Living Cells. Artificial Intelligence in Medicine, 14(1-2), 183-196.<br />
Bovenkamp, E. G. P., J. Dijkstra, et al. (2004). Multi-Agent Segmentation <strong>of</strong> IVUS<br />
images. Pattern Recognition Society, 37, 647-653.<br />
Chen, P., Pavlidis, T. (1980). Image Segmentation as an Estimation Problem. Computer<br />
Graphics and Image Processing, 12(2), 153.<br />
<strong>Chitsaz</strong>, M., Woo, C.S. (2008). The Rise <strong>of</strong> Multi-Agent and R.L. Segmentation Methods<br />
for Biomedical Images. The 4th Malaysian S<strong>of</strong>tware Engineering Conference<br />
(MySEC’08), Kuala Terengganu, Malaysia.<br />
<strong>Chitsaz</strong>, M., Woo, C.S. (2009). Medical Image Segmentation by using Reinforcement<br />
Learning Agent. International Conference on Digital Image Processing<br />
(ICDIP'09), Bangkok, Thailand, IEEE Computer Society Press.<br />
<strong>Chitsaz</strong>, M., Woo, C.S. (2009). A Multi-Agent System Approach for Medical Image<br />
Segmentation. International Conference on Future Computer and<br />
104
Communication (ICFCC'09), Kuala Lumpur, Malaysia, IEEE Computer Society<br />
Press.<br />
Coleman, G. B., Andrews, H.C. (1979). Image Segmentation by Clustering.<br />
Proceedings <strong>of</strong> the IEEE, 67(5), 773-795.<br />
Crevier, D., Lepage, R. (1997). Knowledge-Based Image Understanding Systems: A<br />
Survey. Computer Vision and Image Understanding, 67(2), 161-185.<br />
DICOMsample. "DICOM Files." Retrieved July 2008, from<br />
http://pubimage.hcuge.ch:8080/.<br />
Duchesnay, E., J. J. Montois, et al. (2001). An Agent-based Implementation <strong>of</strong> Irregular<br />
Pyramid for distributed Image Segmentation. Proceedings <strong>of</strong> Eighth<br />
International Conference on Emerging Technologies and Factory, 509-504.<br />
Duchesnay, E., J. J. Montois, et al. (2003). Cooperative Agents Society Organized as an<br />
Irregular Pyramid: a Mammography Segmentation Application. Pattern<br />
Recognition Letters, 24(14), 2435-2445.<br />
Felippa, C. (2007). "Introduction to Finite Element Methods " Retrieved March 2008,<br />
from http://www.colorado.edu/engineering/CAS/courses.d/IFEM.d/.<br />
Fenster, A., Chiu, B. (2005). Evaluation <strong>of</strong> Segmentation algorithms for Medical<br />
Imaging. 27th Annual International Conference <strong>of</strong> the Engineering in Medicine<br />
and Biology Society.<br />
Freixenet, J., X. Mu˜noz, et al. (2002). Yet Another Survey on Image Segmentation:<br />
Region and Boundary Information Integration. Proceedings <strong>of</strong> the 7th European<br />
Conference on Computer Vision-Part III, Springer-Verlag.<br />
Garrobo, A. (2006). "MRI Scanner." Retrieved January, 2009, from<br />
http://media.collegepublisher.com/media/paper871/stills/45359417825c3-89-<br />
2.jpg.<br />
Germond, L., M. Dojat, et al. (2000). A Cooperative Framework for Segmentation <strong>of</strong><br />
MRI Brain Scans. Artificial Intelligence in Medicine, 20(1), 77-93.<br />
Giraldi, G. A., P. S. Rodrigues, et al. (2006). Implicit Dual Snakes for Medical Imaging.<br />
the 28th Annual International Conference <strong>of</strong> the Engineering in Medicine and<br />
Biology Society.<br />
Gray, H. (1918). "Gray's Anatomy <strong>of</strong> Human Body." Retrieved January, 2009, from<br />
http://www.bartleby.com.<br />
Guillaud , A., H. Troadec, et al. (2000). Continuity Perception using a Multi-Agent<br />
System an Application to Growth Ring Detection on Fish Otoliths. The 15th<br />
International Conference on Pattern Recognition Proceedings. , 2, 519-521.<br />
Gyohten, K. (2000). Optimization-Based Image Analysis Dealing with Symbolic<br />
Constraints using Hierarchical Multi-Agent System. IEEE International<br />
Conference on Systems, Man, and Cybernetics, 4, 2794-2799.<br />
105
Hiatt, J. L., Gartner, L.P. (1982). Textbook <strong>of</strong> Head and Neck Anatomy. New York:<br />
Appleton-Century-Cr<strong>of</strong>ts.<br />
Jain, A. K. (1989). Fundamentals <strong>of</strong> Digital Image Processing Printice Hall.<br />
Jennings, N. R., K. Sycara, et al. (1998). A Roadmap <strong>of</strong> Agent Research and<br />
Development. International Journal <strong>of</strong> Autonomous Agents and Multi-Agent<br />
Systems, 1(1), 7-38.<br />
Jeri. (2008). "CT Scanner." Retrieved January, 2009, from<br />
www.smugpuppies.com/category/health/.<br />
Kaelbling, L. P., M. L. Littman, et al. (1996). Reinforcement Learning: A Survey.<br />
Artificial Intelligence Research, 4, 237-285.<br />
Kagawa, H., M. Kinouchi, et al. (1999). Image Segmentation by Artificial Life<br />
Approach using Autonomous Agents. International Joint Conference on Neural<br />
Networks.<br />
Khosla, R., C. Lai, et al. (2003). Optimising the Performance <strong>of</strong> S<strong>of</strong>t Computing Agents<br />
for Classification <strong>of</strong> Unstained Mammalian Cell Images. International<br />
Symposium on Computational Intelligence for Measurement Systems and<br />
Applications, 163-168.<br />
Kincaid , D., Cheney W. (2002). Numerical Analysis: Mathematics <strong>of</strong> Scientific<br />
Computing.<br />
Kirbas, C., Quek, F. K. H. (2003). Vessel Extraction Techniques and Algorithms : A<br />
Survey. The Third IEEE Symposium on BioInformatics and BioEngineering.<br />
Koch, R. M., S. H. M. Roth, et al. (2002). A Framework for Facial Surgery Simulation.<br />
Proceedings <strong>of</strong> the 18th Spring Conference on Computer graphics, Budmerice,<br />
Slovakia, ACM.<br />
Lai, C., R. Khosla, et al. (2003). GA Based Optimisation <strong>of</strong> a Multi-Agent S<strong>of</strong>t<br />
Computing Model for Segmentation and Classification <strong>of</strong> Unstained<br />
Mammalian Cell Images. The Congress on Evolutionary Computation, 2(7),<br />
1192.<br />
Liang, T., Rodriguez, J.J. (1996). MR Cranial Image Segmentation-A Morphological<br />
and Clustering approach. Proceeding <strong>of</strong> the IEEE Southwest Symposium on<br />
Image Analysis and Interpretation.<br />
Liu, J., Tang, Y.Y. (1999). Adaptive Image Segmentation with Distributed Behaviour-<br />
Based Agents. IEEE Transaction on Pattern Analysis and Machine Intelligence,<br />
21(6), 8.<br />
Liu, J. (2006). Synergistic hybrid Image Segmentation: Combining Model and Image-<br />
Based Startegies, <strong>University</strong> <strong>of</strong> Pennsylvania. PhD. <strong>Dissertation</strong>.<br />
106
Lu, H., Bao, S. (2006). An Extended Image Force Model <strong>of</strong> Snakes for Medical Image<br />
Segmentation and Smoothing. 8th International Conference on Signal<br />
Processing.<br />
Marieb, E. N. (2000). Essentials <strong>of</strong> Anatomy and Physiology: Addison Welsely<br />
Longman.<br />
Martini, F. H., C. W. Garrison, et al. (2004). Fundamental <strong>of</strong> Anatomy and Physiology:<br />
Pearson.<br />
Melkemi, K. E., M. Batouche, et al. (2004). MRF and Multi-Agent System bsed<br />
Approach for Image Segmentation. IEEE International Conference on Industrial<br />
Technology.<br />
Melkemi, K. E., M. Batouche, et al. (2005). Chaotic MultiAgent System Approach for<br />
MRF-based Image Segmentation. Proceedings <strong>of</strong> the 4th International<br />
Symposium on Image and Signal Processing and Analysis, 268-273.<br />
Melkemi, K. E., M. Batouche, et al. (2006). A Multi-Agent System Approach for Image<br />
Segmentation using Genetic Algorithms and Extremal Optimization Heuristics.<br />
Pattern Recognition Letters, 27, 1230-1238.<br />
Nikishov, G. P. (2007). Introduction to The Finite Element Method: McGraw-Hill<br />
Science/Engineering/Math.<br />
Obaidellah, U. H. B. (2006). A Finite Element Approach for the Planning and<br />
Simulation <strong>of</strong> 3D Mandibular Osteotomy for Orthoganathic Surgery. <strong>University</strong><br />
<strong>of</strong> <strong>Malaya</strong>. Master <strong>Dissertation</strong>.<br />
Pan, Z., Lu, J. (2007). A Bayes-Based Region-Growing Algorithm for Medical Image<br />
Segmentation. Computing in Science & Engineering, 9(4), 32-38.<br />
Peng, J., Bhanu, B. (1998 a). Delayed Reinforcement Learning for Adaptive Image<br />
Segmentation and Feature Extraction. IEEE Transactions on Systems, Man, and<br />
Cybernetics, Part C: Applications and Reviews, 28(3), 482-488.<br />
Peng, J. and B. Bhanu (1998 b). Closed-Loop Object Recognition Using Reinforcement<br />
Learning. IEEE Transactions on Pattern Analysis and Machine Intelligence,<br />
20(2), 139-154.<br />
Pham, D. L., C. Xu, et al. (2000). A Survey <strong>of</strong> Current Methods in Medical Image<br />
Segmentation. Annual Review <strong>of</strong> Biomedical Engineering, 2, 315-337.<br />
Rares, A., M. J. T. Reinders, et al. (1999). Mapping Image Analysis Problems on Multi-<br />
Agent Systems, Information and Communication Theory Group TU Delft.<br />
Richard, N., M. Dojat, et al. (2004). Automated Segmentation <strong>of</strong> Human Brain MR<br />
Images using a Multi-Agent Approach. Artificial Intelligence in Medicine,<br />
30(2), 24.<br />
107
Rodin, V., A. Benzinou, et al. (2004). An Immune Oriented Multi-Agent System for<br />
Biological Image Processing. Pattern Recognition Society, 37.<br />
Sahba, F., H. R. Tizhoosh, et al. (2006 a). Increasing Object Recognition Rate using<br />
Reinforced Segmentation. IEEE International Conference on Image Processing.<br />
Sahba, F., H. R. Tizhoosh, et al. (2006 b). A Reinforcement Learning Framework for<br />
Medical Image Segmentation. International Joint Conference on Neural<br />
Networks.<br />
Sahba, F., H. R. Tizhoosh, et al. (2007). Application <strong>of</strong> Opposition-Based<br />
Reinforcement Learning in Image Segmentation. IEEE Symposium on<br />
Computational Intelligence in Image and Signal Processing.<br />
Sahba, F., H. R. Tizhoosh, et al. (2008). A Reinforcement Agent for Object<br />
Segmentation in Ultrasound Images. Expert Systems with Applications, 35(3),<br />
772-780.<br />
Shariatpanahi, H. F., N. Batmanghelich, et al. (2006). Distributed behavior-based<br />
Multi-agent System for Automatic Segmentation <strong>of</strong> brain MR Images.<br />
International joint Conference on Neural network.<br />
Shokri, M. and H. R. Tizhoosh (2003). Using Reinforcement Learning for Image<br />
Thresholding. IEEE Canadian Conference on Electrical and Computer<br />
Engineering.<br />
Spinu, C., C. Garbay, et al. (1996). A Multi-Agent Approach to Edge Detection as a<br />
Distributed Optimization Problem. Proceedings <strong>of</strong> the 13th International<br />
Conference on Pattern Recognition.<br />
Tanasoca, R. C. M. (2006). Finite Element Modeling <strong>of</strong> the Anthropoid Mandible:<br />
Mandible Model, Experimental Validation, and Anthropologic Application,<br />
<strong>University</strong> <strong>of</strong> Florida. PhD. <strong>Dissertation</strong>.<br />
Udupa, J. K., V. R. LeBlanc, et al. (2006). A Framework for Evaluating Image<br />
Segmentation Algorithms. Computerized Medical Imaging and Graphics, 30(2),<br />
75-87.<br />
Umbaugh, S. E. (1998). Computer Vision and Image Processing: A Practical Approach<br />
using Cviptools: Prentice Hall.<br />
Valvassori, G. E., M. F. Mafee, et al. (1995). Imaging <strong>of</strong> the Head and Neck. Stuttgart:<br />
Georg Thieme Verlag.<br />
Walter, J. B., J. J. Carr, et al. (2007). Basic Atlas <strong>of</strong> Sectional Anatomy with Correlated<br />
Imaging: Saunders.<br />
Wang, Y., Yuan, B. (2000). Face Detection Using Evolutionary agents. 5th<br />
International Conference on Signal Processing Proceedings.<br />
108
Wang, Y., Yuan, B. (2002 a). Fast Method for Face Location and Tracking by<br />
Distributed Behaviour-Based Agents. Vision, Image and Signal Processing,<br />
149(3), 173- 178.<br />
Wang, Y., Yuan, B. (2002 b). Face Location and Tracking by Distributed Behaviour-<br />
Based Agents. 6th International Conference on Signal Processing Proceedings.<br />
Watkins, C. J. C. H. (1989). Learning from Delayed Rewards. Cambridge, UK,<br />
<strong>University</strong> <strong>of</strong> Cambridge. PhD <strong>Dissertation</strong>.<br />
Withey, D. (2006). Dynamic Edge Tracing: Recurcive Methods for Medical Image<br />
Segmentation, <strong>University</strong> <strong>of</strong> Alberta. PhD <strong>Dissertation</strong>.<br />
Wooldridge, M. (1997). Agent-Based S<strong>of</strong>tware Engineering. IEE Proceedings on<br />
S<strong>of</strong>tware Engineering.<br />
Wooldridge, M. J., Jennings, N. R. (1995). Intelligent Agents: Theory and Practice The<br />
Knowledge Engineering Review, 10(2), 115-152.<br />
Zhang, Y. J. (2001). A Review <strong>of</strong> Recent Evaluation Methods for Image Segmentation.<br />
Sixth International Symposium on Signal Processing and its Applications.<br />
109
Appendix A<br />
Experimental Results <strong>of</strong> the Training Phase - the first data set<br />
Slice No. 0 Slice No. 3<br />
Slice No. 10 Slice No. 13<br />
Slice No. 15 Slice No. 20<br />
Slice No. 23 Slice No. 28<br />
Slice No. 30 Slice No. 33<br />
110
Slice No. 35 Slice No. 38<br />
Slice No. 40 Slice No. 45<br />
Slice No. 47 Slice No. 50<br />
Slice No. 53 Slice No. 55<br />
Slice No. 57 Slice No. 60<br />
111
Experimental Results <strong>of</strong> the Training Phase - the Second data set<br />
Slice No. 253 Slice No. 260<br />
Slice No. 263 Slice No. 269<br />
Slice No. 271 Slice No. 278<br />
Slice No. 284 Slice No. 287<br />
Slice No. 290 Slice No. 301<br />
112
Slice No. 310 Slice No. 319<br />
Slice No. 330 Slice No. 350<br />
Slice No. 360<br />
113
Appendix B<br />
TPVF and FPVF <strong>of</strong> the Experimental Results <strong>of</strong> the Training Phase - the first data<br />
set<br />
Image No.<br />
TPVF (%) FPVF (%)<br />
BG Skin Bone BG Skin Bone<br />
0 99.99 98.29 96.43 0.34 0.22 0.37<br />
3 99.99 97.44 98.34 0.36 0.1 0.56<br />
13 99.99 95.76 94.74 0.45 0.37 0.92<br />
15 99.99 98.16 98.85 0.72 0.08 0.22<br />
20 99.99 95.97 94.65 0.37 0.37 0.89<br />
23 99.99 97.55 97.66 0.55 0.18 0.47<br />
28 99.97 97.99 96.44 0.2 0.34 0.53<br />
30 99.99 97.32 94.08 0.19 0.57 0.78<br />
33 99.99 98.61 97.52 0.1 0.26 0.42<br />
35 99.99 98.83 97.87 0.13 0.22 0.34<br />
40 99.95 98.3 97.41 0.3 0.29 0.46<br />
45 99.94 98.63 96.4 0.26 0.32 0.41<br />
47 99.92 98.7 95.92 0.33 0.33 0.39<br />
50 99.86 96.85 96.43 0.41 0.63 0.91<br />
53 99.73 95.95 97.12 0.43 0.66 1.14<br />
55 99.92 97.8 96.53 0.12 0.84 0.55<br />
57 99.87 97.45 96.48 0.54 0.54 0.75<br />
60 99.86 97.79 97.11 0.37 0.46 0.75<br />
114
TPVF and FPVF <strong>of</strong> the Experimental Results <strong>of</strong> the Training Phase - the second<br />
data set<br />
Image No.<br />
TPVF (%) FPVF (%)<br />
BG Skin Bone BG Skin Bone<br />
253 99.9 99.27 97.23 0.25 0.21 0.12<br />
260 99.99 99.28 96 0.1 0.21 0.14<br />
263 99.96 99.69 97.29 0.14 0.14 0.05<br />
269 99.88 99.74 94.55 0.17 0.19 0.06<br />
271 99.96 99.65 94.95 0.15 0.12 0.07<br />
278 99.99 99.63 96.5 0.11 0.11 0.05<br />
284 99.99 99.74 96.52 0.11 0.09 0.03<br />
287 99.98 99.52 92.41 0.16 0.22 0.06<br />
290 99.99 99.61 95.39 0.17 0.12 0.04<br />
301 99.99 99.69 96.8 0.18 0.09 0.04<br />
310 99.98 99.56 96.1 0.28 0.19 0.08<br />
319 99.91 99.65 96.69 0.2 0.19 0.11<br />
330 99.99 99.6 97.19 0.15 0.21 0.07<br />
350 99.95 99.38 95.6 0.14 0.37 0.15<br />
360 99.97 99.46 98.29 0.27 0.12 0.08<br />
115
Appendix C<br />
Experimental Results <strong>of</strong> the Testing Phase - the first data set<br />
Slice No. 0 Slice No. 3<br />
Slice No. 13 Slice No. 23<br />
Slice No. 30 Slice No. 35<br />
Slice No. 40 Slice No. 47<br />
Slice No. 50 Slice No. 55<br />
116
Slice No. 57 Slice No. 60<br />
Experimental Results <strong>of</strong> the Testing Phase - the second data set<br />
Slice No. 253 Slice No. 260<br />
Slice No. 271 Slice No. 278<br />
Slice No. 284 Slice No. 290<br />
Slice No. 301 Slice No. 310<br />
117
Slice No. 330 Slice No. 360<br />
118
Appendix D<br />
TPVF and FPVF <strong>of</strong> the Experimental Results <strong>of</strong> the Testing Phase - the first data set<br />
Image No.<br />
TPVF (%) FPVF (%)<br />
BG Skin Bone BG Skin Bone<br />
0 99.97 98.95 94.87 0.18 0.34 0.23<br />
3 99.34 97.97 93.54 1.23 0.73 0.34<br />
13 98.55 93.43 92.13 3.08 0.95 1.29<br />
15 98.68 92.07 89.94 2.86 0.89 1.75<br />
23 98.54 97.57 91.94 0.67 1.35 0.87<br />
28 98.62 91.94 97.76 1.64 0.81 97.76<br />
30 98.60 91.26 91.70 1.63 1.25 2.68<br />
35 99.05 98.37 90.09 0.44 1.70 0.46<br />
40 99.66 98.48 90.78 0.52 1.133 0.38<br />
47 99.71 94.05 88.62 0.63 1.04 1.96<br />
50 99.40 95.91 91.86 0.00 1.95 1.32<br />
53 98.99 93.73 89.53 0.00 3.12 1.83<br />
55 99.68 94.60 91.02 067 1.74 1.53<br />
57 99.73 97.51 91.11 0.72 1.46 0.68<br />
60 98.70 96.10 89.60 0.04 2.41 1.69<br />
119
TPVF and FPVF <strong>of</strong> the Experimental Results <strong>of</strong> the Testing Phase - the second data<br />
set<br />
Image No.<br />
TPVF (%) FPVF (%)<br />
BG Skin Bone BG Skin Bone<br />
253 98.41 0 98.17 1.9 91.61 0.43<br />
260 99.35 0.06 98.12 1.39 85.66 0.4<br />
263 99.28 0.01 98.41 1.2 88.14 0.33<br />
269 98.96 0.15 97.86 1.33 84.64 0.43<br />
271 98.8 0.12 97.93 1.43 86.03 0.43<br />
278 99.99 0.12 99.73 0.33 88.25 0.03<br />
284 99.99 0.19 99.54 0.27 88.06 0.05<br />
290 100 0.49 99.07 0.41 82.82 0.08<br />
301 99.96 0.51 98.61 0.51 83.48 0.22<br />
310 99.98 0.57 98.95 0.84 84.94 0.13<br />
319 99.91 0.65 99.27 1.3 79.3 0.02<br />
330 99.99 0.18 99.51 1.5 80.19 0.09<br />
360 99.92 0.24 99.68 1.27 84.05 0.02<br />
120