MICROWAVE-BASED MEDICAL DIAGNOSIS USING PARTICLE SWARM OPTIMIZATION ALGORITHM by Arezoo Modiri APPROVED BY SUPERVISORY COMMITTEE: Kamran Kiasaleh, Chair Lakshman S. Tamil Randall E. Lehmann Rashaunda M. Henderson c 2013 Copyright Arezoo Modiri All rights reserved This Dissertation is dedicated to my beloved parents and brother. MICROWAVE-BASED MEDICAL DIAGNOSIS USING PARTICLE SWARM OPTIMIZATION ALGORITHM by AREZOO MODIRI, BS, MS DISSERTATION Presented to the Faculty of The University of Texas at Dallas in Partial Fulfillment of the Requirements for the Degree of DOCTOR OF PHILOSOPHY IN ELECTRICAL ENGINEERING THE UNIVERSITY OF TEXAS AT DALLAS December 2013 UMI Number: 3606254 All rights reserved INFORMATION TO ALL USERS The quality of this reproduction is dependent upon the quality of the copy submitted. In the unlikely event that the author did not send a complete manuscript and there are missing pages, these will be noted. Also, if material had to be removed, a note will indicate the deletion. UMI 3606254 Published by ProQuest LLC (2013). Copyright in the Dissertation held by the Author. Microform Edition © ProQuest LLC. All rights reserved. This work is protected against unauthorized copying under Title 17, United States Code ProQuest LLC. 789 East Eisenhower Parkway P.O. Box 1346 Ann Arbor, MI 48106 - 1346 ACKNOWLEDGMENTS As the author, it is almost impossible to express gratitude in mere words for the wonderful people who have helped and supported me during my doctoral journey. I feel blessed for having every single one of them in my life. First and foremost, I would like to thank my PhD advisor, Professor Kamran Kiasaleh, for his continuous guidance and support. For me, it was an amazing opportunity to get engaged in several distinct, thrilling, professional activities during my PhD, and Professor Kiasaleh’s leading role as a thoughtful mentor was indeed a significant factor for making these opportunities become reachable. In addition, I am very grateful to professors Lakshman S. Tamil, Randall E. Lehmann and Rashaunda M. Henderson for their persistent, considerate support throughout my PhD and also for serving in my dissertation committee. I am also very thankful to Dr. Robert Glosser and Dr. Ricardo Saad for their sincere, continuous help and Dr. Dennis Miller for his great help to my research project. I have learned much more than course material from my UTD professors, especially those for whom I have served as teacher assistant. Their friendly and meanwhile professional presence in my PhD life has thought me invaluable lessons. I should thank the friendly and thoughtful UTD staff, as well. They have been helping me out with every single step during my education. For sure, being with talented faculty and staff in UTD has been a rewarding experience and wonderful learning opportunity for me, and unfortunately, it is not possible to list the names of all people I want to thank in this page. I am specially thankful to my lab mates, Ali Montazeri, Gaurav Sureka, Anish Nair, Zhengjie Liu, Raffat Khan, Xiajiang Tian, Allen Webb, Damin Cao, Michael Cole and Jaskaran Sing v who have made my PhD journey much more pleasant and memorable. In addition, I am grateful to my very good friends Miad Faezipour and Soudeh Khoubrouy, and of course, to many other friends, the names of whom cannot be mentioned herein. Good friends make life full of blessing. Finally, I would like to emphasize that my educational endeavors have always been strongly supported by my dearly loved parents and brother in any possible way. They have been constantly encouraging me and helping me with the challenges of my life as an international student. No words can express my gratefulness to them. July 2013 vi PREFACE This dissertation was produced in accordance with guidelines which permit the inclusion as part of the dissertation the text of an original paper or papers submitted for publication. The dissertation must still conform to all other requirements explained in the “Guide for the Preparation of Master’s Theses and Doctoral Dissertations at The University of Texas at Dallas.” It must include a comprehensive abstract, a full introduction and literature review, and a final overall conclusion. Additional material (procedural and design data as well as descriptions of equipment) must be provided in sufficient detail to allow a clear and precise judgment to be made of the importance and originality of the research reported. It is acceptable for this dissertation to include as chapters authentic copies of papers already published, provided these meet type size, margin, and legibility requirements. In such cases, connecting texts which provide logical bridges between different manuscripts are mandatory. Where the student is not the sole author of a manuscript, the student is required to make an explicit statement in the introductory material to that manuscript describing the student’s contribution to the work and acknowledging the contribution of the other author(s). The signatures of the Supervising Committee which precede all other material in the dissertation attest to the accuracy of this statement. vii MICROWAVE-BASED MEDICAL DIAGNOSIS USING PARTICLE SWARM OPTIMIZATION ALGORITHM Publication No. Arezoo Modiri, PhD The University of Texas at Dallas, 2013 Supervising Professor: Kamran Kiasaleh This dissertation proposes and investigates a novel architecture intended for microwave-based medical diagnosis (MBMD). Furthermore, this investigation proposes novel modifications of particle swarm optimization algorithm for achieving enhanced convergence performance. MBMD has been investigated through a variety of innovative techniques in the literature since the 1990’s and has shown significant promise in early detection of some specific health threats. In comparison to the X-ray- and gamma-ray-based diagnostic tools, MBMD does not expose patients to ionizing radiation; and due to the maturity of microwave technology, it lends itself to miniaturization of the supporting systems. This modality has been shown to be effective in detecting breast malignancy, and hence, this study focuses on the same modality. A novel radiator device and detection technique is proposed and investigated in this dissertation. As expected, hardware design and implementation are of paramount importance in such a study, and a good deal of research, analysis, and evaluation has been done in this regard which will be reported in ensuing chapters of this dissertation. It is noteworthy that an important element of any detection system is the algorithm used for viii extracting signatures. Herein, the strong intrinsic potential of the swarm-intelligence-based algorithms in solving complicated electromagnetic problems is brought to bear. This task is accomplished through addressing both mathematical and electromagnetic problems. These problems are called benchmark problems throughout this dissertation, since they have known answers. After evaluating the performance of the algorithm for the chosen benchmark problems, the algorithm is applied to MBMD tumor detection problem. The chosen benchmark problems have already been tackled by solution techniques other than particle swarm optimization (PSO) algorithm, the results of which can be found in the literature. However, due to the relatively high level of complexity and randomness inherent to the selection of electromagnetic benchmark problems, a trend to resort to oversimplification in order to arrive at reasonable solutions has been taken in literature when utilizing analytical techniques. Here, an attempt has been made to avoid oversimplification when using the proposed swarm-based optimization algorithms. To summarize the outline of this dissertation, the first chapter introduces the different elements of the comprehensive study which was explained very concisely in the above paragraph. Chapters 2 and 3 present the details of the proposed solutions for the benchmark problems and the MBMD problem, respectively. The remaining chapters are the re-productions of the outcome of this dissertation, which have been published (or submitted for publication) by the candidate and her supervisor during her PhD in scholarly, refereed journals and conference proceedings. The first chapter is mostly taken from their published book chapter plus afterward achievement. ix TABLE OF CONTENTS ACKNOWLEDGMENTS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . PREFACE v . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii ABSTRACT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viii LIST OF FIGURES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xii LIST OF TABLES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xv CHAPTER 1 INTRODUCTION . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 1.1 Particle swarm optimization; development, modifications . . . . . . . . . . . 6 1.1.1 Conventional PSO . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6 1.1.2 PSO variations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10 1.2 Case study 1: PSO in antenna beam forming . . . . . . . . . . . . . . . . . . 16 1.3 Case study 2; PSO in frequency selectivity . . . . . . . . . . . . . . . . . . . 18 1.4 Case study 3: PSO in bend effect compensation . . . . . . . . . . . . . . . . 22 1.5 Breast cancer detection using PSO in MBMD . . . . . . . . . . . . . . . . . 27 1.5.1 Modeling and estimation methodology . . . . . . . . . . . . . . . . . 31 1.5.2 Implementation of microwave diagnosis tool . . . . . . . . . . . . . . 36 1.5.3 Preliminary results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40 CHAPTER 2 PROPOSED PSO IN ANTENNA DESIGN . . . . . . . . . . . . . . . 44 2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44 2.2 PSO in thinned array beam forming1 . . . . . . . . . . . . . . . . . . . . . . 44 2.2.1 Evaluation of the proposed velocity function in RPSO . . . . . . . . . 46 2.2.2 Evaluation of the proposed velocity function in thinned array design . 55 PSO in antenna frequency selectivity2 . . . . . . . . . . . . . . . . . . . . . . 62 2.3.1 Switchable antenna design using modified BPSO . . . . . . . . . . . . 64 2.3.2 Optimization length comparison summary . . . . . . . . . . . . . . . 69 PSO in wearable antenna frequency selectivity3 . . . . . . . . . . . . . . . . 71 2.3 2.4 x 2.5 Modeling antenna bend effect . . . . . . . . . . . . . . . . . . . . . . . . . . 72 2.6 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76 CHAPTER 3 PROPOSED MBMD ARRANGEMENT . . . . . . . . . . . . . . . . 81 PSO in solving the inverse problem4 . . . . . . . . . . . . . . . . . . . . . . . 82 3.1.1 Problem scope and estimation algorithm . . . . . . . . . . . . . . . . 83 3.1.2 Parameter variation study . . . . . . . . . . . . . . . . . . . . . . . . 87 MBMD proposed radiator design5 . . . . . . . . . . . . . . . . . . . . . . . . 92 3.2.1 The array structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93 3.2.2 Analysis results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94 Adding a conductor cover . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97 3.1 3.2 3.3 3.4 3.3.1 Proposed structure and 3D model . . . . . . . . . . . . . . . . . . . . 101 3.3.2 Detection mechanism and analysis . . . . . . . . . . . . . . . . . . . . 104 Experimental results6 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110 3.4.1 Measurement setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114 3.4.2 Results and analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . 118 3.4.3 Comparing with a competing MBMD design . . . . . . . . . . . . . . 121 3.4.4 Calibration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123 CHAPTER 4 CONCLUSIONS AND FUTURE RESEARCH DIRECTIONS . . . . 132 REFERENCES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134 VITA xi LIST OF FIGURES 1.1 Fitness values of the four introduced methods averaged over one thousand runs . 15 1.2 Probability density functions of the final fitness values . . . . . . . . . . . . . . 15 1.3 Fitness value trends of the introduced methods averaged over 700 runs . . . . . 18 1.4 The fifteen and twenty segment microstrip antennas . . . . . . . . . . . . . . . . 21 1.5 Average fitness value trend and the best optimization results . . . . . . . . . . . 24 1.6 Measurement arrangements for the wearable antenna . . . . . . . . . . . . . . . 25 1.7 |S11 | measurement results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27 1.8 Simplified two dimensional model . . . . . . . . . . . . . . . . . . . . . . . . . . 35 1.9 Average fitness function trends for the two scenarios. . . . . . . . . . . . . . . . 36 1.10 Comparison of the actual values with the estimated values . . . . . . . . . . . . 37 1.11 The preliminary radiator structure . . . . . . . . . . . . . . . . . . . . . . . . . 39 1.12 Digital breast phantom along with the proposed radiator structure. . . . . . . . 39 1.13 Cumulative magnitude difference of reflection coefficient . . . . . . . . . . . . . 42 1.14 Cumulative magnitude difference of transmittance coefficient . . . . . . . . . . . 43 2.1 AIN behaviour of the four methods . . . . . . . . . . . . . . . . . . . . . . . . . 48 2.2 AFV behaviour of the four methods . . . . . . . . . . . . . . . . . . . . . . . . . 49 2.3 AIN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50 2.4 AFV . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51 2.5 Fitness value of the four methods averaged over 1000 runs . . . . . . . . . . . . 52 2.6 Average sample particle position . . . . . . . . . . . . . . . . . . . . . . . . . . 53 2.7 Fitness value of the four methods averaged over 500 runs . . . . . . . . . . . . . 54 2.8 Fitness value of the four methods averaged over 500 runs with 30 particles . . . 55 2.9 Symmetric half-wavelength thinned linear array. . . . . . . . . . . . . . . . . . . 57 2.10 AIN behaviour in a 10D problem . . . . . . . . . . . . . . . . . . . . . . . . . . 58 2.11 AFV behaviour in a 10D problem . . . . . . . . . . . . . . . . . . . . . . . . . . 58 xii 2.12 Non-symmetric thinned planar antenna array. . . . . . . . . . . . . . . . . . . . 59 2.13 AIN and AFV behaviour in a 9D problem . . . . . . . . . . . . . . . . . . . . . 60 2.14 Probability density function in a 9D problem . . . . . . . . . . . . . . . . . . . . 61 2.15 Fitness value trend in a 9D problem . . . . . . . . . . . . . . . . . . . . . . . . 61 2.16 Fitness value trend over 500 runs . . . . . . . . . . . . . . . . . . . . . . . . . . 62 2.17 15-segment reconfigurable microstrip antenna structure. . . . . . . . . . . . . . . 65 2.18 Probability density function of BPSO algorithms . . . . . . . . . . . . . . . . . 66 2.19 Final fitness value of BPSO algorithms . . . . . . . . . . . . . . . . . . . . . . . 67 2.20 Particle position for BPSO algorithms . . . . . . . . . . . . . . . . . . . . . . . 68 2.21 Final fitness value of BPSO algorithms sing 4 particles . . . . . . . . . . . . . . 69 2.22 Measurement arrangement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74 2.23 |S11 | measurement results for the flat and bent cases. . . . . . . . . . . . . . . . 74 2.24 Reconfigurable antenna with (a)three and (b) eight switches. . . . . . . . . . . . 75 2.25 |S11 | measurement results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76 2.26 Average fitness value trend . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79 2.27 The final fitness value of each run . . . . . . . . . . . . . . . . . . . . . . . . . . 80 3.1 The block diagram of the proposed PSO . . . . . . . . . . . . . . . . . . . . . . 85 3.2 Mapping function with different smoothing coefficients. . . . . . . . . . . . . . . 86 3.3 The effect of immersion medium . . . . . . . . . . . . . . . . . . . . . . . . . . . 88 3.4 Estimation error . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89 3.5 The error for thicker tissue . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89 3.6 Estimation error assuming different maximum velocities/smoothing coefficients . 91 3.7 Estimation error assuming different number of agents. . . . . . . . . . . . . . . . 91 3.8 The Ansoft digital phantom breast tissue and the antenna array. . . . . . . . . . 94 3.9 The reflection coefficient of the five elements of the array. . . . . . . . . . . . . . 95 3.10 The radiation pattern of the radiator . . . . . . . . . . . . . . . . . . . . . . . . 97 3.11 The polarization-dependent radiation pattern of the radiator . . . . . . . . . . . 98 3.12 The electric field distribution in tissue . . . . . . . . . . . . . . . . . . . . . . . 99 3.13 The transmittance coefficient of the array elements. . . . . . . . . . . . . . . . . 100 xiii 3.14 nsoft female body phantom . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103 3.15 Digital breast phantom along with the proposed radiator structure. . . . . . . . 104 3.16 Tumors’ shapes and locations inside the breast. . . . . . . . . . . . . . . . . . . 107 3.17 The normalized differences of |Ex | . . . . . . . . . . . . . . . . . . . . . . . . . . 108 i i 3.18 |End | and arg(End ) for the superficial spherical tumor . . . . . . . . . . . . . . . 110 i i 3.19 |End | and arg(End ) for the deep spherical tumor . . . . . . . . . . . . . . . . . . 111 i i 3.20 |End | and arg(End ) for the deep small spherical tumor . . . . . . . . . . . . . . . 112 i i 3.21 |End | and arg(End ) for the superficial small spherical tumor . . . . . . . . . . . 113 i i 3.22 |End | and arg(End ) for the superficial tiny spherical tumor . . . . . . . . . . . . 114 i i 3.23 |End | and arg(End ) for the deep tiny spherical tumor . . . . . . . . . . . . . . . 115 i i 3.24 |End | and arg(End ) for the superficial cylindrical tumor. . . . . . . . . . . . . . . 116 i i 3.25 |End | and arg(End ) for the deep cylindrical tumor. . . . . . . . . . . . . . . . . . 117 i i 3.26 |End | and arg(End ) for the superficial gland-shaped tumor. . . . . . . . . . . . . 118 i i 3.27 |End | and arg(End ) for the spherical tumor over 101 receiving points. . . . . . . 119 i i 3.28 |End | and arg(End ) for the cylindrical tumor over 101 receiving points. . . . . . 120 3.29 The creation process of the two breast phantoms . . . . . . . . . . . . . . . . . 121 3.30 The measurement setup using an ENA. . . . . . . . . . . . . . . . . . . . . . . . 125 3.31 The measured |S11 | of the sixteen antennas . . . . . . . . . . . . . . . . . . . . . 126 3.32 Cumulative phase difference of reflection coefficient . . . . . . . . . . . . . . . . 126 3.33 Cumulative phase difference of transmittance coefficient . . . . . . . . . . . . . . 127 3.34 The MBMD design from Duke University and their phantom. . . . . . . . . . . 128 3.35 Comparing the phantom measurement results . . . . . . . . . . . . . . . . . . . 129 3.36 Comparing signatures achieved from tumor with misplacement . . . . . . . . . . 130 3.37 Digital breast phantom modeling misplacement. . . . . . . . . . . . . . . . . . . 130 3.38 Tumor signature and movement effect on adjacent antennas. . . . . . . . . . . . 131 xiv LIST OF TABLES 1.1 Optimization cost . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34 2.1 Required optimization time for a single 25-iteration run in minutes. . . . . . . . 70 3.1 Optimization length comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . 92 xv CHAPTER 1 INTRODUCTION1 Authors - Arezoo Modiri and Kamran Kiasaleh Electrical Engineering Department The University of Texas at Dallas 800 West Campbell Road Richardson, Texas 75080-3021 1 This chapter is adapted from chapter 5 in <Swarm Intelligence for Electric and Electronic Engineering> c 2013, IGI Global, www.igi-global.com. edited by <Girolamo Fornarelli and Luciano Mescia> Copyright Posted by permission of the publisher. 1 2 Microwave-based medical diagnosis (MBMD) is a tremendously complicated electromagnetic problem both hardware wise and software wise. This study reports the promising results of a preliminary solution to this complex problem. For the sake of clarity, the author prefers to start with explaining the proposed algorithm, its variants and its responses to various problems, and, then, proceed with the radiator design, implementation aspects and measurement results. A quick survey of the literature indicates that distinct algorithms have been selected to solve MBMD problem; however, a reasonable approach is to test the chosen algorithm in some known problems prior to introducing it to the main unknown one. This chapter underscores the flexibility and the power of the particle swarm optimization (PSO) algorithm in handling a variety of analytically intractable problems in order to pave the path for delving into the foremost objective MBMD one. To that end, specific scenarios, which are representatives of a larger class of problems, are explored using PSO. The chapter begins by providing a detailed definition of PSO and its modified versions and then continues with the discussions regarding the challenges of solving realistic problems by means of this algorithm. In each case study, along with the problem definition, the key characteristics of the problem are brought to the forefront. This, in fact, tells the reader what other problems can be solved in the same or similar way. The conventional algorithm, then, is modified, exclusively to fit the requirements of that specific problem, and, finally, the performance of the resulting algorithm is presented and evaluated. Although the benchmark problems seem to be quite distinct from the MBMD one, they possess major common properties, which make a swarm-based algorithm a viable search mechanism for them. Those properties can be summarized as follows: • The inherent random nature of the key variables of the problem. • The large number of unknowns as compared to the available information. 3 • The necessity of expeditious solution discovery due to the ultimate real-time nature of the corresponding application. The outline of this chapter is as follows: In section 1.1, PSO algorithm is introduced, and several modifications to the conventional format of the algorithm are presented and compared in terms of the final result accuracy and optimization cost. In addition to the results presented in (Modiri, et. al., 2011b), this section brings together a summary of the other discoveries regarding PSO, which are presented in the literature, such as those introduced by (Nakano, et. al., 2007; Clerc, et. al,. 2002). A good list of the articles in this area can be found in the reference list of (Modiri, et. al., 2011b). When studying optimization algorithms, it is important to bear in mind that, in general, there are two approaches one may take. The first approach is dedicated to a suitable solution found expeditiously, whereas the second approach targets the best possible solution regardless of the optimization time or cost. One needs to compromise with respect to the features of the problem in hand. Both real-number PSO (RPSO) and binary PSO (BPSO) algorithms are taken into account. In order to rationalize the introduced modifications, wellknown mathematical benchmark problems are evaluated. This allows the author to set the stage for the more challenging problems of later sections. In section 1.2-1.4, based on the studies presented in (Modiri, et. al., 2011a; Modiri, et. al., 2011d; Modiri, et. al., 2010), the author concisely illustrates how PSO is utilized in a more challenging (and of course more appealing) scenarios. In particular, the design of reconfigurable and wearable antennas are considered (Liu, et. al., 2005; Jin, et. al., 2008). These are, in fact, the electromagnetic problems chosen to test PSO prior to introducing it to the main problem. In order to illustrate how PSO as a viable technique enters the picture, a brief description of the problem is presented in the next paragraphs. Reconfigurable antenna problems which are studied herein are useful in a variety of applications where software defined control over the antenna performance is required. Particularly, 4 the ability to change the frequency performance of the antenna (in terms of resonance frequency) is investigated. This feature is appropriate in multi-application devices where each application may have its own frequency assignment. Instead of incorporating multiple antennas for supporting individual applications, a single reconfigurable antenna can be designed. The advantage of using PSO algorithm in this problem becomes even more obvious when a level of unpredictability in the start time of an application is present in the overall design; see (Jin, et. al., 2008) as an example. In addition to frequency selectivity, the antenna radiation pattern can also be controlled in reconfigurable antennas, resulting in yet another demanding application (Gies, et. al., 2003). System designers are often faced with the problem of designing systems in areas which are jam-packed with wireless networks, such as hospitals. In such environments, there is a high chance of interference between adjacent communication links. Therefore, it is more convenient to be able to block the radiation in a certain direction. Here, PSO algorithm is used to control the radiation pattern of the array antennas with the goal of beam blocking, see (Jin, et. al., 2007). Another topic studied here is wearable antennas (WA). Mounting antennas on the fiber of daily clothing in the form of wearable antennas has recently attracted a great deal of interest. This is mainly due to the growing demand for body sensor networks. Sensors are dependent on the unobtrusive radiators in and around the human body in order to constantly trace bio signatures over the entire monitoring period. This monitoring state, in some cases, may continue for several months. However, it has been shown in the literature that various seemingly unnoticeable factors, including habitual activities of the user or common daily events can degrade the WA performance substantially, see chapter 6 in (Hall, et. al., 2006). More crucially, telecare devices are at risk when used for controlling the vital bio-signs of young, active children remotely. Hence, circumventing the performance degradation caused by the environmental variations can play an important role in preserving a high quality active monitoring. This task is done via applying a real-time, self-tuning algorithm. As a possible 5 variation, in section 1.4, we investigate the problem of bending the antenna and its effect on the antenna performance. A solution technique is introduced for real-time compensation of this phenomenon using PSO algorithm. The solution to this problem will also pave the way for solving similar problems with antenna modification. In the section 1.5, the usefulness of PSO in an even more challenging and attractive scenario is discussed. In that scenario, the algorithm is used to estimate the electrical properties of a lossy, multilayer structure. This study becomes more appealing when the medium under investigation is the human body tissue. The main application of this study in real world, as it is evident from the title of this dissertation, is detecting malignancy. The author’s research on breast cancer detection via microwave technology, partly published in (Modiri, et. al. 2011c, 2012, and 2013), is the main focus of this section. Dealing with live tissue has its intrinsic challenges; many articles in the literature concerning live tissue, e.g., (Lazebnik, et. al., 2007), have demonstrated the semi-randomness of the electric properties of human tissue. Taking advantage of these studies, in section 1.5, a PSO-based tumor detection method is reviewed, particularly for breast cancer detection through MBMD at frequencies below 3GHz. In fact, the proposed detection technique is not dependent on the type of the tissue under investigation, and hence, can be used in other similar applications as well. However, the limitation is imposed by the penetration depth of electromagnetic energy at the desired frequency (herein, microwave) in different tissues. In this scenario, both complex number PSO (CPSO) and discrete PSO (DPSO) algorithms are studied and redesigned in order to address the requirements of the problem at hand. The main goal is to demonstrate that PSO can successfully distinguish between the permittivity values of various layers of the human tissue, enabling one to verify the existence and the possible location of a tumor. It should be noted that, in addition to real number, binary, complex number and discrete PSO algorithms studied in this dissertation, other versions of this algorithm have also been introduced and utilized in distinct electromagnetic problems successfully in the literature. 6 Boolean PSO is an interesting example. Some more examples are introduced by Mahmoud (2008), Janson (2006), Garcia-Nieto (2011), Jin (2010), Cai (2009), Zhao (2010), Xu (2009), Ibrahim (2012) and Epitropakis (2012). However, those PSO variants are out of the scope of this manuscript and the interested reader is referred to the available articles in the literature, such as (Afshinmanesh, et. al., 2008). It is also noteworthy to mention that, different references do not necessarily use common abbreviations for distinct PSO variations. For example, some articles use BPSO as the abbreviation for Binary PSO, and some for Boolean PSO. Later in section 1.5, the radiator structure proposed for MBMD is briefly introduced along with the measurement setup and test environment. The proposed structure consists of an array of curved planar dipole antennas incorporated on the inner surface of a hemispherical dielectric body. The array is designed so that it creates complete radiation coverage all over the tissue under the test. It is also shown that adding a conductor reflector sheet improves the detection chance for most tumor cases and also acts as an electromagnetic shield. Digital female body phantom is used for the sake of creating accurate simulation ground. For measurement trials, however, tissue phantoms are built and actual breast tumor donated from a patient is placed inside the cancerous phantom. Finally, concluding remarks along with a summary of the potential of PSO and the conditions under which such potentials can be achieved are provided in the conclusion section of this chapter. It is important to emphasize that this chapter reviews all the aforementioned subjects in brief. More details are presented in chapters 2 and 3. 1.1 1.1.1 Particle swarm optimization; development, modifications Conventional PSO As it is explained by Das (2008), Menhas (2011), Liu (2009), Sedighizadeh (2009), Banks (2007) and many others, PSO is a member of the evolutionary algorithm (EA) family. The 7 idea of EAs originated from the concept of stochastic variation of a group of ’individuals’, also named ’particles’, in a solution space. Various algorithms have been developed based on this very concept since 1990’s. The main difference between the existing algorithms, however, resides in the manner by which their operators are defined. A thorough comparison of EA’s in electromagnetics was presented by Hoorfar (2007), where the following classifications of EA’s were also introduced: evolutionary programming, evolution strategies, and genetic algorithms. Each of these classes was introduced to address a specific group of problems for which such algorithm yields an effective solution. Therein, the author concluded that the evolutionary programming (EP) class was a very valuable tool for the efficient design of microwave devices due to its simplicity, ease of implementation, and flexibility. PSO algorithm belongs to EP class, and hence, the aforementioned statement applies to PSO as well. At this point, let us have a brief review of PSO history. Kennedy and Eberhart (1995) formulated the PSO algorithm, for the first time, by emulating the behavior of group-living creatures during their search activities. The algorithm was further developed by (Shi, et. al., 1999; Eberhart, et. al., 2001; Clerk, et. al., 2002), among others. As shown by Jin (2007), PSO, with the aid of a variety of analytical and numerical tools, becomes a viable technique for various electromagnetic applications. Here are some examples of using PSO; Matekovits (2005), Mikki(2005) and Robinson (2004) have studied how PSO generally can be used in electromagnetics, while Cui (2005), Jin (2005), Hwang (2009), Chamaani (2010) and Boeringer (2004) have focused on some specific applications, such as the design of absorbers, waveguides and antennas. Wang (2005), He (2009), Ko (2009) and Sun (2010) have used PSO in filter design. Benedetti (2006), Bayraktar (2006), Abu-Al-Nadi (2012) and Ismail (2010) have shown how PSO is utilized in array design, and of course, many more interesting articles can be found in the literature that use PSO variants in various electromagnetic applications. In spite of a relatively large number of articles on EAs, the design of EA algorithms, in 8 general, and PSO, in particular, remains problem dependent. Initial conditions also have a substantial impact on the performance, especially when optimization length is a significant concern. Eberhart (2001), Nakano (2007), and Camci (2008) effectively depicted the abovementioned properties of PSO in their investigations. In other words, it is always recommended to choose the optimization technique and set the parameters of the algorithm accordingly to meet the unique attributes of the problem at hand. The fundamentals of PSO algorithm have been extensively presented in the literature; for example, see (Eberhart, et. al., 2001; Jin, et. al. 2007). For the benefit of the reader, a concise, yet complete definition of the algorithm is provided here. Let us assume a problem with n unknowns or parameters. In fact, n defines the dimensionality of the problem at hand, which means that the solution space is n-dimensional. The optimum solution is a single point in this n-dimensional space. The ranges over which problem parameters vary define how wide the solution space is. A search group, consisting of m individuals, tries to find the optimum solution. Similar to other optimization techniques, a fitness (or objective) function determines how ’good’ each location in the solution space is. Although multiple objective functions can exist for one problem, it is often possible to merge all of them into one. Hence, single-objective optimization is studied herein. In PSO, each particle remembers the best location, in terms of fitness function, that it has discovered during its own exploration. This is named the ’personal best’. At each iteration cycle, the best location found by all particles (the best personal best) is named the ’global best’. Following steps describe the basic building blocks of the PSO algorithm: • The ’m’ particles are spread randomly inside the solution space. • Fitness values are calculated and personal bests and global best are identified. • With respect to the personal and global bests, a ’velocity’ vector is assigned to each particle. 9 • The locations of the particles are updated based on the velocity vectors. • The algorithm loops back to step 2 until the desired fitness value is achieved. It is important to note that the velocity vector is the only important operator in the PSO algorithm and this is the main reason for the ease of implementation of PSO. A definition for the velocity vector and the impact of its variation on the overall performance will be presented in the ensuing discussion. Having m particles and n parameters, the positions and velocities are stored in mn matrices. In the conventional PSO, the velocity is calculated using following equation. vt = ωvt−1 + c1 η1 (pbestt−1 − xt−1 ) + c2 η2 (gbestt−1 − xt−1 ). (1.1) In the above equation, xt , vt , pbest and gbest denote the position, velocity, personal best, and global best matrices, respectively. t is the iteration number. η1 and η1 are two random variables uniformly distributed on [0,1], and lastly, ω, c1 and c2 are constants used to move the algorithm towards achieving the desired goals. As Eberhart (2001) depicts, ω, c1 and c2 have been, in large part, obtained through trial and error. This also means that these parameters can be redesigned according to the problem at hand. In order to initiate Eq.(1.1), it is necessary to know the initial velocity. In the conventional PSO, the initial velocity matrix, v0 , is composed of the elements shown in Eq.(1.2). j j v0 (i, j) = random number unif ormly distributed on [vminimum , vmaximum ]. iǫ1, 2, ...m jǫ1, 2, ...n (1.2) 10 j j In the above, vmaximum and vminimum are the maximum and minimum permissible velocj j ities over j th dimension. Often, vminimum is assumed to be equal to vmaximum in intensity, j but in the opposite direction. vmaximum is generally assumed to be equal to the parameter range or a fraction of it, for RPSO. As shown by Jin (2007), a mapping function is required in the case of binary and discrete PSO’s for the sake of quantization. Another part of the algorithm is the boundary condition, which is necessary in order to set a bound on the limits of the solution space. Robinson (2004) has explained different boundary conditions and has claimed that ’invisible wall’ showed a superior performance. Therefore, invisible wall is mainly used in this chapter. Using this boundary condition, the fitness function calculation will not be performed for the particles that ’fly out’ of the solution space. A fairly undesired fitness value is assigned to these particles instead. This way, although all the particles remain intact, the computation complexity is decreased slightly. Upon the return of the escaped particles, the objective function calculation will be restarted for them. Yet as a second option, ’absorbing wall’ boundary condition is also taken into account in this section. An absorbing wall simply stops particles right at the boundary if they try to fly out of the solution space. 1.1.2 PSO variations When deciding to employ PSO in time-sensitive electromagnetic problems, there are two main issues to consider. First, similar to the other global optimization techniques, it is possible that the particles are trapped by a local best solution which, ultimately, prevents them from reaching the actual best solution. Second, in many demanding electromagnetic problems, the optimization cost can get unacceptably high due to the large number of required iterations and/or particles. Due to the aforementioned two facts, basic PSO has been modified in variety of ways to compensate for its drawbacks. A review of the different modifications existing in the literature 11 can be found in (Nakano, et. al.,2007) along with a comparison of their performances. There, the author has also introduced a new method, namely, tabu-searching PSO or TS-PSO. To briefly illustrate, in TS-PSO, a second swarm and a number of ’tabu’ conditions are added to the algorithm. Equations (5)-(7) in (Nakano, et. al., 2007) show that the particles of the second swarm are led away from the personal best and/or global best positions which satisfy the tabu condition. In this way, the particles are prevented from getting locked in their local bests. However, optimization length is shown to increase significantly in TS-PSO due to the increased computation complexity. More modified PSO algorithms are briefly introduced in the ensuing paragraphs in which personal best coefficient (c1 ) and initial velocity (v0 ) definition are particularly varied. The advantage of the following modifications resides in the non-increasing computation cost. Looking back at the steps followed in the PSO algorithm, one can observe that, at each iteration cycle, one of the personal bests is qualified to become the global best of that iteration. The other personal bests are actually undesired local bests, which can only be saved in the memory of the particles in order to guide them in their future trajectories. Therefore, the idea of ignoring undesired personal best positions, as an alternate to the move-away procedure in TS-PSO, seems to be reasonable. Hence, the following velocity function was proposed in (Modiri, et. al., 2011b). There, the authors also changed the manner by which the initial velocity was selected to a bi-state condition, as follows in (1.3). vt = ωvt−1 + c2 η2 (gbestt−1 − xt−1 ). j j v0 (i, j) = randomly selected to be equal to vminimum or vmaximum . iǫ1, 2, ...m jǫ1, 2, ...n (1.3) 12 The reason behind the second modification can be found in nature. Bees, as an example, start their initial flower-searching activities with a speed that is higher than that of the subsequent search steps. In the conventional RPSO, ω is decreased from 0.9 to 0.4 during optimization. This way, the algorithm emulates the mentioned reducing search speed concept. However, such an approach has shown no impact on BPSO (Jin, et. al., 2007), therefore, ω is generally set to a constant value of 1 in the conventional BPSO. In (Modiri, et. al., 2011b), however, this large initial velocity was introduced through (1.3), whereas in j j the case of BPSO, vminimum was set to 0 and vmaximum was assumed to be 6. The answer to the question ”why 6?” can be found in (Jin, 2007). More explanations are given in section 2.2. Many articles in the literature, such as (Eberhart, et. al., 2001; Carlisle, et. al., 2002; Hoorfar, 2007; Jin, et. al., 2007; Camci, 2008), have investigated the best values for c1 and c2 . According to the most recommended results, both c1 and c2 are either set to 1.49 or 2. Jin (2007), however, has shown better convergence performance when c1 and c2 are both set to 2. Therefore, c1 = c2 = 2 is used in this study as well. In the next step, a brief comparison of the conventional and modified algorithms when used in optimizing a set of well-known benchmark fitness functions is reviewed. Rastigrin, Griewank and Rosenbrock benchmark functions are studied in many articles in the literature, such as in (Hoorfar, 2007; Nakano, et. al., 2007). These three functions possess a global minimum that occurs at the origin of the solution space. The assessment results of the following four methods are reviewed here. It is noteworthy that the initial random solutions should be kept identical for all methods during each run in order to have reliable comparison results. Method.1: c1 = c2 = 2, and v0 as shown in (1.1) (conventional method). Method.2: c1 = c2 = 2, and v0 as shown in (1.3). Method.3: c1 = 0, c2 = 2, and v0 as shown in (1.1). Method.4: c1 = 0, c2 = 2, and v0 as shown in (1.3). 13 In (Modiri, et. al., 2011b), the convergence performances of the aforementioned methods were studied in 500 iterations averaged over 1000 runs, and the results were presented in terms of the average final value (AFV), and the average iteration number (AIN) required to attain a given AFV. The average convergence trend over iterations, and the distribution of the final values to which the algorithm has converged were also shown. A 10D scenario (n=10) was analyzed with m= 5, 10, and 20 particles. Each parameter was allowed to vary either over [-5,5], [-500,500], or [-50000,50000]. Optimization cost which was defined as the time required for the algorithm to converge to an acceptable solution (predefined by a threshold) was claimed to be the highest priority in the study. Due to the fact that decreasing the number of particles can shorten the optimization time considerably, no more than 2n particles were employed. A short summary of the conclusions is depicted in this chapter. For more details, readers are referred to chapter 2. It was shown that, Methods 2, 3 and 4 outperformed the conventional one (Method. 1) in terms of the convergence speed for Rastigrin function better than the two other functions. In terms of the best final value, Rosenbrock function ended up with more improved results. The very important point is that, the resulting error was only acceptable when the parameters were allowed to vary in a wide variation span. Figure 1.1 depicts the fitness function trend averaged over 1000 runs when n = m = 10. Evidently, the modified methods get far better than the conventional one as the variation range becomes wider. The termination criterion was set to the maximum iteration number of 500. Interestingly, for Griewank function, all methods had similar performances. Let us, again, highlight an important fact at this point. That is, the performances of the EAs are problem-dependent. In terms of the computation complexity, it is obvious that Methods 3 and 4 are lower in complexity due to elimination of a term from fitness function without any increase in the required number of iteration cycles; however, more features are needed to be compared. In Figure 1.2, the probability density functions of the final fitness values are plotted for Rastigrin. The results are achieved over 1000 runs. Each run is a 500-iteration procedure. In 14 the figure, MFFV denotes the maximum final fitness value. To better illustrate, the four bars at 0.1MFFV show the probability associated to the final fitness value being less than or equal to 0.1MFFV. Each bar represents one of the variants. Similarly, the bars at 0.2MFFV show the probability of the final fitness value being less than or equal to 0.2MFFV and more than 0.1MFFV. By studying the distribution of the final values to which the algorithms converge, Modiri (2011) concluded that the probability of ending with lower final fitness values increased in the modified methods as the variation span assigned to the parameters increased. This means that the superiority of the convergence rate for the PSO variants is more pronounced in the scenarios where the solution space is wider. The interested reader is referred to (Modiri, et. al., 2011b), for more details. Similar trend is claimed to be observed by studying the other two functions. Moreover, in order to have a set of results comparable to that presented by Nakano (2007), particle movements inside the solution space were traced in (Modiri, et. al., 2011b), and it was shown that the modifications offer the particles a greater degree of freedom in the initial stages of the exploration and faster convergence afterwards. Similar results were achieved when the absorbing wall boundary condition was used. The aforementioned versions of PSO (Method.1 to Method.4) are further addressed in pursuing case studies. It is worth reminding the reader that the main goal of this part of the chapter is to give a picture of the flexibility of the PSO algorithm and show how it can be redesigned according to the application. We are not trying to prove the viability of a single variant of PSO for all applications. Interested readers are also referred to other variations of PSO introduced in the literature, such as in (Ho, et. al., 2008; Gao, et. al., 2011; Luan, et. al., 2012; Montes de Oca, et. al., 2009). 15 Figure 1.1. Fitness values of the four introduced methods averaged over one thousand runs in a ten dimensional problem with ten particles for (a) Rastigrin, (b) Griewank and (c) Rosenbrock functions. The parameter range is considered to be [-50000,50000]. Figure 1.2. Probability density functions of the final fitness values for the four introduced variants considering one thousand runs in a ten dimensional problem for Rastigrin function with parameter ranges of (a) [-50000,50000], (b) [-500,500] , and (c) [-5,5]. 16 1.2 Particle swarm optimization-Reconfigurable radiators; Case study 1: PSO in antenna beam forming The previous section summarized the results achieved by some of the variations of PSO for mathematical benchmark problems. Here, modified BPSO is tested in a more realistic and demanding electromagnetic problem, namely, ’thinned array beam forming’, using the velocity function modifications described in the previous section and presented in (Modiri, et. al., 2011b). The goal was to be able to block the radiation of an array antenna in a given direction in an expeditious manner. Thinned arrays are periodic antennas with a number of non-illuminating elements. Such antennas are in demand for scenarios where software-defined pattern shaping is of interest; see (Jin, et. al., 2007). The applications of this type of beam forming were described in the Introduction of this chapter. Forty-element and 200-element linear arrays, as well as 33 , 44 and 55 planar arrays, were studied in the aforementioned paper and the results of this study are summarized in the ensuing paragraphs. Details can be found in chapter 2. BPSO, as it is obvious from its name, is quite attractive for problems with two-state parameters; see (Camci, 2008). However, in order to extend its utility to real-time smart applications, again, modifications are required. The algorithm is set to minimize the side lobe level (SLL) in the array broadside (as compared to the main-lobe strength). In other words, the following fitness function is to be minimized: f = max 20 log |arrayf actor|at the desired direction or directions (1.4) The array factor formula can be found in many references, such as Eq. (26) in (Jin, et. al., 2007). AFV, AIN, the average convergence trend, and the distribution of the final values averaged over more than 700 runs were studied in (Modiri, et. al., 2011b) for this problem. 17 Leaving the pure mathematical benchmark problems of previous section behind and tackling the realistic electromagnetic problems, it is shown that the complexity of the problem as well as the necessity of speeding up the design process increases. Generally speaking, an increase in the number of particles increases the computation time while reducing the probability of failure. Therefore, it is observed that a problem-dependent trade-off is normally taken into account. Interesting results are achieved when the previously mentioned four modifications are applied to BPSO. The performances of Method.2 and Method.4 in terms of AIN and AFV were shown to be superior to that of Method 1 (see chapter 2). However, Method 3 ended up with an inconsistent performance advantage over the conventional method. This modification (Method.3) was shown to even degrade the convergence performance. However, by reshaping Method.3 to arrive at Method.4, improved results were observed. The reader should note that, since the distinction between RPSO and BPSO is mainly originated from the underlying quantization strategy (Jin, et. al., 2007), by changing the quantization strategy, the results may also vary to some degree. Furthermore, it is important to mention that, since in BPSO each parameter is either 1 or 0, the parameter range, which was an underscored factor in the previous section for RPSO, is quite limited. Figure 1.3 demonstrates the convergence trend in the case of 44 planar array when m = 2. The dominance of Method.2 and Method.4 in terms of the convergence speed and final results, over the two other techniques are apparent. The radiation patterns are calculated in various φ-planes. It is also shown in chapter 2 that a similar set of results emerge for the case of the absorbing wall boundary condition. An interesting comparison of the optimization speed is also demonstrated in chapter 2. There, it is shown that, by adding two more particles in the 6-φ-plane scenario, the optimization time jumps from 2.8 hours, on a standard 2.83 GHz and 3.23GB personal computer, 18 Figure 1.3. Fitness value trends of the four introduced methods averaged over seven hundred runs in a sixteen dimensional problem considering two particles. Optimization study is performed in (a) six φ-planes, (b) eight φ-planes, and (c) ten φ-planes to 5.25 hours. Therefore, the significance of a method, which is capable of achieving reliable results expeditiously, becomes blaringly clear. 1.3 Particle swarm optimization-Reconfigurable radiators; Case study 2; PSO in frequency selectivity Yet as another demanding application, frequency selectivity of switchable microstrip antennas is studied in this section. Microstrip antennas (MAs) are inexpensive and relatively easy to manufacture. They are also conformal and flexible in shape. Therefore, the demand for MA applications has been constantly growing. However, the design of MA structure using the available analytical techniques becomes too complicated when one goes beyond the regularly shaped patches. This restriction is imposed by the high computational complexity of irregular configurations. In recent years, irregular structures are becoming more attractive due to their potential of 19 achieving large bandwidths or operating in a multi-band environment. However, existing articles in the literature point to a weakness of the existing approaches; i.e., the lengthy design and optimization processes, see (Balanis, 2008). It is interesting though that, EAs, such as PSO, have also been used for the MA design, see (Jin, et. al., 2008) as an example. As it was shown in the previous section, PSO is easy to implement and control, making it exceptionally appealing for multi-dimensional electromagnetic problems, such as those introduced in (Jin, et. al., 2005; Jin, et. al., 2007; Camci, 2008). The modifications applied to the velocity vector by Modiri, et. al. (2011) and explained in the previous sections are used in this section, as well. Here, the results of the cooperation between modified BPSO (implemented in MATLAB) and EM-SONNET, which is an MOM electromagnetic solver, are reviewed with the purpose of underscoring the performance improvement of a PSO-based optimization technique in problems where the real-time adjustment of the frequency response of the antenna is of paramount significance. In fact, this problem possesses a complexity higher than that of the previous case study. Here the electromagnetic solver calculates the fitness function. To have a more reasonable analysis, the comparison results of the convergence behavior for the conventional and the modified algorithms are shown in terms of the average optimization trend. For a further discussion of the performance metrics of this technique, such as the particle activity and success rate, the reader is referred to chapter 2. Switchable multi-segment antennas are one of the most versatile reconfigurable antennas due to the fact that their frequency responses can be simply varied by connecting some of the segments to the main antenna body and disconnecting the others. These types of antennas are widely used where frequency-selectivity is required. Yet, the reliability of the switching operation remains an issue. Figure 1.4 depicts the 15-segment MA, etched over a 3mm duroid substrate, which was analyzed in (Modiri, et. al. 2011). The antenna feeding connection was performed through 20 the middle patch, marked by a dot in the figure. The optimization algorithm does not make any changes to this middle segment. In this article, BPSO was employed to design irregular structures out of both 15-segment and 20-segment MAs (Figure 1.4). In this example, each segment was assumed to be a 66mm2 patch. The demonstrated results are obtained by averaging over 50 runs. A magnitude of -10 dB for the reflection coefficient, S11, at the design frequency (or frequencies) of interest is assumed as the success criterion. Also, in order to keep the optimization length reasonable for a real-time application, a maximum iteration number of T=50 is imposed as the termination criterion. It is noteworthy to mention that Jin (2005) has recommended a parallel strategy in order to expedite the MA design procedure using PSO. There, FDTD solver was used along with PSO. However, the parallel processing, although extremely effective in common applications, may not be valid for most software-defined reconfigurable structures were portable platforms are considered. The reason is that, parallel strategy requires large memory and processing capacity, which are difficult conditions to satisfy for the case of small size portable devices, which actually are the main focus of this chapter. Both single-frequency and multiple-frequency resonances are studied in (Modiri, et. al.,2011b) using Eq. (1.5) as the objective function. f= N X (S11 )atfi (1.5) i=1 where N depicts the number of resonant frequencies considered. It should be noted that, although Eq. (1.5) seems to be simple and trivial, S11 (reflection coefficient) cannot even be written in the form of closed formulae for irregular shaped patches. This renders this problem mathematically intractable. The frequency response and convergence trends of the conventional BPSO and the modified one are shown in Figure 1.5 when considering single frequency resonance at f1 = 5.1GHz. BPSO employs 7-particle swarm in this example. A close look at Figure 1.5 reveals the 21 Figure 1.4. The fifteen and twenty segment microstrip antennas used for simulation purpose are shown in (a). The best design achieved by the conventional and modified methods for the fifteen-segment scenario is shown in (b). (c) and (d) show the best designs of twenty-segment antenna achieved by the conventional and modified methods, respectively. convergence superiority of the modified BPSO over the conventional one. Regardless of the termination point, the modified BPSO offers a gain of at least 1.21 dB over the conventional one. More notably, it meets the design success condition of S11 =-10 dB faster than its counterpart with an average speed improvement of 50%. This result is of great interest to those who are engaged in the design of reconfigurable antennas with the computation time as a major constraint. The final irregular shaped structures associated with the best designs, achieved by the two methods, are also shown in Figure 1.5. The discrete probability density functions associated with the final fitness values, AFV, and AIN are also studied in (Modiri, et. al., 2011a). In this study, it is shown that the modified methods presented here achieve higher success rate. Furthermore, it was shown that the improvements stay valid for triple resonant design (4.8 GHz, 5 GHz and 5.2 GHz) as well. Obviously, a triple frequency scenario presents higher levels of computation complexity. For 22 instance, looking back at Eq. (1.5), it is clear that in a multi-resonance scenario more care should be taken in order to avoid the algorithm from getting locked in a cycle which reduces S11 at one of the frequencies while leaving the other frequencies with unacceptable S11 values. As a solution to this problem, any S11 value of less than -15 dB is fixed at -15 dB in (Modiri, et. al., 2011a). As a result, the swarm would stop reducing S11 at the frequencies for which -15 dB has already been achieved. It is shown there that the modified algorithm outperforms the conventional one in terms of the convergence speed, assuming a termination criterion of T=50 iterations. The two algorithms converge to identical configurations for this case. This configuration, along with the two other configurations, achieved in a singleresonance 20-segment MA, is shown in Figure 4b and 4c. For a more comprehensive study of the optimization time, a run-time table was presented in (Modiri, et. al., 2011a), in which the noticeable impact of the initial values on the optimization length was highlighted. Chapter 2 presents this information. 1.4 Particle swarm optimization-Reconfigurable radiators; Case study 3: PSO in bend effect compensation In this section, one more application of PSO, which is an interesting antenna design problem, is investigated. Wearable antennas (WA’s), as the name implies, are intended to be a part of everyday clothing. The earliest demand for WA’s was established by military with the purpose of reducing the chance of the identification of a radioman within a squad. The conventional military communications person often has to carry long antennas and such antennas are visible from a distance (Hall et al., 2006). Later on, WA’s received wide interest for sports, emergency, space, and medical applications too. Among these, medical applications are the most demanding so far. As a critical components of body area networks (BAN’s), WA’s need to maintain communications among in-body, on-body, and off-body devices. Therefore, WA’s have been developed with a variety of frequency and radiation 23 specifications. Nevertheless, it has been shown that several factors can deteriorate the performance of a WA, and ultimately, create health hazards in telemedicine applications. Thus, a serious need for real-time reconfigurable WA’s has emerged, especially for scenarios where changing posture due to movement imposes serious performance degradation. According to the application requirements, on-body textile antennas need to be flexible and conformal in structure, light in weight, and reasonably small in size. Therefore, MA’s seem to be one of the best candidates for such applications. Considering BAN’s, two performance criteria are generally expected to be met: 1. The patient comfort should be satisfied, especially in long term monitoring situations, 2. The antenna should preserve an acceptable performance for various performancedegrading scenarios. Wrinkles and bends on the antenna surface or moisture absorbed by substrate textile are some of such scenarios (Hall, et. al., 2006). Textile materials, in general, have low dielectric constants (below 2), and, as a result, a textile MA potentially has a relatively large bandwidth which can compensate for slight variations in the resonant frequency. However, the environmental variations may easily go beyond this limitation. Here, the effect of one of these variations, i.e., bending, is studied. Other variations can be treated in a similar way. Again, owing to the algorithm’s flexibility and high convergence speed, modified BPSO is used here in order to reconfigure a 2.45GHz MA in a real time manner. In (Modiri, et. al., 2011d), a patch antenna was proposed as the 2.45 GHz reconfigurable WA and was simulated in EM Sonnet. BPSO kernel was implemented in MATLAB. At this point, let us have an in depth discussion of practical details of a few implemented reconfigurable WAs. In (Hall et al., 2006), a 56mm 51mm textile patch antenna was fabricated on a 3mm fleece substrate with resonant frequency of 2.45 GHz. The ground plane size was 24 Figure 1.5. (a) Average fitness value trend and (b) the best optimization results achieved by the conventional and the modified BPSO algorithms using seven-particle swarm. considered to be 7671mm2 and the coaxial feeding point was located at the distance of 19 mm from the edge. The same textile antenna was re-fabricated by Modiri et al., and measured using Agilent E5071C VNA , see Figure 1.6. According to the measurements, it was claimed that by creating bends of 45 and 90 degrees on the antenna structure frequency jumps as large as 144 MHz and 250MHz, respectively, would occur in the frequency behavior of the antenna (see section 2.4). It is worth noting that a structure modification (i.e., bending) in only one dimension has resulted in such significant deviation in the frequency response of the MA. In order to compensate for this undesired frequency deviation, parasitic narrow patch elements were appended to the main antenna body. Once bending has occurred, some 25 Figure 1.6. Measurement arrangements for the wearable antenna in (a) flat, and (b) bent modes. Switchable antenna with compensating parasitic element in bent mode is shown in (c). of these parasitic elements would be connected to the main antenna body via switches in order to enlarge the actual length of the antenna, and thus, compensate for the shrinkage which has occurred in the electrical length of the element. Experimental verification is shown in chapter 2. Two scenarios were investigated in (Modiri, et. al., 2011d). In one scenario, the reconfigurable antenna consisted of three switches and a single parasitic element. While in the other one, eight switches were considered along with three parasitic elements. Of course, the higher the number of switches and parasitic elements, the better would be the optimization performance of the frequency response. Yet, it is impractical to integrate a large number of switches in a wearable device. The overall idea of this case study is quite similar to the previous one, although they differ in details and application. In the current scenario, a monitoring system is required to check the signal power at the desired resonant frequency by sampling the power period- 26 ically. As soon as the power falls below a predefined threshold (due to some environmental variations), BPSO is invoked in order to find the best switching arrangement to retrieve the original operation frequency. The objective function in these types of problems is typically the reflection coefficient at the desired frequency, which, of course, should be minimized, see Eq.(1.5). Yet, as another solution to this problem, it seems feasible to generate a bank of predefined switching arrangements to compensate for a set of environmental variations. However, in practice, it is quite possible that an unexpected variation, such as a twisting (multidimensional bend), can occur for which no solution is provided. Moreover, such a solution can only be valid if it is customized to a single person due to the exclusive personal body styles. Therefore, the first solution, i.e., real-time random decision making, was chosen by the author. To give the readers an idea of the final results achieved by the algorithm, Figure 1.7 shows the retrieved frequency response in the case of 30 and 45 degree single-dimensional bends when 8-switch, 3-element scenario is considered. Since, in practice, reflection coefficient magnitude of -10 dB is acceptable in most antenna applications, a success threshold of -10 dB was imposed in this study. It was also shown therein, that, overall, the modified BPSO outperformed the conventional one in maintaining the desired frequency performance. This is a result similar to that of the case studies 1 and 2. More details, such as convergence trends associated with the conventional and modified BPSOs, distribution curves related to the final fitness values, and frequency response curves of the two studied scenarios, can be found in the next chapter. 27 Figure 1.7. |S11 | measurement results of the flat and bent modes for the eight-switch, threeelement scenario. 1.5 Breast cancer detection using particle swarm optimization in MBMD In the last part of this chapter, PSO algorithm is tested in a more complex and challenging problem where the makeup of an unknown multi-layer structure has to be estimated. Gandhi (2010) and Yeung (2009) successfully used PSO to estimate the composition of a multi-layer structure using simulations. The object under the test was exposed to electromagnetic radiation and the scattering field was processed. In these studies, the medium under the test was assumed to either be lossless or low-loss in order to decrease the calculation complexity. However, the application of PSO in these types of problems is not limited to low-loss structures. This concept can, in fact, be utilized in more demanding and, at the same time, interesting fields, such as tumor detection. As shown by Golnabi (2011), a typical MBMD system consists of an array of antennas, which encircle the breast tissue. The antennas radiate the electromagnetic energy, one at a time, and the receivers gather the 28 scattered field. A set of measurements is performed at distinct frequencies and/or different transmitter positions. The inputs to the algorithm are these measurement results. Then, through a challenging inverse problem analysis, the algorithm has to get relatively close to an acceptable guess for the tissue structure. Here, we particularly investigate the breast cancer detection through MBMD. The makeup of the tissue layers can be estimated with respect to their dielectric constants. Therefore, the optimization goal can be simplified in words to ’estimating the dielectric constants of the tissue layers’. Similar to many other cancer cases, it is well proven that the early detection of the breast cancer drastically raises the survival rate of the patient. Moreover, the treatment and recovery becomes more tolerable, both physically and financially, when the tumor is detected soon enough. Although, X-ray mammography is available as the most recommended breast cancer detection technique, it has a dismal record of 20% detection failure, which cannot be easily ignored (Lazebnik, et. al., 2007). In addition, X-ray is ionizing, rendering this modality somewhat unsafe for scenarios where frequent checkups are required or highly recommended, see (Hassan, et. al., 2011; Nikolova, 2011). There have been thorough attempts toward developing more convenient, non-invasive breast cancer detection techniques that pose little risk to the patient when routine checkups are performed. Ultrasound and MRI seem to be two promising examples. However, these two techniques are either low in accuracy, in the case of ultrasound, or exceedingly expensive, in the case of MRI, when compared with X-ray mammography. In other words, they cannot compete with X-ray in cost/accuracy, especially when routine checkups come into the picture. That is why they are not used as widely as X-ray in clinical practice for breast cancer. Moreover, remarkable initial studies have been published, introducing more innovative methods of cancer detection at microwave, near-infra-red, optical, and tera-hertz frequency bands, see (Sabouni, 2010; Gamagami, 1997; Lazebnik, 2008; Khan, 2007; Jensen, 2011; Zhurbenko, 2010; Winters, 2009; Delbary, 2010; Zhurbenko, 2011; O’Halloran, 2009). Reviewing these 29 articles sheds light on the promise of MBMD as the diagnostic tool of choice for a noninvasive, safe, and accurate detection of breast cancer. The reason behind this fact is, simply, the practical aspects of electromagnetic imaging techniques. For instance, one important factor is the detection of tumors deep within the tissue. It is well known that, higher is the radiation frequency, smaller will be the penetration depth inside the human body. Therefore, higher frequency radiation (optics and terahertz) can hardly reach subcutaneous tumors. Also, the standard sizes of the electrical components at microwave frequencies are small enough to make a non-invasive, portable detector feasible. Yet as another fact, microwave radiation is quite responsive to the liquid content of the tissue. This would be of significant importance since the overall ratio of the blood flow in the breast tumor is between 4.7 and 5.5 times larger than that of the normal tissue (Dellile, et. al., 2002). As a result, employing non-ionizing microwave radiation in cancer detection and treatment monitoring has received a great deal of attention. A list of research studies in this area can be found in (Lazebnik, et. al., 2007). However, a hybrid technique, may ultimately be shown to be the most effective means of detecting breast cancers, see (Paulsen, et. al., 2005). Reviewing the existing articles on the subject of breast cancer detection using microwave radiation, one can identify the following key observations: 1. A lack of large-scale, in-vivo measurements for both validation and/or revision of the theoretically proposed techniques. 2. Low success rates of malignancy detection in non-superficial areas of the breast, even in the simulation stage, especially, when distinguishing between glandular tissue and tumor becomes critical. A closer examination of the above factors reveals the internal connectivity between them. That is, in order to conduct extensive in-vivo measurement studies, one has to obtain approval from regulatory bodies, which rely heavily on theoretical viability of the proposed 30 techniques and their supporting studies. In (Modiri, et. al., 2011c), it is argued that the inadequacy in modeling the tissue is the main culprit behind one arriving at low success rate in detecting cancerous tissues. In particular, non-negligible variations in the tissue characteristics from person to person, and even in a single person at different times of the month or under diverse experimental conditions, must be taken into account for the purpose of theoretical modeling. In one of the largest scale studies, reported by Lazebnik (2007), 807 samples were investigated over the frequency band of 0.5 GHz-20 GHz. There, it was practically demonstrated that the electrical properties of the tissue extend over a wide range. Cho (2006) also illustrated that these electrical properties were significantly influenced by the measurement setup. Therefore, the authors, in (Modiri, et. al., 2011c), proposed modeling permittivities as random variables. In addition, PSO, as a random-based EA algorithm, was chosen to find the breast tissue composition through solving the inverse scattering problem. It should be noted that dealing with inverse problems with a large number of unknowns has its inherent difficulty of arriving at nonunique solutions. Hence, the unknowns have to be managed carefully. In order to determine the solution technique, we need to model the measurement setup. It should be noted that, typically, array radiators are used for cancer detection, see (Golnabi et al., 2011), and the transmittance between two selected antennas is studied at a time. Therefore, the electromagnetic analysis can be simplified to the interaction of a narrow beam with a multi-layer structure, and this can be further simplified to planar wave scattering analysis in a lossy multi-layer structure (Orfanidis, 2008; Thakur, et. al., 2002). As it is shown in Eq.(1.6), the accumulated error, considering both amplitudes and phases of the reflection and transmittance coefficients, are utilized as the objective function (f) of the search algorithm. Nm is the total number of measurements. R and T are the reflection 31 and transmittance coefficients, respectively. f= X {||Ractual | − |Restimated || + (1.6) |phase(Ractual ) − phase(Restimated )| + ||Tactual | − |Testimated || + |phase(Tactual ) − phase(Testimated )|}; where ’arg’ refers to the phase operator. To illustrate how PSO can be used in such a problem, let us review the methodology and summarize the results demonstrated in (Modiri, et. al., 2011c) in the next section. More details are given in chapter 3. 1.5.1 Modeling and estimation methodology The first step in solving this problem is to generate a model for the object under the test. The breast tissue model was generated with respect to the breast physiology. The breast model structure is composed of glandular tissue, fat, and skin. For the case of cancerous tissue, tumor should be also added to the picture. As shown by Lazebnik (2007), the electrical properties of normal breast tissue are primarily dependent on its adipose content. Therefore, the model shown in Figure 1.8 takes the adipose contents into account. According to the breast physiology, various models of breast, in terms of the number of layers, can be suggested. This, in fact, defines the accuracy and computation simplicity levels. It is also required to assume a value as the total thickness of the tissue. Of course, this parameter is not a fixed value for all subjects; nevertheless, 4.4 cm was used in (Modiri, et. al., 2011c). Since the breast is typically in a encompassing radiator structure which would slightly squeeze the tissue, 4.4 cm was determined to be reasonable. The microwave radiation was analyzed in the frequency band of 1-2.25 GHz. The following two scenarios were studied: 32 1. Assuming there was no a priori knowledge available of the type and composition of the tissue under the test. 2. Assuming there was some reasonable a priori knowledge of the tissue under the test. Of course, the first scenario was the most difficult one in terms of the computation complexity. It should be noted that, the complexity of the problem is not due to the fitness function calculation. It is mainly due to the difficulty of solving the inverse problem in such a large solution space. Two PSO variants were proposed for the two desired scenarios. In both scenarios c1 and c2 are assumed to be equal to 2. ω is set to 1 for discrete PSO and gradually decreasing from 0.9 to 0.4 for the complex one. Here, it is noteworthy to emphasize again the problem dependency of PSO variants. The variant introduced for the previous case studies was not the best choice for this case study. In scenario 1, due to the lack of a priori knowledge, the random variables representing the dielectric constants were free to choose any values between the lowest possible and largest possible constants related to the human body tissues. In this case, Complex number PSO (CPSO) was used for which the variation range was relatively wide; CPSO is similar to RPSO, except for dealing with complex numbers instead of real ones. In fact, by implementing this scenario, it was shown that PSO could be utilized in estimation of the composition of any tissue. This can be achieved if the penetration depth and receiver sensitivity satisfy the detection requirements. In the second scenario, it was known that the tissue under the test is the breast tissue. There, at each iteration cycle, two optimization procedures were performed. First, for each layer, a tissue type was randomly assigned out of a set of six tissues, including blood, skin, fat-dominated tissue, glandular tissue, 31-84% fatty tissue, and malignant tissue (Lazebnik 2 et al.,2007). These are, in fact, the six tissue types which can exist in human breast. Each 33 of these six classes was represented by a random variable whose variation range was defined according to the measurements reported in the literature. In the second optimization step of scenario 2, after defining the arrangement of the tissues, the exact value of the dielectric constant related to each tissue layer was chosen out of its corresponding variation range, so that the optimum matching to the measured scattering fields could be achieved. For this scenario, in order to better manage the problem, discrete PSO (DPSO) was proposed. In principle, DPSO is similar to BPSO. The differences reside in the following two features: 1. The number of data levels is two for BPSO, while it is generally more than two for DPSO (six for the aforementioned DPSO). 2. The values which can be assumed at each level are restricted to 0 and 1 in BPSO, while it is extended to a continuous variation span in the above DPSO. For more detailed information, the interested reader is referred to chapter 3. The permittivity estimation results in the case of 7-layer modeling when DPSO is employed are shown in Figure 1.8. A 3 mm thick tumor was added to the model in two distinct positions. First, the tumor was added inside the fatty tissue, and then it was moved to the glandular tissue. Obviously, due to similarity of the permittivity values of the glandular tissue and the tumor, it is somewhat difficult for the swarm to make an accurate estimation for the second case. Nonetheless, DPSO can successfully identify the abnormal heterogeneity in the tissue characteristic in both cases. Figure 1.9 and Table 1.1 shows the comparison results for the two cases of 5-layer and 7-layer tissue models along with the convergence trend. Figure 1.10 shows the estimated permittivities as compared to actual ones at the studied frequencies. The variation of the permittivity as a function of frequency is clearly illustrated in this figure. In addition to many parameters existing in the solution domain, incident beam angle can also be considered for the narrow beam radiation case. Since the number of unknowns 34 is relatively high in the inverse problem equations, adding more equations by adding more incident angles seemed to be helpful; however, the author didn’t find this approach as a reliable remedy. In fact, in many situations adding more equations leads the swarm to a wrong direction, please see (Modiri, et. al., 2011c; Modiri, et. al., 2012). To summarize, it is depicted that PSO-based algorithm proposed herein successfully identifies the contrast between the permittivities of the tissue layers with an average estimation error below 10%. This accuracy is increased to an average error rate smaller than 5% in the second scenario. Obviously, by increasing the optimization cost (for instance, increasing the number of agents), the accuracy of the final responses is improved. Table 1.1. Optimization cost Layers (Scenario) Scenario 1 5-layers Scenario 1 7-layers Scenario 2 5-layers Scenario 2 7-layers Number of agents Number of frequencies Number of incident angles Convergence time (min) 40 2 2 7.81 280 6 3 185.3 40 1 2 0.7 280 1 3 9.2 35 Figure 1.8. (a) Simplified two dimensional model of the breast tissue. (b) Actual permittivity values and the estimated ones in the second scenario, assuming a 3 mm thick tumor in the fatty area (top) and in the glandular area (bottom). 36 5−layers (Second Scenario) 5−layers (First Scenario) 7−layers (Second Scenario) 7−layers (First Scenario) 3 10 2 Fitness Function 10 1 10 0 10 First Scenario −1 10 −2 10 0 Second Scenario 2000 4000 6000 Number of Iterations 8000 10000 Figure 1.9. Average fitness function trends for the two scenarios. 1.5.2 Implementation of microwave diagnosis tool Although medical diagnosis using microwave technology has been under investigation for almost a decade, there is no trace of a popular clinical device or an approved medical tool under this technology. The reason resides in the accuracy, reliability, user-friendliness and cost issues of the previous attempts. Other existing mass-screening diagnosis solutions suffer from major deficiencies. MBMD has proved to be able to detect breast malignancy and bone fractures, see (Hynes, 2012) and (Zhu, 2010) as two examples; however, for the case of breast cancer, a low-cost, reliable MI platform for critical routine checkups is not available to the public at this time. This serious issue became a motivation for the author and her supervisor to study and introduce a preliminary implementable microwave radiating structure for breast cancer detection. Encouraging simulation results, in terms of detection probability, were demonstrated for the proposed structure. The simulation study was followed by preliminary experimental results using an in-house manufactured breast phantom and cancerous tissue Imaginary(dielectric constant) F/m Real(dielectric constant) F/m 37 60 40 20 0 1 2 3 4 5 layer number 1 2 3 6 7 6 7 20 Actual @ 1GHz Actual @ 1.25GHz Actual @ 1.5GHz Actual @ 1.75GHz Actual @ 2GHz Actual @ 2.25GHz Estimated 15 10 5 0 4 5 layer number Figure 1.10. Comparison of the actual values at different frequencies with the estimated values in the first scenario for the 7-layer model. in the form of a donated breast cancer tumor. Taking advantage of the results shown in literature, an easily implementable structure which was, in fact, a crude model of a breast examination device was designed. The difference between the scenario in which a tumor existed inside the breast tissue and the scenario in which the breast tissue was healthy were demonstrated. 3D simulations were performed using Ansoft HFSS which is a high frequency electromagnetic (EM) solver based on the finite element method. To model the human body, Ansoft digital phantom was used. This phantom had a fixed structure and could not be used to model arbitrary body styles. For the experimental trials, a simple prototype was implemented and the scattering parameters (SP) were measured. Provided 38 that such parameters can be measured readily using network analyzers, they serve as an important tool for any noninvasive diagnostic modality. The measurement results of the implemented radiator structure are reviewed in the ensuing sections. More details on the proposed MBMD design and its associated measurements and simulations are presented in chapter 3. The following section summarizes the authors studies published in (Modiri, et. al., 2013a; 2013b; 2013c). Our preliminary proposed architecture The proposed device and its digitally simulated model are shown in Figures 1.11 and 3.15. The hemisphere-shaped base structure was created by 3D printing using inflexible ABS plastic and later covered by copper tape. The thickness of the plastic was 5 mm and it had sixteen specially designed holes to hold the SMA connectors considered for feeding of the antennas. The hemisphere structure was designed in SolidWorks and then printed using Dimension Elite 3D printer. The hemisphere base had an inner radius of 68 mm. Our preliminary microwave measurements were performed using ENA E5071C. Due to the relatively irregular shape of the breast model, the size and shape of the radiator structure were chosen by trial and error. The model in Ansoft for the female breast phantom, as it is shown in Figure 3.15, was used for simulation and it consisted of glands and fat. Skin, although very important, was not considered in the model. The glands were spread inside the breast in a web-shaped random structure. In fact, this phantom represented a single scenario in which the breast tissue was densely filled with glandular tissue. There were no phantoms for other body styles available to us. For tumor model, the dielectric properties were imported into HFSS. The Debye model corresponding to the dielectric constants of the normal and cancer tissues are available in the literature. Mechanism and analysis We defined the following four metrics: 39 Figure 1.11. The preliminary radiator structure which includes a novel design of sixteen antennae and a conductor cover. X Bent Dipole Antennas Y Z Figure 1.12. Digital breast phantom along with the proposed radiator structure. 11 |Snd | 11 arg(Snd ) = 16 X ||S jj j=1 = 16 X | arg(S jj 21 arg(Snd )= jj − arg(Stumor )| jj arg(Snormal ) normal ) j=1 21 |Snd |= jj normal | − |Stumor || jj |Snormal | 15 X 16 ij ij X ||Snormal | − |Stumor || ij |Snormal | i=1 j=i+1 15 X 16 ij ij X | arg(Snormal ) − arg(Stumor )| ij arg(Snormal ) i=1 j=i+1 (1.7) (1.8) (1.9) (1.10) 40 jj jj are the S1 1 (rewhere the operator ’arg’ is the phase operator. S1,normal and S1,tumor flection) parameters for the j th terminal for the normal and cancerous cases, respectively. Similarly, S2 ’s are the S21 (transmittance) parameters for the ith and j th terminals. Equations (1.7)-(1.10) depict how the measured data was analyzed using the scattering parameters (SP). The main goal of the experiment was to isolate signatures due to the tumor. The signatures were detected as the result of comparing normal and cancerous tissues. Equations (1.7)-(1.10), in fact, show the cumulative normalized difference between normal and cancer tissues in terms of magnitude and phase for the reflection and transmittance coefficients at a single frequency (continuous wave exposure). The magnitudes of the S parameters for normal and cancerous breasts used in equations (1.7)-(1.10) were expressed in dB and the phases were expressed in radian. 1.5.3 Preliminary results After a number of trial attempts, the author was able to create reliable breast phantoms following the recipes in the literature. There were different recipes available; however, one was found to be more concise and precise (Hahn, 2012; Ostadrahimi, 2009). Creating tissue phantoms which can model the body accurately was and is an important challenge in this study. All tissue solutions are gelatin-type materials and degrade by time. Building the complete phantom requires careful temperature control. Fortunately, we received a tumor tissue from a breast cancer patient to use in our phantom. This was, in fact, a major step toward a realistic study of the effectiveness of the proposed structure. We performed all our measurements, summarized herein, within 48 hours of the completion of the creation of the phantoms. When not used, the phantoms were kept in a refrigerator, as it was recommended. Here, the results from two breast phantoms are reviewed; a normal breast phantom and a cancer breast phantom. The cancer breast phantom contained the donated tumor tissue. The two phantoms were created in exactly same way, at the same time and in exact same 41 containers. An attempt was done to make the two phantoms identical in any way other than the existence of the tumor. However, operator mistakes cannot be prevented easily and they add some levels of uncertainty to the solution. The tumor tissue was placed inside the glandular tissue. The phantoms had a base diameter of 12 cm and height of 4.5 cm. The tumor tissue donated to us was a relatively thin piece, the thickness of which was not more than 3 mm; however, it had an irregular spread shape with the largest dimension of 1 cm. The phantom and the radiator were placed on a bag of ice water during the measurements for two reasons; first, for mimicking the chest wall muscle backing the breast tissue in real scenarios, and second, for keeping the phantom cool throughout the measurements. This was not indeed the most accurate way to mimic the chest wall muscle, yet it was our best available option. Since the radiator structure was entirely covering the phantom, we were less concerned about the phantom dehydration during the measurement. Flexible RF cables were used to connect the ENA arms to the SMA connectors. Having sixteen radiating antennas and using a two port ENA, the two set of measurements using the two previously described phantoms took almost an work entire day performed by two people. Due to the nature of our measurement technique, we were able to measure the reflection coefficient for each antenna fifteen times and the transmittance coefficient between each two antennas two times for each phantom. This was the only way to have less concern about the reliability of our measurements as we had only one chance to go through the measurement steps due to phantom degradation. The antennas were originally designed to resonate at 1.2 GHz; however, as expected, the experimental setup, the environmental factors, and the presence of the phantom led to a slightly different resonating frequency. The impact of environment is another important issue when studying for such diagnostic modalities. The other important point was that, the sixteen antennas were cut by hand out of copper tape and then attached to the ABS structure’s inner body (see Figure 1.11), and it was well 42 600 500 Star: Frequency step of 42.5MHz Circle: Frequency Step of 85MHz Cross: Frequency Step of 127.5MHz Square: Single frequency of 1.354GHz |S11nd| 400 300 200 100 0 0 0.5 1 1.5 2 2.5 Frequency (up to which the results are added) (Hz) 3 9 x 10 11 Figure 1.13. Cumulative magnitude difference of reflection coefficient (|Snd |) for singlefrequency and multi-frequency studies. expected that there would be slight difference between their performances. Yet, since all the measurements were performed with the same radiating structure, the results were dependable for a preliminary study. It was observed that the optimum frequency was 1.354 GHz instead of 1.2 GHz. The effect of this deviation on the ultimate results was negligible since the new resonance frequency was still inside the frequency range which was recommended for MI measurements in the literature. Figures 1.13 and 1.14 demonstrate the cumulative normalized difference between normal and cancer tissues in terms of magnitude for the SP. Using ENA settings, we were able to gather the data with steps of 42.5 MHz; however, in order to show the significant advantage of multi-frequency measurement over a single-frequency one, the results for multi-frequency studies using frequency steps of 42.5, 85, and 127.5 MHz are shown in Figures 1.13 and 1.14. 43 600 Star: Frequency step of 42.5MHz Circle: Frequency Step of 85MHz Cross: Frequency Step of 127.5MHz 500 Square: Single frequency of 1.354GHz |S21nd| 400 300 200 100 0 0 0.5 1 1.5 2 2.5 Frequency (up to which the results are added) (Hz) 3 9 x 10 21 Figure 1.14. Cumulative magnitude difference of transmittance coefficient (|Snd |) for singlefrequency and multi-frequency studies. The results for a single-frequency measurement at 1.354 GHz are shown, as well. In multi-frequency analysis, the cumulative differences achieved at all considered measurement frequencies were simply added up. It is evident that the tumor signatures are noticeable in all the figures. Phase responses are shown in chapter 3. In Figures 1.13 and 1.14, as expected, the slope of the magnitude signature decreases as the measurement frequency gets far from the optimum resonance frequency of the antennas. At 1.354 GHz, the values achieved for equations (1.7)-(1.10) are 33.7610, 19.8374, 3.6856 and 69.0095, respectively. The results were quite encouraging since these signatures were easily detectable using a common ENA. However, these were the results obtained from a single set of measurement trial and needed to be verified in several cases for the sake of reliability and further improvement. Chapter 3 reviews the details of this study. CHAPTER 2 PROPOSED PSO AND ITS APPLICATIONS IN RECONFIGURABLE ANTENNA DESIGN 2.1 Introduction As it was mentioned in the previous chapter, a set of electromagnetic problems with known solutions were chosen to test and validate PSO algorithm prior to introducing the complex and unknown objective problem of MBMD. Chapter one presented a summary of these problems and how PSO was customized to tackle them. This chapter, however, brings the details of those studies to the forefront. The papers which were published by the author and her supervisor on this subject are the building blocks of this chapter. 2.2 PSO in thinned array beam forming1 Here, PSO variants are used to block the radiation pattern a thinned antenna array in a certain direction. To recall, some modifications in the velocity calculation of the Particle Swarm Optimization (PSO) algorithm were introduced in section 1.1. The suggested modifications aimed to arrive at a more straightforward and robust search procedure as compared to the conventional method. Here those modifications are assessed in more details. Two main factors, i.e., personal best influence and initial velocity values, are evaluated. It is shown in this section that in problems with wide-range parameters, the effect of personal best locations is intrinsically encompassed by that of global best locations, thereby allowing for 1 c 2011 IEEE. Adapted, with permission, from Modiri,A. and Kiasaleh, K., Modification of Real-Number and Binary PSO Algorithms for Accelerated Convergence, IEEE Transactions on Antennas and Propagation, Jan./2011 for section 2.2 . 44 45 further simplification of the PSO algorithm by eliminating the factor which accounts for the personal best solutions in the velocity calculation. This simplification expedites the convergence procedure in real PSO. It is also shown that the initial velocity values can be modified in order to enhance performance in terms of achieving better solution when compared with the existing algorithms, particularly in binary PSO. In order to validate the viability of the proposed procedure, the performances of the real-number and binary PSO algorithms with different velocity calculations are assessed in 1000-run sets, and pros and cons are studied. In particular, the performance of the proposed velocity function shown in Eq.(1.3), when used to design software defined thinned array antennas, is shown to be superior to that of the conventional velocity function shown in equations (1.1) and (1.2). The viability of Eq.(1.3) can even be proved mathematically by evaluating the average position of particles in each iteration cycle. Let x(i,j)t and xv(i,j)t denote the position and velocity of the ith particle in the j th dimension during the tth iteration, respectively. In PSO algorithm, this position is achieved by x(i,j)t = x(i,j)t−1 + v(i,j)t . (2.1) Using the conventional velocity function Eq.(1.1), the average particle velocities are given by vt = ωvt−1 + c1 ( pbestt−1 2 − gbestt−1 xt−1 xt−1 ) + c2 ( − ). 2 2 2 (2.2) Where pbestt and gbestt denote the average personal best position vector of all existing particles and the global best position vector in the tt h iteration, respectively. xt and vt are the position and velocity vectors averaged over i. Substituting c1 = c2 = 2 in Eq.(2.1) and using Eq.(2.2), the average position vector is simplified to xt = ωvt−1 + pbestt−1 + gbestt−1 − xt−1 . (2.3) 46 If the algorithm moves the particles in the correct direction, xt−1 should approach pbestt−1 . Therefore, the average position of the particles in a successful iteration cycle in which the majority of particles have moved to positions with better fitness values (compared to previous iteration) can be approximated as Eq.(2.4). xt = ωvt−1 + gbestt−1 . (2.4) One can observe that the average position of particles in Eq.(2.4) may also be achieved by using the proposed velocity function, given by Eq.(1.3) in above procedure. 2.2.1 Evaluation of the proposed velocity function in RPSO In order to have comparable results, the benchmark fitness functions used in (Hoorfar, 2007; Nakano, 2007), shown in equations (2.5), (2.6) and (2.7), are used. To elaborate, these three functions all possess a global known minimum that occurs at the origin of the n-dimensional solution space. So the goal of optimization is to approach this zero value by moving m particles in an n-dimensional solution space. Rastigrin function: f (x) = n X i=1 (x2i − 10 cos(2π xi + 10)). (2.5) Griewank Function: n n 1 X 2 Y xi f (x) = xi − cos( √ ) + 1. 4000 i=1 i i=1 (2.6) Rosenbrock function: f (x) = n X i=1 (100(xi+1 − x2i )2 + (xi − 1)2 ). (2.7) Where x = [x1 , x2 , , xn ]. To assess the impact of proposed modifications separately, the results of the following four 47 methods are investigated (the initial solutions are identical for all methods during each run throughout this paper): • Method1: c1 = c2 = 2, and the initial velocity is given by Eq.(1.1). • Method2: c1 = c2 = 2, and the initial velocity is given by Eq.(1.3). • Method3: c1 = 0, c2 = 2, and the initial velocity is given by Eq.(1.1). • Method4: c1 = 0, c2 = 2, and the initial velocity is given by Eq.(1.3). In order to have comprehensive statistical study, the convergence behaviors in different cases are observed in 500 iterations during 1000 runs, and the average results associated with each benchmark function are depicted in terms of the achieved average final value (AFV) and the average iteration number (AIN) required to attain a given AFV. The results are averaged over a set of runs. The number of sample runs in the set increases by 20, i.e. the results of 20, 40 1000 sample runs are averaged and plotted. In this manner, in addition to achieving consistent statistical results, we are also able to measure the number of samples required for attaining reliable results. A 10D scenario (n=10) is assumed with particle numbers of 5, 10, and 20. The parameters can vary in the ranges of [-5,5], [-500,500], and [-50000,50000]. Although a higher number of particles are generally considered in RPSO, two objectives are obtained here by these selections. First, the (m,n) values are chosen so that they match the parameters for the ensuing binary PSO study, allowing for a comparative study. Second, since speed of convergence is a priority in this study, key solutions for accelerating the convergence while preserving PSO’s intrinsic robustness are sought. In real cases, it is almost impossible to reduce the number of dimensions in order to speed up the solution. However, decreasing the number of particles can drastically shorten the optimization time. Thus, we do not step beyond m=2n particles in the first step of our study. AIN/AFV results are shown in Figures 2.1-2.4. 48 Figure 2.1. AIN behaviour of the four methods over gradually increasing sample runs in a 10D problem with 5, 10 and 20 particles for (a) Rastigrin, (b) Griewank and (c) Rosenbrock functions and parameter range of [-500,500]. From Figures 2.1-2.4, two interesting trends are discovered. The first one is the high similarity between the results of the three modified methods, particularly in terms of convergence speed. The second common trend is that the modified methods mostly expedite the convergence by almost 18% at the expense of higher errors for lower parameter ranges. However, as the parameter range increases, the modified methods begin to slow down, while getting much lower errors with respect to Method1. For Rastigrin and Griewank, this transition happens somewhere between 500 and 50000, whereas for Rosenbrock it occurs somewhere between 5 and 500; i.e, in addition to exhibiting encouraging convergence behavior, the modified methods outperform the conventional one, both in speed and error rate, provided that the proper parameter ranges are selected. For Griewank, 50000 is such a range. 49 In order to cope with the realistic computation limitations of a standard 2.83 GHz and 3.23GB RAM PC running MATLAB, the procedure is performed for each function separately and the resulting figures are grouped together as in Figures 2.1-2.4. For the sake of clarity, the figure in the right column of Figure 2.4(c) as well as some other figures which will appear later in this section are shown in logarithmic scale. Figure 2.2. AFV behaviour of the four methods over gradually increasing sample runs in a 10D problem with 5, 10 and 20 particles for (a) Rastigrin, (b) Griewank and (c) Rosenbrock functions and parameter range of [-500,500]. To summarize the results, the modified methods outperform the conventional one in terms of convergence speed for Rastigrin function better than the two others, and in terms 50 of the best final value for Rosenbrock function. However, the cost of final value error is only acceptable for high parameter ranges. Figure 2.3. AIN behavior of the four methods over gradually increasing sample runs in a 10D problem with 10 particles for (a) Rastigrin, (b) Griewank and (c) Rosenbrock functions and parameter range of [-5,5], and [-50000,50000]. Figures 2.1-2.4 reveal the AFV and the required AIN, but they do not explain the gradual convergence behavior, which is of great importance in this study. Figure 2.5 demonstrates the fitness value variation during optimization process, averaged over 1000 runs for 10D problem with 10 particles. It is obvious that modified methods are superior for the typical range 50000 with termination criterion of 500 iterations. For the case of Griewank, all methods 51 Figure 2.4. AFV behavior of the four methods over gradually increasing sample runs in a 10D problem with 10 particles for (a) Rastigrin, (b) Griewank and (c) Rosenbrock functions and parameter range of [-5,5], and [-50000,50000]. perform almost similarly. Since average fitness trends may not be adequate for presenting the success rate, in Figure 1.2 the probability density function of the final fitness value in 1000 runs for 500 iteration scenario was plotted for Rastigrin function. It was shown that the probability of ending with smaller fitness values increased in modified methods as the parameter range increased. Similar trend was observed by studying other functions. For the sake of conciseness, we have omitted those results. 52 Figure 2.5. Fitness value of the four methods averaged over 1000 runs in a 10D problem with 10 particles for (a) Rastigrin, (b) Griewank and (c) Rosenbrock functions and parameter range of [-50000,50000]. In addition to convergence performance, the behavior of particles has also been studied by Nakano (2007) yet as another performance criterion. Figure 2.6 depicts how the average position of a sample particle in a sample dimension varies. The results are averaged over 1000 runs for Rastigrin function assuming n = 10, m = 20, and parameter range of 500. In order to have a better view, two areas in Figure 2.6(a) are zoomed out in figures (b) and (c). It is noteworthy that the modifications offer the particles a slightly greater degree of freedom in the initial phase of the exploration, but faster convergence in later iterations. 53 Similar trend is observed by studying other sample particles, other dimensions, and other functions. Again for the sake of brevity, those results are not included. Figure 2.6. (a) Average sample particle position in an arbitrary dimension for the four methods over 1000 runs with 10 particles in a 10D problem for Rastigrin function with parameter range of [-500,500], (b) zoom of 1-100 iterations, and (c) zoom of 200-300 iterations. A similar procedure is repeated using absorbing wall boundary condition for a 10D problem with 5, 10 and 20 particles. Figure 2.7 shows the convergence behavior for Rosenbrock function in the same 10D problem with 5 particles. It is apparent that the results are similar to those of the invisible wall. Figures related to the other cases and functions are eliminated since they also show analogous outcomes. Since it was concluded from previous study that 54 Figure 2.7. Fitness value of the four methods averaged over 500 runs in a 10D problem with 5 particles for Rosenbrock function with parameter range of [-50000,50000] using absorbing boundary condition. 500 runs are enough for collecting reliable average results, the sample number is reduced to 500 runs. Finally, to provide a comparable analysis with studies available in the literature, the parameters of the search are chosen similar to those presented in (Hoorfar, 2007; Nakano, 2007). It is assumed that PSO has a population size of m=30 particles, and the fitness functions have three parameters (n=3) ranging in [−50000, 50000]. The termination criterion is defined by a maximum iteration number of 500. Figure 2.8 depicts the convergence behavior averaged over 500 runs. It is observed that the modified methods yield much better performance in terms of the convergence speed and final fitness value for Rastigrin and Rosenbrock, but for Griewank, the results of all methods are similar. In comparison, TSPSO algorithm requires more than 5000 iterations to achieve similar results for parameters with shorter range ([-5,5]) (Hoorfar, 2007). It was shown in this section that the two modifications proposed for PSO yield similar improvements in the RPSO algorithm behavior. The modifications generally speed up the 55 Figure 2.8. Fitness value of the four methods averaged over 500 runs with 30 particles for (a) Rastigrin, (b) Griewank and (c) Rosenbrock functions with parameter range of [50000,50000]. algorithm. As the parameter range increases, the superiority of the proposed modifications improves greatly both in convergence speed and final value, particularly for Rastigrin and Rosenbrock functions. It is also shown that run number of less than hundreds is not sufficient for making statistical decision. 2.2.2 Evaluation of the proposed velocity function in thinned array design In this section, we assess the performance of the proposed algorithm in BPSO scenario, which has been focus of some recent studies (Camci, 2008). BPSO is highly useful in problems with two-state parameters, such as irregular shaped planar antenna design and array antenna design (Jin, 2007), but to extend its appliance to smart type of applications requiring fast resulting, it is not appropriate yet. This paper studies thinned antenna arrays which are periodic arrays with a number of non-illuminating elements. In addition to general array 56 applications, these types of antennas can be used where software-defined beam-forming or pattern-shaping is required by using the currents of elements as tuning factors. In this section, we first study 2n-element linear half-wavelength thinned antenna array design, using BPSO, assuming c1 = c2 = 2. It has been proven for BPSO that the gradual decrease of ω has no effect on performance, thus, as recommended in literature, we assume ω to be constantly equal to ”1” (Jin, 2007). The optimization goal is to minimize the side lobe level (SLL) in the antenna’s broadside. The lobes are assessed in ±90 degrees angular distance from main lobe center. The fitness function, shown in Eq.(2.8), is minimized. f = max 20 log |array f actor| (2.8) The array factor formula can be found in many references such as equation (26) in (Jin, 2007). Assuming the antenna in xy plane, in this paper, the SLL is assessed in the plane including the array itself and broadside (θ = 0) for linear array, but for planar arrays a number of φ-planes are considered. In order to have consistent study, first, the convergence behaviors in cases similar to those encountered in RPSO are observed in 1000 runs, and the average results associated with each case are gathered in two figures depicting AIN and AFV values. Then, planar array antenna cases are studied. As the complexity of fitness function evaluation increases, the necessity of achieving better results in a fewer number of iterations becomes more obvious. In addition, as the number of particles reduces, the calculation time reduces significantly, although the chance of convergence to appropriate results also drops. Thus, a problem-dependent trade-off should be considered. Since this study considers large number of runs, the dimensionality of the problems are selected according to address memory and time limitations. 57 Figure 2.9. Symmetric half-wavelength thinned linear array. 2n-element symmetric linear array The linear arrays are assumed to be symmetric arrays with 2n elements, as shown in Figure 2.9. A 10D scenario (n=10) with particle numbers of 5, 10 and 20 are considered and AIN and AFV curves are shown in Figures 2.10-2.11. The performance of Method2 and Method4 in terms of AIN and AFV for BPSO (similar to the RPSO case) is superior to Method1. However, although Methods2 and Method3 for RPSO yielded superior performance as compared to the conventional method, only Method2 yielded superior performance consistently for BPSO case, while Method3 offered inconsistent performance advantage over the conventional method in our studies. The major reason for this distinction between RPSO and BPSO is due to insertion of the two-level (binary) quantization (Jin, 2007). In fact, c1 = 0 (Method3) may even degrade the convergence performance unless it is used together with proposed initial velocity (as in Method4). The figures clearly show that the convergence speed (AIN) is significantly improved by Method2 and Method4, whereas the AFV gain is just slightly enhanced. Also, it is important to mention that since in BPSO each parameter can either be 1 or 0, parameter range will not be a considerable factor. 58 Figure 2.10. AIN behaviour of the four methods over gradually increasing sample runs in a 10D problem with 5, 10 and 20 particles. Figure 2.11. AFV behaviour of the four methods over gradually increasing sample runs in a 10D problem with 5, 10 and 20 particles. Non-symmetric planar array Similar study is repeated for the more complicated case of planar half-wavelength, nonsymmetric thinned array antennas (Figure 2.12). A 44 (n=16) problem is first considered. In order to shorten the total run time, 100 iterations are performed in 700 runs, and the 59 Figure 2.12. Non-symmetric thinned planar antenna array. swarm size is limited to two particles. The number of runs is reduced to 700 from 1000 according to the study results in previous section, showing reliable run number. The pattern is assessed in 6, 8 and 10 φ-planes (φ=30:30:180, φ=0:20:180, and φ=0:15:180, respectively). The superiority of Method2 and Method4 both in terms of convergence speed and final result is evident in Figure 1.3 which is not repeated here. Similarly, 33 , 44 and 55 planar arrays are studied with 3, 4 and 5 particles, respectively. This time, the pattern is assessed in no more than 6 φ-planes (φ=30:30:180), 200 iterations are performed in each of 700 runs. In order to alter the fitness function, a success boundary of -20dB is added to the fitness function introduced in 2.9, resulting in the following fitness function: f = max 20 log |array f actor| + 20 (2.9) Figures 2.13-2.15 show the AIN/AFV, density, and average fitness trend for the 33 array, respectively. The results confirm the previous conclusions. Method2 and Method4 are superior to the conventional method and Method3, both in terms of convergence speed and final value. Figure 2.14 shows that the probability of achieving a smaller final fitness values is higher for Method2 and Method4 as compared to other methods. An important point underscored by Figure 2.15 is that wherever the optimization procedure is terminated, 60 Figure 2.13. AIN and AFV behaviour of the four methods averaged over gradually increasing sample runs in a 9D problem with 3 particles. Method2 and Method4 outperform the other two techniques. Similar trend is observed by studying other cases, but, for the sake of conciseness, the results are not included here. Finally, as in RPSO, absorbing-wall scenario is investigated, and similar outcomes are achieved. In Figure 2.16, the optimization results associated with the 33 array with three particles averaged over 500 runs, using absorbing boundary condition, are depicted. It is apparent that the conclusions of invisible boundary condition can be extended to absorbing wall case. The overall results confirm that the proposed methods (Method2 and Method4) yield superior performance for BPSO. Finally, to underscore the time-consuming nature of the optimization process, 2.8, 3.4, and 5.5 hours were required for the 700 runs of simple 44 array with only 2 particles for pattern assessment in 6, 8 and 10 φ-planes, respectively, and angular resolution of ∆θ=0.167 degrees. To underscore the computing requirement, it is notable that adding two more particles to 6 φ-plane case converts 2.8 hours of computing on a standard 2.83 GHz and 61 Figure 2.14. Probability density function of the final fitness value for four methods, with 700 runs in a 9D problem with 3 particles in 6 φ-planes. Figure 2.15. Fitness value trend of the four methods averaged over 700 runs in a 9D problem with 3 particles in 6 φ-planes. 3.23GB RAM PC to 5.25 hours. Assuming a realistic software-defined thinned array scenario, a substantially smaller number of runs will be required. Thus, significance of a method or 62 Figure 2.16. Fitness value trend of the four methods averaged over 500 runs in a 9D problem with 3 particles in 6 φ-planes using absorbing wall boundary condition. methods with achieved satisfactory results in a smaller number of iteration becomes quite evident. 2.3 PSO in antenna frequency selectivity2 In this section, several irregularly shaped microstrip antenna structures are designed using the new version of Binary Particle Swarm Optimization (BPSO) algorithm explained previously, particularly aimed for software-defined radio (SDR) applications. The optimization results are compared using both the modified and the conventional BPSO algorithms. Pros and cons are studied in terms of optimization length, convergence speed and final design conformability to desired objectives. It is depicted that the modified BPSO achieves the design criterion considerably faster than the conventional one, at the cost of slightly limiting 2 c 2011 IEEE. Adapted, with permission, from Modiri,A. and Kiasaleh, K., Efficient Design of Microstrip Antennas for SDR Applications Using Modified PSO Algorithm, IEEE Transactions on Magnetics, May/2011 for section 2.3 . 63 particle exploration ability. This section, in fact, presents the details of the study summarized in section 1.3. The demand for microstrip antenna (MA) application has been continuously increasing owing to its well known unique properties. Although there are analytic models available for the design of MA structures, these models are merely proper for simple regularly shaped patches. This limitation is imposed due to the high computational complexity of irregular structures. In recent years, however, irregular structures are gaining popularity due to their ability to achieve large bandwidth or multi-band operation. Nonetheless, available studies have shown lengthy optimization procedures for such designs. The PSO’s simplicity, ease of implementation, and flexibility make it extremely appealing for multi-dimensional electromagnetic designs (Jin, 2007; Jin, 2005; Camci, 2008). Modifications in PSO velocity vector, proposed by the authors in (Modiri, et. el., 2011b) and explained previously are applied to MA design herein. In (Modiri, et. el., 2011b), the problem-dependent characteristic of PSO was underscored which is a critical feature required to be considered while customizing the algorithm to fit different problems. As it was mentioned in section 2.3, BPSO (implemented in MATLAB) is used along with EM-SONNET, which is an electromagnetic simulator based on the method of moments (MoM), to design irregularly-shaped MA structures. The main goal is to depict the performance enhancement of a PSO-based optimization method in optimizing antenna design for software-defined radio (SDR) application where prompt frequency response characteristic of the antenna is of paramount importance. The convergence behavior of the conventional and the proposed methods are compared in terms of the average optimization behavior, particle activity and success rate. The uniqueness of the study demonstrated in this paper resides in analyzing the proposed algorithm in an actual real time electromagnetic design, instead of the benchmark math problems of 2.5-2.7. Utilizing both MATLAB (as the optimizer tool) and SONNET (as the electromagnetic analyzer tool), the optimization time lengths are also investigated. 64 2.3.1 Switchable antenna design using modified BPSO In this section, the performance of the proposed algorithm is evaluated in several MA design scenarios. BPSO, in general, is greatly advantageous in optimization problems with two-state parameters; however, it is not appropriate for real-time applications in its conventional format. The design of MAs consisting of several switchable segments are studied here. Connecting or disconnecting any number of the segments to the main antenna body results in distinct frequency behaviour. These types of antennas can be used where frequencyselectivity is desired. This study focuses on reconfiguring the antenna structure by defining different frequency behaviour objectives (in SDR scenarios). The issues related to the switching system and switch types are beyond the scope of the paper and are not pursued here. Figure 2.17 shows the structure of a 15-segment MA on 3mm duroid substrate. The middle patch contains the feed and will not be affected by the algorithm. BPSO is used to design several irregularly shaped structures using 14 segments. The feeding position on the middle patch is deviated from the patch center to better accord irregular structures. In order to remove a segment, BPSO assigns 0’s for both the segment length and width, resulting in the disconnection of the patch from the main antenna body in SONNET simulation. In addition to 15-segment structures, 20-segment ones have also been designed. 15-Segment single-frequency design The final results of a single frequency optimization problem for a 15-segment antenna was presented in section 1.3. The S11 curves related to the best designs achieved by the two methods with termination criterion of T=50 iterations were shown in Figure 1.5. The irregular shaped structures designed by each method were also added to the figure. Figure 2.18 demonstrates the discrete probability density function of the final fitness values, as a measure of success. In this figure, MFFV stands for the maximum final fitness value achieved in the 50 runs. The optimization goal is to minimize the fitness function. Thus, if 65 Figure 2.17. 15-segment reconfigurable microstrip antenna structure. the probability of ending with lower final fitness values is higher for one method, the success rate will also be higher for that method. It is apparent that the proposed method shows higher success rate. In order to further complete the comparison study, the achieved average final value (AFV) is shown in Figure 2.19. The results are averaged over an increasing number of runs. The number of runs is increased by 2; i.e., the averaged results are obtained over 2, 4 50 runs. In this manner, in addition to performing more consistent study, we are also able to recognize the number of sample runs required for attaining reliable results. Converging to a smooth mode in Figure 2.19 confirms that 50 runs are acceptable for observing the convergence behavior of AFV. Of course, the more runs, the more reliable statistical results will be. However, since runs can be quite time-consuming (as will be seen later), 50 runs are selected herein. In order to observe the particle trajectory, the optimization procedure with a single particle was studied in (Janson, 2007). For the same reason, Figure 2.20 depicts the average particle trajectory in single-particle scenario for the fitness function given by Eq.(1.5) in one randomly selected dimension (out of 14 available ones). Similar overall exploration activities are observed for the two algorithms in this figure. It is depicted that in the proposed algorithm, the particle starts with slightly higher activity due to the initial velocity definition of 66 Figure 2.18. Probability density function of the final fitness value for the conventional and the proposed BPSO algorithms using 7-particle swarm. Eq.(1.3), but slows down faster, and consequently, converges faster due to the elimination of the personal best term. Same trends are observed in other dimensions which are not shown here. 15-Segment triple-frequency design As the second goal, in order to achieve wider bandwidth, a design optimization at three frequencies of f1=4.8GHz, f2=5GHz and f3=5.2GHz is studied with a 4-particle swarm. This time the fitness function of Eq.(1.5) has three frequency terms; i.e., N=3 (triple-term). In this case, it is necessary to prohibit the optimization from getting locked in a loop, which decreases S11 at one of the frequencies while leaving the other frequencies with high S11 values. Accordingly, a condition has been added to the code which substitutes any S11 value of less than -15dB with -15dB. Consequently, the swarm will not attempt to improve S11 values at frequencies for which -15dB threshold has been achieved. Figure 2.21 depicts the convergence behavior of the two methods averaged over R=50 runs. 67 Figure 2.19. Final fitness value averaged over increasing number of runs for the conventional and the proposed BPSO algorithms using 7-particle swarm. The superiority, in terms of convergence, of the proposed method is again confirmed. The proposed method gains at least 2.32dB when compared with the conventional method. It is noteworthy that the proposed algorithm achieves -30dB at the 4t h iteration, whereas the conventional one requires 8 iterations to achieve the same. However, at the 8t h iteration, where both algorithms have satisfied f −30dB, the proposed algorithm outperforms the conventional one by 5.47dB. The optimization resolution is assumed to be 66 mm2 . Therefore, it is not possible to further refine the results of triple-frequency designs by either of the algorithms, unless better resolution is considered at the cost of higher optimization length, which is not studied here. In Figure 1.5(a), the best design achieved by the two methods with termination criterion of T=50 iterations is shown. This time, both methods have designed the same irregular shaped structure, shown in the figure, as their best design (in terms of minimum final fitness value) among 50 runs. However, the proposed algorithm achieves the desired objective faster, as it can also be observed from the average trend shown in Figure 2.21. 68 Figure 2.20. Particle position averaged over 50 runs in an arbitrary dimension for the conventional and the proposed BPSO algorithms using 7-particle swarm. Similar results are achieved from 7-particle scenario which is not demonstrated here for the sake of conciseness. 20-Segment single-frequency design The same single resonance optimization of Eq.(1.5) with N=1 and f1=5.1GHz is performed for a 20-segment structure. Five more segments are added in a row on top of the structure shown in Figure 2.17, to convert the 15-segment antenna to a 20-segment one. Thus, the problem dimensionality increases to 19 from 14. Results similar to those of the previous subsections are observed using 2, 3 and 6 particles, which are not shown here due to limited space. In Figure 1.5(b) and (c), the best irregular shaped designs achieved by the two methods, for the 6-particle scenario, are demonstrated. 69 Figure 2.21. Fitness value trend averaged over 50 runs for the conventional and the proposed BPSO algorithms, using 4-particle swarm. 2.3.2 Optimization length comparison summary In previous parts, the optimization length dependency on the fitness function complexity, termination criteria and particle number was briefly addressed in several MA designs, particularly aiming at real-time applications of BPSO. It is shown that the optimization length is almost proportional to the number of particles. Thus, decreasing the swarm population is desired, but the possibility of the optimization goal being satisfied should also be taken into account. This possibility is highly dependent on the problem itself. In order to quantify optimization length, Table 2.1 depicts the required run-time in several design scenarios. From the exhibited 50-run results, one can arrive at the conclusion that the proposed method achieves an optimization time that is at least 40% shorter than the conventional method. To elaborate, in Table 2.1, a time period of 9.5min is estimated for a 25-iteration run in a 2-particle, single-term fitness function scenario in the case of the 20segment design. However, on the average, 12 and 6 iterations are enough for the conventional 70 and the proposed methods, respectively, to meet the success criterion of -10 dB. Thus, the real design times are roughly 4.75min and 2.37min for the conventional and the proposed methods, respectively. Table 2.1. Required optimization time for a single 25-iteration run in minutes. Single-Term fitness function (N=1) Triple-Term fitness function (N=3) Number of particles m=2 m=4 m=6 m=8 m=2 m=4 m=6 m=8 Run Time for 15-segment Scenario (min) 5 8.5 18 27 5.2 12.5 18.8 29 Run Time for 20-segment Scenario (min) 5.5 9.5 19.5 26 9.5 16 20 31 Two important points have been considered while preparing data for Table 2.1: 1. The convergence length is affected by the initial values. Therefore, to make the evaluation consistent, same initial values are used for similar cases. 2. For the problems considered in this paper, on the average, the convergence occurs before the 25t h iteration for both methods. Thus, the iteration number is limited to 25. It is observed in Table 2.1 that, if the number of particles is increased by α, the run time increases by β which falls in the range [0.6α, 1.2α]. In addition, moving from 15-segment design to 20-segment design, one certainly observes an increase in the optimization length, but the time increase is not easily predictable, since the convergence time is dependent on the number of particles. The same situation is encountered when moving from single-term to triple-term functions. 71 Finally, it should be noted that simulators with more efficient processing management can be customized to be used in practical SDR applications in order to guarantee lower optimization lengths as compare to those achieved by MATLAB/SONNET arrangement. 2.4 PSO in wearable antenna frequency selectivity3 This section introduces the details of the study which was summarized in section 1.4. This study aims at the need for real time reconfiguration of the wearable antennas (WA’s) as a result of being positioned on an evermoving mass (human body) with changing postures and exclusive electrical specifications. Owing to the algorithm’s flexibility and high convergence speed, modified Binary Particle Swarm Optimization (BPSO) is utilized for the real time reconfiguration of a proposed wearable microstrip antenna (MA) consisting of switchable parasitic elements. The viability of the modifications is validated in terms of convergence speed and ultimate fitness value for a proposed simple microstrip antenna designed to maintain resonating at 2.45GHz, ISM band, in different situations. In addition to eliminating clumsy devices that can tangle in surroundings, the incorporation of the antenna into the clothing has the added benefit of minimum discomfort in long-term usage. This feature is particularly critical in biomedical remote monitoring applications, where bio signatures should be tracked continuously. In telecare procedures that last for months or years, the significance of the patient comfort increases even more. In addition to patient comfort, WA’s should maintain acceptable performance during monitoring period. It has been shown in literature that various factors, such as wrinkles and bends of the antenna surface or moisture absorbed by substrate textile, degrade the WA 3 c 2011 IEEE. Adapted, with permission, from Modiri,A. and Kiasaleh, K., Real time reconfiguration of wearable antennas, Proceedings of IEEE Topical conference on Biomedical Wireless Technologies, Networks and Sensing Systems (BioWireleSS), Jan./2011 for section 2.4 . 72 performance Hall et al., 2006). These incidents are easily caused by different body movements, poses, environmental alternations, and accidents. It should be noted that the telecare devices are even used by children who don’t realize the safety rules of the device. Therefore, real time self-tuning feature of the antenna can largely decrease the health risks created by these types of variations causing antenna malfunctioning. 2.5 Modeling antenna bend effect In (Hall et al., 2006), a 56x51mm2 textile patch antenna is fabricated on a 3mm fleece substrate to resonate at 2.45GHz, the ground plane size is mentioned to be 76x71mm2 and the patch is fed by coaxial line at distance of 19mm from the edge. As it is shown in Figure 1.6, the antenna was refabricated and its frequency behavior was tested using Agilent E5071C vector network analyzer (VNA) and Agilent 53708 coax cable. VNA calibration was done by Agilent 85033E 3.5mm calibration kit. The resonance frequency of the fabricated antenna is lower than 2.45GHz (2.155GHz ), which is the result of using different textile from the one mentioned in (? ). The dielectric constant of the textile used here is larger than 1.1 which is mentioned in (Hall et al., 2006). Figure 2.23 depicts the measurement results related to the effect of antenna bending on its frequency behavior. It is observed that bends of 45 and 90 degrees result in frequency shifts of 144MHz and 250MHz. It is noteworthy that only due to a one-directional bend, the antenna completely repels the original frequency band for which it was designed. This experiment simply confirms the need for antenna reconfiguration. In order to compensate for the frequency hop resulted from antenna bending, parasitic narrow patch elements are proposed to be added to the main antenna body as it is shown in Figure 2.24. In the case of the antenna bending, these parasitic patches should be connected to the main antenna body in order to increase the actual length of the antenna, and consequently, compensate for the decrease in electrical length of the antenna.This method has been verified by experiment. Figure 2.25 depicts that, the three switches shown in Figure 2.24(a) are able 73 to return the resonant frequency back to the original one. In fact, one parasitic element and three switches, as shown in Figure 2.24(a), can not support more than five cases (one flat and four bend cases). Of course, it is not straightforward to simulate a variable bend by Sonnet or any other electromagnetic simulation 2D or 2.5D tool. However, it is necessary to input a variation for the optimizer to imitate the practical case. In reality, a bend or a twist causes power drop at the desired frequency. This effect is recognized by power sampler and then the optimizer is enabled. Thus, for the sake of creating a test environment, in this paper, the mechanical variation is modeled by an electrical variation which is decided to be dielectric constant variation in the substrate. For sure, this doesn’t mean that a bend can be easily and completely modeled by permittivity change. Nevertheless, it was shown in Figure 2.23 that bending the antenna results in its resonating at a higher frequency which means that the electrical length of the antenna has decreased. Since the power at the desired frequency is the only required metric, the bending phenomenon can be imitated by changing the antenna substrate by a lower permittivity material. Re-designing and optimizing the same antenna in Sonnet to resonate at 2.45GHz, patch lengths of 53mm and 44mm are achieved for substrate permittivity values of 1.1 and 1.7, respectively. The optimized patch width is 49mm and the feeding point distance from the patch edge is 14.5mm. The permittivity of the original substrate of the antenna is selected to be 1.7. Four parasitic elements and eight switches are assumed as it is shown in Figure 2.24(b). PSO algorithm is used to decide which one/ones of the switches should be on in order to shift the resonant frequency of the antenna back to 2.45GHz. Of course, the higher the number of switches is, the better the frequency response becomes. However, it is not practically reliable to incorporate large number of switches in a wearable device. Therefore, we have assumed eight switches in this paper. Each switch is modeled by a 2x1mm2 conductive tape in its on mode. 74 (b) (a) (c) Figure 2.22. Measurement arrangement for the antenna in (a)flat,and (b)bent modes. (c)Switchable antenna with compensating parasitic element in fixed bent mode. 0 −5 |S11|[dB] −10 −15 Flat (OFF−OFF−OFF) 45deg (OFF−OFF−OFF) 90deg (OFF−OFF−OFF) −20 −25 −30 −35 2 2.1 2.2 2.3 2.4 2.5 2.6 Frequency(GHz) 2.7 2.8 2.9 Figure 2.23. |S11 | measurement results for the flat and bent cases. Here, an eight-dimensional problem is solved by BPSO in which the swarm population is assumed to be 4. It was shown before that the optimization length is almost directly 75 Parasitic elements 3mm 2mm Switch S1 S2 S3 1mm 9mm S1 S3 S2 S4 1mm 11mm 44mm 56mm 49mm S6 19mm 51mm (a) S5 14.5mm S8 S7 Switch (b) Feeding point Figure 2.24. Reconfigurable antenna with (a)three and (b) eight switches. proportional to the swarm population. Thus, decreasing the swarm population speeds up the convergence trend, but the optimization goal may not be satisfied with low number of individuals. Therefore, swarm size is highly dependent on the problem itself. The optimization goal is to minimize the magnitude of the reflection coefficient of the antenna at the feeding point, at 2.45GHz.In order to compare the convergence performance of the proposed BPSO with the conventional BPSO, three different cases are optimized using both methods. Since high optimization time cost is not acceptable in real time applications, the termination criterion is set to be the maximum iteration number of 50. It will become clear in the next section that 50 iterations are more than enough for this problem. In order to present more consistent results, 20 runs have been executed for each of the three cases and both average and individual results are depicted. 76 0 −5 |S11|[dB] −10 −15 −20 −25 −30 −35 −40 1.5 20deg (OFF−ON−OFF) 45deg (ON−OFF−ON) 70deg (ON−ON−OFF) 90deg (ON−ON−ON) 2 Frequency (GHz) 2.5 3 Figure 2.25. |S11 | measurement results for different switch arrangements compensating for different bends 2.6 Results Modeling the effect of different bend angles with different dielectric constants of the antenna substrate, the optimization algorithm decides what the arrangement of the ON/OFF switching modes should be for the eight switches of Figure 2.24(b) in order to retain 2.45GHz as the resonant frequency of the antenna. Figure 2.26 shows the average convergence behavior of both optimization algorithms in three different cases of ǫr = 1.1, ǫr = 1.4, and ǫr = 1.7, respectively. It is observed that the algorithm proposed in Eq.(1.3) outperforms the conventional one in the two first cases by 8.5dB and 3dB, respectively. In the third case, both algorithms have similar behavior and their difference is limited to tenths of dB. It is important to mention that, an |S11 | (reflection coefficient magnitude) of -10dB is practically acceptable in most applications. Therefore, a termination criterion of -10dB had to be added to the algorithm. However, since the convergence behaviors of the two algorithms 77 were supposed to be compared, this criterion was not utilized in this paper. Nonetheless, Figure 2.26 depicts that the proposed algorithm meets the success criterion much faster. Figure 2.27 shows the final fitness values achieved by each algorithm after 50 iterations in all 20 runs. It is observed that the proposed algorithm converges to lower values with higher probability. Studying Figure 2.26 and Figure 2.27, one can conclude that the modification of the initial velocity has offered higher chance to the swarm particles to start the optimization procedure from local solutions close to the best solution. On the other hand, omitting the personal best term speeds up the convergence trend by slightly limiting the exploration ability of the swarm. Although both algorithms eventually converge to almost the same fitness values, it is noteworthy that the proposed algorithm satisfies the design goal in less number of iterations, and therefore, it is more suitable for real-time applications. The best designs (in terms of low final fitness value) achieved by the two algorithms for the eight switches shown in Figure 2.24(b) are [on on on on on on on on], [on on off on off on off on], and [on on off off on off on off] for the three permittivity cases of 1.1, 1.4, and 1.7, respectively. The time period required for each iteration varies by permittivity value and the arrangement of the switches. However, in this problem, an average of 33sec of computing is required for each iteration on a standard 2.83GHz and 3.23GB RAM PC. In fact, simulators with more efficient processing management than that of Sonnet-Matlab arrangement can be customized for practical real time applications in order to guarantee faster computations. Nonetheless, the comparison results related to the optimization algorithms are valid regardless of the computation speed of the computer. In fact, Figure 1.7 in chapter one depicts the measurement results of the antenna structure shown in Figure 2.24(b)in a number of flat and bent cases. It is observed that, with all switches in off mode, the antenna resonates at 2.42GHz in its flat mode. Since the antenna length (44mm) should almost be equal to half wavelength at its resonant frequency, it can 78 be concluded that the substrate permittivity is almost 1.98. It is observed that bending the antenna by 30deg and 45deg shifts the resonant frequency to 2.55GHz and 2.63GHz, respectively. With respect to the known relation between permittivity and resonant frequency, the 2.63GHz resonance can be modeled by a permittivity value of almost 1.7. Therefore, the switch arrangement achieved for permittivity of 1.7 in previous paragraph is used to compensate for the bend effect. It is observed that 45deg shift effect can be successfully recovered by the predicted switch arrangement but this arrangement can not compensate for 30deg bend properly. It is also shown in Figure 1.7 that if the same switch arrangement is used while the antenna is in its flat mode, the resonant frequency jumps down to 2.37GHz, as expected. Fitness Value(dB) Fitness Value(dB) Fitness Value(dB) 79 Conventional BPSO Proposed BPSO 0 −5 −10 −15 −20 0 10 20 30 Iteration Number 40 50 (a) 10 20 30 Iteration Number 40 50 (b) 10 20 30 Iteration Number 40 50 (c) −5 −10 −15 −20 −25 0 −18.5 −19 −19.5 −20 −20.5 0 Figure 2.26. Average fitness value trend over 20 runs for (a)ǫr = 1.1, (b)ǫr = 1.4, and (c)ǫr = 1.7 80 Conventional BPSO Proposed BPSO 0 −10 Final Fitness Value (dB) −20 −30 0 5 10 Run Number 15 20 (a) 5 10 Run Number 15 20 (b) 5 10 Run Number 15 20 (c) 0 −10 −20 −30 0 −20 −20.5 −21 0 Figure 2.27. The final fitness value of each run for (a)ǫr = 1.1, (b)ǫr = 1.4, and (c)ǫr = 1.7 CHAPTER 3 PROPOSED MBMD ARRANGEMENT FOR BREAST CANCER DETECTION This chapter reviews the author’s studies on MBMD technology and describes the novel MBMD technique which was proposed by the author and her supervisor. It is worthy of acclaim and emphasis that since microwave radiation is non-ionizing and microwave technology is relatively cheap and reliable, MBMD is a promising modality for development of self-/physician-assisted screening tools in the case of breast cancer detection where routine checkups are required. However, MBMD is in its early stages of development and many unknowns and ambiguities pop up when one tries to approach any of its various aspects. The author has studied both hardware and software aspects of the problem. The ultimate goal of the study has been to introduce a preliminary model of a self examination breast monitoring tool. The proceeding sections present the challenges faced and accomplishments gained in the course of this investigation. This chapter starts with the PSO-based algorithm proposed to solve the inverse problem for MBMD technique and continues with explaining the proposed MBMD radiator structure and its simulation and measurement results. A summary of this study was given in section 1.5. Here, more details are presented. 81 82 3.1 PSO in solving the inverse problem1 Basically, in MBMD, the scattered/transmitted fields from/through an inhomogeneous structure at microwave frequency band are collected and analyzed in order to define the composition of the object under the test. The dielectric constants of the contents are often calculated as the indicators of the material type. A comprehensive review of microwave diagnosis applications are found in (Hassan, 2011; Nikilova, 2011; Golnabi, 2011). In these studies, MBMD is used to create an image of the human body by highlighting the contrast between dielectric constants of different tissues. MBMD techniques deal with inverse electromagnetic problems, for which a variety of algorithms have been proposed and studied in the literature (Golnabi, 2011). It is quite evident that the algorithm needs to be relatively fast and reasonably accurate for biomedical diagnostics. However, such complex inverse problems potentially end up with non-unique results. For that reason, global optimization algorithms prove to be more appropriate computational techniques as compared to the analytical optimization methods. There are some articles in literature which use evolutionary algorithms for non-biomedical problems, see (Cui, 2005) and (Genovesi, 2007). These problems are considerably simpler that the MBMD problem due to dealing with low loss or lossless materials. When using evolutionary algorithms, the main challenge is to analyze and classify the results obtained by a large number of agents at each iteration, and to lead the algorithm in the correct path accordingly, in order to ultimately achieve the best estimation. That is the correct composition of the tissue under the test. A burden in this study is the penetration depth which is restricted by the high absorption rate of the electromagnetic energy by body tissue at frequencies of interest to the MBMD studies. Yet, some parts of human body, such as 1 c 2012 IEEE. Adapted, with permission, from Modiri, A. and Kiasaleh, K., A novel discrete particle swarm optimization algorithm for estimating dielectric constants of tissue, Proceedings of 2012 Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), Sep./2012 for section 3.1 . 83 breast, stay suitable candidates for MBMD, firstly because they are not consisted of muscle or bone. It is well known that global optimization algorithms are inherently problem-dependent and need to be customized according to the problem introduced to them. As it was reported earlier in this manuscript, several variants of PSO were developed for different electromagnetic problems. Here, the PSO variant proposed for MBMD estimates the permittivity arrangements of lossy multi-layer structures of body tissue models. The impact of various parameters, namely, the MBMD frequency, the immersion medium, the number of agents, the smoothing coefficient, and the maximum velocity are studied on the estimation performance, in terms of the maximum estimation error. It is demonstrated that by choosing the parameters correctly, one can expeditiously achieve estimation results with a maximum error less that 10%. In (Modiri, et. al., 2011c), it was demonstrated that PSO variants are able to successfully act as detection algorithms in a complex MBMD problem. There, two proposed versions of PSO were used based on complex number and discrete PSO. Here an even improved PSO variant is introduced based on (Modiri, et. al., 2012). Two main factors that make this study distinct from the other ones available in literature are • The new versions of PSO introduced, • The measurement-based exact loss factors used (see (Lazebnik, 2007)). The loss factors have generally been ignored or over simplified in order to decrease the complexity of the problem in other studies (Gandhi, 2010; Yeung, 2009). 3.1.1 Problem scope and estimation algorithm The proposed DPSO algorithm finds the tissue-types and their corresponding dielectric constants in a hierarchical manner (see Figure 3.1). The makeup of each tissue layer is first 84 selected from a set of six possible tissue groups, namely, (t1: skin), (t2: adipose dominant tissue), (t3: 30%-80% fatty tissue), (t4: glandular tissue), (t5: blood) and (t6: malignant tissue), see (Lazebnik, 2007) and (Clegg, 2010). For the sake of conciseness, the tissue-types will be called with their t-numbers. These six groups are, in fact, the tissue-types one can expect to find inside the breast. After tissue-type selection, the best complex values of the dielectric constants are found inside the variation spans assigned to those tissues in the literature, (Lazebnik 2 , 2007). In other words, here, the inherent variations of the dielectric constants of the tissues are taken into account. These variations are created by differences which exist either between different people or even in a single person, under different conditions. Error threshold is considered to be 10% as it is shown in Figure 3.1. This means that, if all the estimated dielectric constants have an error of lower than 10%, then the second optimization cycle is activated. It is noteworthy that the threshold of 10% does not address a perfect estimation, however it is acceptable in MBMD. The reason for this claim is that the chance of detecting the tissue-type incorrectly has been shown to be very low at this threshold. In addition, 10% variation can even happen due to inherent inaccuracies in the process. It was explained in chapter one that the popular PSO velocity function is used for binary PSO, as well. The only requirement is a mapping function which maps the real values to zeros and ones (see equation 12 in (Jin, 2007)). Likewise, for DPSO, a similar mapping function is introduced in (Modiri, et. al., 2012) to address the six required levels. The mapping function is shown in Eq.(3.1). S = 0.5 + NT /(1 + exp(−ν × sc)) (3.1) where ’sc’ is the smoothing coefficient and NT is the number of tissue types which is equal to six for the breast tissue considered in this study. ν is the velocity and is a real number. Figure 3.2 shows the mapping function using different smoothing coefficients. It 85 Optimization1: Estimate the tissue−type for each layer No Error < Threshold? Yes Optimization2: Estimate the complex dielectric constants Figure 3.1. The block diagram of the proposed PSO-based optimization procedure consisting of two customized optimization cycles. should be noted that, in order to have a correct mapping, the maximum allowed velocity should be selected with respect to ’sc’. Any velocity value mapped inside [a − 0.5, a + 0.5) in Figure 3.2 is interpreted as ’a’ in order to create discrete values. Equation (3.2) shows the fitness function that is minimized by DPSO. ’R’ and ’T’ depict the reflectance and transmittance coefficients. f = X {||Ractual | − |Restimated || + |arg(Ractual ) − arg(Restimated )| + ||Tactual | − |Testimated || + | arg(Tactual ) − arg(Testimated )|}; (3.2) 86 6.5 5.5 S 4.5 Vmax=1.5 & sc=5 Vmax=6 & sc=2 Vmax=12 & sc=0.5 Vmax=9 & sc=1 3.5 2.5 1.5 0.5 −15 −10 −5 0 Velocity 5 10 15 Figure 3.2. Mapping function with different smoothing coefficients. where arg () denotes the phase of the enclosed complex number. The normal incidence of plane wave to multi-layer structure is considered to solve the forward problem (Orfanidis, 2008). Here, both magnitude and phase errors are taken into account in Eq.(3.2), as such quantities bear distinct information relevant to the optimization process (Orfanidis, 2008). An immersion medium is considered to surround the tissue, as it is recommended in many other research articles to improve MBMD detection performance (Nikolova, 2011; Gonabi, 2011). The immersion medium is assumed to be similar in properties to the fatty tissue (t2). Two types of controlling parameters are involved in this problem: • MBMD parameters • DPSO parameters MBMD parameters are the immersion medium, tissue thickness and the radiation frequency. The controlling gears of DPSO are the maximum allowed velocity (Vmax), which is related to the smoothing coefficient in the mapping function, and the number of agents. The effects 87 of these parameters are studied in the following section in terms of the estimation error in 100 independent runs per analysis. Each run is limited to 1000 iterations for the first optimization cycle shown in Figure 3.1 and 100 iterations for the second one. These restrictions were imposed by the knowledge gained from trials and errors for the sake of optimizing both the final error and the computation time. To be able to compare the results with those presented in (Modiri, et. al., 2011c), the tissue is modeled as a 5-layer structure (t1, t2, t4, t2, t1) for which it is shown later that 40-agent swarm (m=40) is fairly strong to gain an accuracy better than 10%. Reviewing the model, one can easily recognize that this is a model of normal breast tissue. The thicknesses of the tissues are t1:2mm, t2:15mm, and t4:10mm. The total thickness is 4.4cm, which represents a good model for the breast tissue when the tissue is squeezed for MBMD measurement. Although other frequencies are also studied, the reference frequency in this paper is 1GHz, which has been proved to be appropriate for MBMD (Hassan, 2011; Nikolova, 2011). 3.1.2 Parameter variation study MBMD parameters As the first MBMD parameter, the immersion medium is studied. This medium acts as matching interface for the fields to enter the body tissue with lower reflections. Figure 3.3 shows the effect of using different types of immersion media on the detection error in 100 runs. It should be noted that, unlike previous studies associated with MBMD immersion medium, here, the impact of this intermediate medium on the optimization trend of the search algorithm is highlighted. The dielectric constants of air, water and fat at 1GHz are assumed to be 1, 78.2 − 3.796j and 4.79 − 0.8j, respectively. As it is evident in Figure 3.3, the probability of achieving a final estimation with error higher than 10% increases to 6% in the case of water. It is interesting that this increase in error happens for imaginary part of the dielectric constant which has lower significance in identifying the tissue type. Maximum Error Percentage 88 15 Maximum Error Percentage Fat Imaginary 10 5 0 10 20 30 40 50 60 Number of Runs 70 80 90 100 Air Imaginary 15 Air Real 10 5 0 Maximum Error Percentage Fat Real 10 20 30 40 50 60 Number of Runs 70 80 90 100 Water Real 20 Water Imaginary 10 0 0 10 20 30 40 50 60 Number of Runs 70 80 90 100 Figure 3.3. The effect of immersion medium on detection error for a 44mm-thick tissue at 1 GHz. Vmax, sc and m are considered to be 1.5, 5 and 40, respectively. However, the mean error percentage for all three cases remains between 7% and 9%. It is worth mentioning that for the case of fat, the maximum required iteration number is 191, which is almost 20% of the maximum iteration number of 1000 (see Table 3.1). This study simply confirms that oily and fatty substances are preferred for MBMD. Figure 3.4 shows the estimation error at two other frequencies, 500 MHz, and 5 GHz. 5 GHz case, apparently, has the highest probability of error while 1 GHz (Figure 3.3) has the lowest probability of error. It is interesting that the standard deviation of the results achieved from 5 GHz study is more than four times that of 1 GHz study. The mean error percentages, however, for all the cases remain between 7.5% and 12% (see Table 3.1). Although 500 MHz radiation has higher chance of penetration than its 1 GHz counterpart, the relative thicknesses of the tissue-layers also comea into picture and introduces a new trend. Maximum Error Percentage Maximum Error Percentage 89 500MHz Real 20 500MHz Imaginary 10 0 0 10 20 30 40 50 60 Number of Runs 40 70 80 90 100 5GHz Real 5GHz Imaginary 20 0 0 10 20 30 40 50 60 Number of Runs 70 80 90 100 Maximum Error Percentage Figure 3.4. Estimation error at some popular MBMD frequencies for a 44mm-thick tissue (V max = 1.5, sc = 5, m = 40). Immersion medium is considered to be fat. 20 59mm Real 15 59mm Imaginary 10 5 0 0 10 20 30 40 50 60 Number of Runs 70 80 90 100 Figure 3.5. The error for thicker tissue(f = 1GHz, V max = 1.5, sc = 5, m = 40). In practice, the tissue thickness is defined by the person’s body style (fat, skinny, etc.). Here, we repeat the study for another case in which the thicknesses of the t2 and t4 tissues are increased by 5 mm. The total tissue thickness is 59 mm in this case. As shown in Figure 3.5, the probability of achieving an estimation with an error level higher than 10% increases to 3% when considering fat as the immersion medium. Nonetheless, the mean error percentage for all the cases stays in the range of 7.5-8.5%. DPSO parameters MBMD parameters studied in previous section are very important; however, they are not fully under the control of the MBMD operator. Optimization algorithm, on the other hand, 90 can be controlled with less limitations. In this section, the impacts of some of the optimization parameters on the estimation process are studied. It should be noted that, correctly mapping the real velocity values to discrete values is quite challenging. The first step is to define the smoothing coefficient of the mapping function. By decreasing sc, introduced in 3.1, the probability of achieving errors higher than 10% increases. Further increasing of sc above 5 will not result in any further improvement either. Therefore, here we demonstrate our results mostly considering sc =5. The second issue is the maximum velocity. In discrete versions of PSO, due to intrinsic limitations imposed by mapping function, V max does not have the same direct impact on the performance as it has in the real number PSO case. Reviewing Figure 3.2, one can simply conclude that V max should be selected with respect to sc, not only to cover all six levels, but also to preserve a smooth variation. As shown in Figure 3.6 and Figure 3.3, the average errors for sc = 0.5, 1, 2, 5 and 10 are 8.5, 8.1, 7.8, 8.1 and 7.7, respectively. Lower V max offers higher estimation accuracy; however, it also increases the required number of iterations in order to converge. On the other hand, higher velocities decrease the chance of finding the solution. sc = 5 and V max = 1.5 are the best parameter set for keeping the error below 10% in all 100 runs for this problem. Finally, the number of agents is studied. It was shown in chapter 2 that, although having large number of agents improves the search activity, it also increases the optimization cost in terms of computation resources and optimization length. Therefore, it is always desired to find an optimum agent-number. As it is shown in Figure 3.7, lowering the number of agents from 40 to 30 increases both the probability of achieving higher errors than the desired 10% and the average maximum error. Furthermore, increasing the number of agents to 50 decreases the average error from 8% in the case of 40 agents to 7.8%, which is not a significant improvement. Maximum Error Maximum Error Percentage Percentage Maximum Error Maximum Error Percentage Percentage 91 40 sc=0.5 − Vmax=12 − Real sc=0.5 − Vmax=12 − Imaginary 20 0 0 10 20 30 40 50 60 Number of Runs 70 80 90 100 sc=1 − Vmax=9 − Real sc=1 − Vmax=9 −Imaginary 20 10 0 0 10 20 30 40 50 60 Number of Runs 70 80 90 100 sc=2 − Vmax=6 − Real sc=2 − Vmax=6 − Imaginary 20 10 0 0 10 40 20 30 40 50 60 Number of Runs 70 80 90 100 70 80 90 100 sc=10 − Vmax=0.65 − Real sc=10 − Vmax=0.65 − Imaginary 20 0 0 10 20 30 40 50 60 Number of Runs Maximum Error Percentage 20 Maximum Error Percentage Figure 3.6. Estimation error for a 44mm-thick tissue at 1GHz when assuming different maximum velocities and smoothing coefficients. 40 agents are considered. 15 30 agents − Real 30 agents − Imaginary 10 0 0 10 20 30 40 50 60 Number of Runs 70 80 90 100 50 agents − Real 50 agents − Imaginary 10 5 0 10 20 30 40 50 60 Number of Runs 70 80 90 100 Figure 3.7. Estimation error for a 44mm-thick tissue at 1GHz when assuming different number of agents. V max is 1.5 and sc is 5 Table 3.1 summarizes the optimization length for some of the cases studied. Comparing these results with those shown in (Modiri, et. al., 2011c), one can simply conclude that the 92 Table 3.1. Optimization length comparison Medium air water fat fat fat fat fat 1 2 3 F(GHz) 1 1 1 5 0.5 1 1 m1 40 40 40 40 40 30 50 max(iter 2 ) 33 1000 191 1000 1000 1000 190 mean(iter) 222 158 50 454 370 93 47 P 3 (high error) 0 6 0 40 26 0 0 Number of agents Required number of iterations for lower than 10% error Probability of achieving higher than 10% error new version of DPSO introduced herein has significantly enhanced the detection performance by decreasing the estimation time by 85%. 3.2 MBMD proposed radiator design2 A new MBMD structure proposed by the author and her supervisor is studied in this section (Modiri, et. al., 2013a). The novelty of this design resides in introducing an easily implementable structure for in-vivo, non-invasive, breast monitoring with great potential of supporting self-examination and promising detection performance. MBMD has been widely studied for cancer detection recently. In general, either an array or a scanning antenna is used in MBMD. Employing an array has the advantage of avoiding the mechanical issues of a scanning antenna. Moreover, a scanning antenna is not a practical arrangement for self-examination purposes. However, designing the array still remains an important subject in MBMD technology. 2 c 2013 IEEE. Adapted, with permission, from Modiri,A. and Kiasaleh, K., Characterizing a proposed sixteen-element array antenna designed for microwave imaging of breast cancer, Proceedings of IEEE Topical conference on Biomedical Wireless Technologies, Networks and Sensing Systems (BioWireleSS), Jan./2013 for section 3.2 . 93 The deficiencies of X-ray, such as relatively large false negative rates and ionizing property, have made scientists seek an alternative modality, especially for the situations where successive imaging tests are required (Nikolova, 2011; Hassan 2011; Fear, 2003). Breast cancer monitoring tests fit in this very scenario; hence, MBMD has been receiving a great deal of interest and attention in the case of breast cancer (Paulsen, 2005; Fear, 2002; Irishina, 2009). In the ensuing parts of this dissertation, the development stages of the radiator structure are explained first in simulation and then in practice. The 3D simulations are performed in Ansoft HFSS. Since the array had to be characterized in an environment similar to the actual measurement, Ansoft digital female phantom is used for the majority of simulation studies. The Ansoft digital phantom is a reliable model for microwave studies as it provides a relatively accurate model of the frequency dependent glandular and fatty tissues in the breast. For measurement studies, however, the radiator structure is built and breast tissue phantoms are created. A donated tumor tissue was used as the target to be detected. 3.2.1 The array structure As it is shown in Figure 3.8, the antenna array consists of sixteen identical bent dipoles, separated from each other by 22.5 degrees. These antennas are located on the perimeter of a circle which goes around the breast tissue. The size of the circle is dependent upon the size of the breast. The phantom breast size is considered here in order to define the diameter of this circle (100 mm). To better clarify the dimension and the shape of the antennas in a 2D space, by fixing the feeding point at the center of the coordinate system (0,0), the top most and the bottom most edges of the antenna arms will be located at (8.5mm,27.5mm) and (19 mm, 20.5mm), respectively. The radius of the antenna wire is assumed to be 0.4mm. This antenna is designed to resonate at 1.2GHz in the presence of the tissue. The spaces in between the antennas and the tissue are filled with a fat-type material as immersion medium (Golnabi, 2011). The dielectric constant of fat is used for the immersion medium. This layer 94 Figure 3.8. The Ansoft digital phantom breast tissue and the antenna array. of fat even covers the antennas. The thickness of this fat layer is assumed to be 17 mm. The fat medium builds a hemisphere with the radius of 67 mm, which covers the antennas and the breast tissue, see Figure 3.8. Due to the relatively irregular shape of the breast tissue, there is no systematic way to calculate these dimensions; therefore, they are achieved by trial and error. 3.2.2 Analysis results In practice, the elements of the array are excited one at a time while the other antenna elements act as receivers (Sill, 2005). Figure 3.9 shows the reflection coefficient for five elements of the array. Parameter t in Figure 3.9 denotes the angular location of each antenna. 95 Figure 3.9. The reflection coefficient of the five elements of the array. The sixteen antennas have almost similar resonance behavior. The curves related to the other eleven elements of the array are not shown in Figure 3.9 in order to keep the figure easily readable. However, for these remaining eleven elements, the reflection coefficient magnitudes (—S11—) are shown only at 1.2GHz which is the desired radiating frequency. The straight dipole structure is bent and optimized in order to present a structure that almost follows natural body curves and is also smaller in size when compared to the conventional straight dipoles. Figures 3.10 and 3.11 show the near field radiation patterns of the antenna at two different distances (10 mm and 50 mm). In HFSS, two spheres with the aforementioned radii are created around the radiating antenna and the fields are calculated on those spheres and plotted as it is seen in these figures. It should be noted that in these types of applications, near field study is always necessary. Figure 3.10 shows the total Electric field pattern and 96 Figure 3.10 shows the polarization-based patterns. Two important points are observed in the figures: • A relatively large portion of the power is reflected back from the tissue and scattered in the surrounding space. Although the immersion medium is present, the major part of electromagnetic energy is not able to bypass the material discontinuity. • As expected, the largest component of the electric field (Ey) is the polarization, which better complies with the orientation of the dipole. Ex and Ez are lower in intensity by almost 6 dB and 3dB, respectively. However, the radiation at different polarizations, due to both the bent structure of the antenna and scattering phenomenon, sets the stage for detecting polarization dependent malignancies. To have a more comprehensive study, the electric field distributions, inside and outside the tissue, are shown in Figure 3.12 from two different views. This figure actually gives an idea to the interested reader of how far the radiation from the sixteen antennas can penetrate inside the tissue, what the electric field distribution is, and how much energy is scattered in the surrounding environment. The values shown in the color bar of Figure 3.12 are for the case of having an excitation of 1W; however, they can be scaled to any desired intensities. To better illustrate, Figure 3.12c shows the field distribution in the fatty tissue and glandular tissue separately. The bent antenna shown in the figure is the transmitter. Finally, Figure 3.13 shows the transmission coefficient when one of the antennas is transmitting and the other fifteen are receiving. Many parameters, mainly, the multipath effect, surface waves traveling on the skin, absorbance and scattering phenomena create such an |S21| distribution. Although transmittance is not the only detection factor considered and reflection also plays an important role, |S21|′s do not seem to be large enough to support a reliable detection. Therefore, in the pursuing section, the radiator structure is modified toward a better performance. 97 Figure 3.10. The radiation pattern of the radiator when only one of the antennas is radiating in the plane of θ = 90o . Total electric field is shown for the two desired scenarios. 3.3 MBMD proposed radiator design with conductor cover3 As it was shown in the previous section, the transmittance coefficient needs to be enhanced in the proposed radiator. There are a few articles that introduce a conductor enclosure in order to improve the detection sensitivity of non-invasive MBMD. In these articles, the enclosure encompasses both the entire human body and the radiating setup (Gilmore, 2008; Mojabi, 2010). In other words, these designs require that the patient or the tissue under the test is placed in a box similar to a Faraday cage in order to enhance the detection chance. Here, it is shown that a conductor cover, or dome, added to the proposed radiator design 3 c 2013 IEEE. Adapted, with permission, from Modiri, A. and Kiasaleh, K., Non-invasive Microwave Breast cancer detection - A Comparison Study, will appear in proceedings of 2013 IEEE Global Humanitarian Technology Conference (GHTC), Oct./2013 for section 3.3 . 98 100 200 300 400 500 600 100 200 300 400 500 600 700 Figure 3.11. The radiation pattern of the radiator when only one of the antennas is radiating in the plane of θ = 90o . Electric field components are shown separately. ameliorates the performance of the proposed MBMD structure noticeably. The proposed conductor cover does not need to encompass the entire breast (which is impossible for an in-vivo measurement). Instead, it is designed to cover the frontal section of the breast. This way, a preliminary model of a portable breast examination device with a great potential as a self-examination tool is introduced. With the aid of simulation, the difference between the scenario in which a tumor exists inside the breast tissue and the scenario in which the breast tissue is healthy are demonstrated for the two cases of having the conductor dome in the measurement setup and when it is absent. The tumor detection sensitivity, as the performance indicator, is defined as the cumulative difference between the received electric field intensities of the normal and cancerous tissues at different measurement points. The phase sensitivity is also calculated in a similar way. Each simulation is conducted when only 99 Figure 3.12. The electric field distribution (a) inside the tissue in xy plane, (b) outside the tissue in xy plane and (c) inside the tissue in zy plane when the dome is not present. one of the antennas is set to be transmitting. Then, the same procedure is repeated for the other elements of the array, allowing for each element of array to act as a transmitter at some point in the experiment. It should be noted that similar field measurements, both magnitude and phase, have been 100 Figure 3.13. The transmittance coefficient of the array elements. used in the past for cancer detection purposes, see (Klemm, 2010; Dun, 2003). This contribution follows on a similar path. As the main goal of this section, however, the performance improvement using the proposed radiator structure along with the conductor dome is underscored. It is also shown that the detection chance is affected by the polarization of the electric field (Cho, 2006; Arunachalam, 2008; Woten 2008). We observe a correlation between the detection parameters and the polarization of the transmitted wave in this study which is similar to those of the previous studies. It is noteworthy that, due to the high memory requirement of the simulation, a workstation with very high RAM is employed. The simulations are demonstrated at 1.2 GHz. This frequency is within the frequency range which allows one to achieve the required resolution and the sufficient penetration depth for breast cancer detection, see (Golnabi, 2011). 101 3.3.1 Proposed structure and 3D model The structure of the proposed radiator is shown in Figures. 3.14 and 3.15 along with the Ansoft female body and breast model. The radiator consists of an array of sixteen dipole antennas inside the proposed conductor dome. In order to create the dome model in HFSS, a sphere was plotted and cut at the borders where it was intersecting the phantom body exterior. The radius of the sphere was assumed to be 67 mm. The following rules are followed to achieve a better implementable design: • (1) The dome has to be as small as possible and yet capable of encompassing the breast tissue and the 16 antennas completely. • (2) An immersion medium is assumed to be separating the antennas and the tissue from each other. The center of the dome sphere is assumed to be on the X axis, shown in Figure 3.15. One of the issues faced in this study is the large computational complexity. Some solutions are needed to make the simulation feasible and still achieve a reliable results. To this end, the following simplifications are considered: • (1) As it is shown in Figure 3.14, the female human body phantom is cut horizontally at two planes and only the breasts and the tissues surrounding them or very close to them are kept in the model. Since human body tissue significantly absorbs the microwave radiation, the effect of this modification is negligible. The dimensions of the model and the surrounding boundary are shown in Figure 3.14. Due to the HFSS modeling rule, the enclosure boundary is set to be more than λ/4, where λ refers to the free space wavelength at the desired frequency of 1.2 GHz. This implies that, there exists a gap of at least λ/4 between the model of the body and the simulation boundaries (see Figure 3.14). 102 • (2) Instead of letting all the sixteen antennas be present in the model at the same time and then letting them radiate one at a time, the simulation process is split into sixteen steps. At each step, only one of the antennas is present, and the electric field is simulated on 101 points on the perimeter of a circle surrounding the breast tissue. This circle is called field-measurement circle. The electric field plots are achieved by creating field reports in HFSS. Optimetrics analysis is performed in HFSS for the sixteen steps. This second simplification reduces memory resource requirement by a sizable margin. However, the author was able to keep all the antennas in the simulation in the studies which will be reported in the later sections. This opportunity was created after upgrading both the computer and HFSS version. As it is shown in Figure 3.15, the feeding points of the sixteen antennas are located on the perimeter of a circle at steps of 22.5 degrees. It should be noted that, this circle is not identical to the field-measurement circle; however, the two circles are located close to each other. The reason for choosing two distinct circles instead of one was to avoid very high field intensities at the vicinity of the antenna feeding points. The field-measurement circle is shown in Figure 3.15. It should be noted that the field-measurement circle encompasses the breast tissue without touching it or intersecting it. Therefore, it is a good representative for data reading points in a realistic scenario. The radius of the field-measurement circle is 50 mm. The breast model in Ansoft female body phantom, as it is shown in Figure 3.14, consists of glands and fat. The glands are spread inside the breast in a web-shaped random structure. In fact, this phantom represents a worst case scenario in which the breast tissue is densely filled with glandular tissue. In the next part, a variety of tumors with different shapes and sizes are introduced to the model and the analysis results of the previously explained 16-step simulation procedure is demonstrated for the two following scenarios: 1. When the outer boundary of the dome is perfect electric conductor, 103 1200mm 2200mm 1815mm 190mm 560mm 180mm 190mm Figure 3.14. Ansoft female body phantom and our separated section of it for the breast cancer study. 2. When the outer boundary of the dome is air. For the tumor, Eq.(3.3) is imported into HFSS in order to model its frequency dependent dielectric constant. The Debye model corresponding to the dielectric constants of the normal and cancer tissues are available in (Lazebnik, 2007). ǫT umor = 7.44048 + 48.5595/(1 + 4.72414 × 10−021 × f 2 ) (3.3) Although the sixteen antennas are identical, the relatively irregular shape of the breast phantom creates some discrepancy in their performances in terms of reflection coefficient. 104 X Bent Dipole Antennas Y Z Figure 3.15. Digital breast phantom along with the proposed radiator structure. 3.3.2 Detection mechanism and analysis In order to compare the two cases of having and not having the conductor dome, the field strengths and phases are found in simulation, and the following two equations are computed: i |End | i arg(End ) = = N X ||E ij j=1 ij normal | − |Etumor || ij |Enormal | N X | arg(E ij ij − arg(Etumor )| ij arg(Enormal ) normal ) j=1 (3.4) (3.5) In the above two equations, The E-field magnitudes are in dB and phases are in radians. The operator ’arg’ is the phase operator and ’i’ refers to any of the coordinate axes x, y or z. The summations in equations (3.4) and (3.5) are over the receiving points N. In fact, the detection probability increases by increasing N. In our simulations, the results are calculated over 101 equally spaced points on the perimeter of the field-measurement circle shown in Figure 3.15; however, considering the ease of implementation, it is more reasonable to use the existing fifteen antennas as the receiving ports (one antenna transmitting and fifteen antennas receiving). The results for both studies (101 and 16 ports) are shown in the ensuing paragraphs. In order to overcome the impact of noise and other factors causing fluctuations in the received field strength, a threshold dependent summation has been plotted; i.e., a 105 threshold is introduced so that only the normalized difference values above that threshold are added up. This way, one can set measurement sensitivity on a case by case basis in order to yield the best performance. In the proposed algorithm, a decision in favor of the presence of an anomaly is made if either of the two measurement results obtained using equations (3.4) and (3.5) show that the field differences in the summation exceed the threshold. It should be mentioned that, in order to model the cancerous tissue, a tumor is added to the normal tissue model. Here, the detection hinges upon changes in the electric field measurements at the receiving antennas in the presence and absence of the cancer tissue. In other words, using this method, an attempt is made to establish the ”normal” response when the cancer is absent and compare the cancer case to that of the normal case. In practice, the normal scenario may be established by repeated measurements of a healthy individual. This information may then be stored and maintained by the individual under investigation. However, if the person starts using this device when the cancer is in its initial stages, the changes due to cancer growth are going to be detected. It should be noted that, this change-based study is done at the absence of the PSO-based algorithm explained previously, just to emphasize the effect of modifying the radiator structure by adding the conductor dome. The simulations’ results for the following tumors are summarized herein: • (1) Spherical tumor with a diameter of D=10mm. • (2) Spherical tumors with diameters of D=5mm. • (3) Spherical tumors with diameters of D=1mm. • (4) Cylindrical tumor with a diameter of D=10mm and height of h=10mm. • (5) Gland-shaped tumor. 106 Results Figure 3.16 shows the locations of the tumors (discussed above) inside the breast phantom from two different views (front and top). An attempt has been made to introduce a variety of shapes and locations in order to have a comprehensive study. In addition to the tumor cases mentioned in the previous section, a number of other cases involving tumors with different shapes and sizes with varying locations have been studied; however, for the sake of conciseness, and also due to the similarity of the results, one set of results, which represent the overall performance of the proposed structure, is chosen carefully. As it is shown in Figure 3.17, for each tumor scenario, there are sixteen simulation results showing the normalized differences between the normal and cancerous cases on the fieldmeasurement circle. Figure 3.17 shows the results for the magnitude of the electric field in the x direction, i.e., |Ex |. Nonetheless, similar results are achieved for magnitudes and phases at the other two coordinates, which means 96 analysis results for each case. Then, equations (3.4) and (3.5) are calculated for different threshold values. In Figure 3.17, the x axis of the sixteen plots shows the distance from a reference point on the field-measurement circle. Since the perimeter of the field-measurement circle is 314mm, the last x-axis value is 314. It should be noted that in some of the plots in Figure 3.17, the dashed curves are very close to the x-axis (almost zero). The thresholds start with zero and do not exceed the peak level of the magnitude of the field for each data set. As it is depicted in Figure 3.16 and other figures of this part, at each step, two scenarios are simulated. In one scenario. the proposed dome structure exists and in the other one it is absent. The curves related to the latter scenario are shown in dashed lines and solid lines are chosen for the former scenario. Figures 3.18 to 3.26 show the 3.4 and 3.5 plots for the tumor scenarios shown in Figure 3.16 when only sixteen receivers are considered on the field-measurement circle. 107 Figure 3.16. Tumors’ shapes and locations inside the breast. Examining Figures 3.18 to 3.26, one can easily conclude that the detection chance is noticeably higher when the proposed structure is used with the conductor dome. Although the results are highly polarization dependent, the aforementioned conclusion is still valid. It should be noted that the dipole antennas are mainly oriented in the x direction. Since the dipoles have bent shape, they are not in fact parallel to the x axis. The main conclusion of the analysis of both x and y axes is that the conductor dome improves the possibility of detecting the presence of an anomaly. However, for many z direction results adding the conductor dome decreases the detection chance. Nonetheless, it should be noted that y direction improvement is much larger than the z direction degradation. Figure 3.18 is related to a spherical tumor with a diameter of 10 mm at distance of 15 mm from the skin surface. This location is referred to as a superficial tumor location. Figures 3.19 and 3.20 show the results for the same tumor when it is pushed back inside the breast tissue by 2.5 cm and 3.5 cm on y direction, respectively. In Figure 3.20, the new tumor location is closer to the chest wall by 10 mm. In Figures 3.21 and 3.22 the tumor diameter shrinks to 5 mm and 1 mm, respectively. The centers of both tumors are placed at a distance of 15 mm from the skin surface. The D=1 mm tumor is studied at the depth of 30 mm from the skin surface in Figure 3.23. To better underscore the polarization dependency, superficial Gland−shape Tumor at Location1 TX Antenna#1 15 10 Normalized |Ex| Difference 0.4 0.2 0 0 100 200 3 300 0 0 400 100 0.2 TX Antenna #3 200 300 400 300 400 300 400 300 400 300 400 300 400 300 400 300 400 TX Antenna#4 2 0.1 1 0 0 Normalized |Ex| Difference TX Antenna #2 0.6 Without dome With dome 5 Normalized |Ex| Difference Normalized |Ex| Difference 108 100 200 1.5 300 0 0 400 TX Antenna#5 1 1 0.5 0.5 0 0 100 15 200 300 100 200 TX Antenna#7 200 TX Antenna#6 0 0 400 10 200 TX Antenna#8 100 5 0 0 100 200 Distance(mm) 1.5 Normalized |Ex| Difference Normalized |Ex| Difference 100 1.5 300 400 0 0 100 2 TX Antenna#9 200 Distance(mm) TX Antenna#10 1 1 0.5 0 0 100 200 200 300 400 100 Normalized |Ex| Difference 200 300 400 0 0 TX Antenna#13 4 2 0.5 100 80 200 100 1.5 1 0 0 200 TX Antenna#12 1 6 Normalized |Ex| Difference 100 2 TX Antenna#11 100 0 0 0 0 300 400 0 0 TX Antenna#14 100 4 TX Antenna#15 200 200 TX Antenna#16 60 40 2 20 0 0 100 200 Distance(mm) 300 400 0 0 100 200 Distance(mm) Figure 3.17. The normalized differences of |Ex | for the sixteen transmitting antennas. 109 and deep cylindrical tumors are also analyzed in Figures 3.24 and 3.25. For the latter case, the tumor is pushed back by 35 mm in the y direction shown in Figure 3.15. Finally, a superficial tumor with the exact shape of a gland, placed very close to an actual gland, is investigated in Figure 3.26. Studying the aforementioned figures, one can easily realize the signature degradation in the case of deeper tumors. Cylindrical tumor reveals this fact in a more obvious manner. Although detectable signatures still exist in the case of deep tumors, the detection probability reduces significantly. Also, looking at the tumors with the same shape, as the size of the tumor decreases, the overall detectable signatures get weaker, as one would expect. One other conclusion is that, although the gland-shape tumor is larger than the other tumors considered here, the signature for this type of tumor is not better than those of other types with smaller size. This observation can be attributed to the similarity of the electrical properties of the glandular tissue to those of the tumor. The aforementioned results are all y expected; yet, what is not expected is the relatively large values of |End | for some scenarios. The reason for this phenomenon is the radiation nulls created in the setup in which the conductor dome is considered. To further illustrate, when using the conductor dome, the radiated electric field pattern shows nulls at several locations. By adding the tumor, due to scattering, some low levels of radiation appears in those locations. This creates a relatively large signature for those cases. Obviously, as the number of the receiving points increases, the detection probability will increase accordingly. Thus, in the next part, the results for the two of the aforementioned scenarios are studied when the number of receiving points are increased to 101 points. It should be mentioned that this is not a feasible study; however, it sheds light to few more dark corners of MBMD technique studied here. It is clear from Figures 3.27 and 3.28 that by increasing the number of receivers, the signatures become more pronounced; however, having such a high number of receivers can create implementation issues for a self-examination device. 15 10 With dome 5 Without dome 0 0 2 4 6 Summation Threshold 8 With dome 200 100 Without dome 100 200 300 Summation Threshold 0 0 15 10 Without dome 5 0 0 1 With dome 2 3 4 5 Summation Threshold Cumulative Ex Normalized Phase Difference 20 Cumulative Ey Normalized Phase Difference 25 Cumulative Ez Normalized Phase Difference Tumor Diameter=1cm Location1 Magnitude 300 Cumulative |Ez| Normalized Difference Cumulative |Ey| Normalized Difference Cumulative |Ex| Normalized Difference 110 8 Tumor Diameter=1cm Location1 Phase 6 4 2 0 0 With dome Without dome 0.5 1 1.5 Summation Threshold 2 4 3 2 With dome 1 0 0 Without dome 0.5 1 1.5 2 Summation Threshold 6 4 2 0 0 Without dome With dome 0.5 1 1.5 Summation Threshold 2 i i Figure 3.18. |End | and arg(End ) for the superficial spherical tumor with diameter of 10mm. The summation is over sixteen receiving points. 3.4 Experimental results4 Finally, the performance of the proposed breast cancer detection tool based on microwaves is studied through experiment. Single-frequency and multi-frequency detection strategies are investigated. The normal magnitude and phase differences of the scattering parameters 4 c 2013 IEEE. Adapted, with permission, from Modiri, A. and Kiasaleh, K., Experimental Results for a Novel Microwave Radiator Structure Targeting Non-invasive Breast Cancer Detection, will appear in proceedings of 2013 IEEE Global Humanitarian Technology Conference (GHTC), Oct./2013 for section 3.4 . 10 5 0 0 With dome Without dome 0.5 1 1.5 Summation Threshold 2 25 20 15 With dome 10 5 0 0 Without dome 5 10 15 Summation Threshold 15 10 Without dome 5 0 0 With dome 0.2 0.4 0.6 0.8 Summation Threshold 1 Cumulative Ex Normalized Phase Difference 15 Cumulative Ey Normalized Phase Difference 20 Tumor Diameter=1cm Location2 Magnitude Cumulative Ez Normalized Phase Difference Cumulative |Ez| Normalized Difference Cumulative |Ey| Normalized Difference Cumulative |Ex| Normalized Difference 111 6 Tumor Diameter=1cm Location2 Phase 4 2 0 0 Without dome With dome 0.5 1 1.5 2 Summation Threshold 4 3 2 With dome 1 0 0 Without dome 0.5 1 1.5 Summation Threshold 2 6 4 Without dome 2 0 0 With dome 0.5 1 1.5 Summation Threshold 2 i i Figure 3.19. |End | and arg(End ) for the deep spherical tumor with diameter of 10mm. The summation is over sixteen receiving points. between normal tissue and cancerous ones are used as preliminary indicators. It is shown that this practically implementable and yet relatively simple structure clearly highlights the existence of a deep tumor via a set of in-vivo non-invasive measurements. Earlier quite encouraging simulation results of using the proposed tool as a tumor detector were explained. Here, the measurement results of an implemented radiator structure are demonstrated which in fact, validate the previous simulation results. The implemented structure is improved in terms of antenna design as compared to the previous ones intro- 10 Without dome 5 0 0 With dome 2 4 6 Summation Threshold 8 20 15 10 With dome 5 0 0 Without dome 2 4 6 Summation Threshold 15 10 5 Without dome 0 0 With dome 0.5 1 1.5 2 Summation Threshold 2.5 Cumulative Ex Normalized Phase Difference 15 Cumulative Ey Normalized Phase Difference 20 Tumor Diameter=1cm Location3 Magnitude Cumulative Ez Normalized Phase Difference Cumulative |Ez| Normalized Difference Cumulative |Ey| Normalized Difference Cumulative |Ex| Normalized Difference 112 6 Tumor Diameter=1cm Location3 Phase 4 2 0 0 Without dome With dome 0.5 1 1.5 2 Summation Threshold 4 3 2 With dome 1 0 0 Without dome 0.5 1 1.5 2 Summation Threshold 6 4 2 0 0 Without dome With dome 0.5 1 1.5 Summation Threshold 2 i i Figure 3.20. |End | and arg(End ) for the deep spherical tumor with diameter of 10mm close to the chest wall. The summation is over sixteen receiving points. duced in (Modiri, et. al., 2013a). It is worthy of claim that the ultimate goal of this study is that the patients be able to use the device themselves with no need to any physical presence in a clinic. The device needs to be used periodically for routine checkups. Here, the experimental results of a set of preliminary tests of the implemented device are demonstrated. The implemented device was shown in Fig.1.11 of chapter 1. The hemisphere shaped base structure is created by 3D printing. For the sake of realistic experiment, breast 5 0 0 With dome Without dome 1 2 3 4 5 Summation Threshold With dome 150 100 50 0 0 Without dome 50 100 150 200 Summation Threshold 10 5 Without dome 0 0 With dome 0.5 1 Summation Threshold 1.5 Cumulative Ex Normalized Phase Difference 10 Cumulative Ey Normalized Phase Difference 15 Cumulative Ez Normalized Phase Difference 20 Tumor Diameter=5mm Location1 Magnitude 200 Cumulative |Ez| Normalized Difference Cumulative |Ey| Normalized Difference Cumulative |Ex| Normalized Difference 113 5 Tumor Diameter=5mm Location1 Phase 4 3 2 With dome 1 0 0 Without dome 0.5 1 1.5 2 Summation Threshold 4 3 2 With dome 1 0 0 Without dome 0.5 1 1.5 2 Summation Threshold 4 3 2 Without dome 1 0 0 With dome 0.5 1 1.5 2 Summation Threshold i i Figure 3.21. |End | and arg(End ) for the superficial spherical tumor with diameter of 5mm. The summation is over sixteen receiving points. tissue phantoms are built. The very useful recipes introduced in (Hahn, 2012) and (Ostadrahimi, 2009) are followed for this purpose. The measurement results are gathered as .S2P files from ENA and analyzed in MATLAB. 15 10 With dome 5 0 0 Without dome 1 2 Summation Threshold 3 25 20 15 10 With dome 5 0 0 Without dome 2 4 6 8 Summation Threshold 8 6 4 Without dome 2 0 0 With dome 0.2 0.4 0.6 0.8 Summation Threshold 1 Cumulative Ex Normalized Phase Difference 20 Cumulative Ey Normalized Phase Difference Tumor Diameter=1mm Location1 Magnitude Cumulative Ez Normalized Phase Difference 25 Cumulative |Ez| Normalized Difference Cumulative |Ey| Normalized Difference Cumulative |Ex| Normalized Difference 114 5 Tumor Diameter=1mm Location1 Phase 4 3 2 With dome 1 0 0 Without dome 0.5 1 1.5 2 Summation Threshold 4 3 With dome 2 1 0 0 Without dome 0.5 1 1.5 2 Summation Threshold 4 3 Without dome 2 1 0 0 With dome 0.5 1 1.5 Summation Threshold 2 i i Figure 3.22. |End | and arg(End ) for the superficial spherical tumor with diameter of 1mm. The summation is over sixteen receiving points. 3.4.1 Measurement setup After a number of trial attempts, reliable breast phantoms are created. Although there are some other phantom recipes in the literature, as well, the above mentioned ones are found to be more concise and precise. Fortunately, a donated tumor tissue from a breast cancer patient has become available to the author to use in phantom studies. The recipe authors have mentioned that the phantom doesn’t follow the behavior of the real breast tissue if it is not used in almost five days after being built. Thus, all the measurements summarized 20 Tumor Diameter=0.1cm Location2 Magnitude 15 10 5 0 0 With dome Without dome 1 2 3 Summation Threshold 4 30 20 10 0 0 With dome Without dome 5 10 Summation Threshold 15 15 10 5 0 0 Without dome With dome 0.5 1 Summation Threshold 1.5 Cumulative Ez Cumulative Ey Cumulative Ex Normalized Phase Difference Normalized Phase Difference Normalized Phase Difference Cumulative |Ez| Normalized Difference Cumulative |Ey| Normalized Difference Cumulative |Ex| Normalized Difference 115 6 Tumor Diameter=0.1cm Location2 Phase 4 With dome 2 0 0 Without dome 0.5 1 1.5 2 Summation Threshold 4 3 With dome 2 1 0 0 Without dome 0.5 1 1.5 2 Summation Threshold 8 6 Without dome 4 2 0 0 With dome 0.5 1 1.5 Summation Threshold 2 i i Figure 3.23. |End | and arg(End ) for the deep spherical tumor with diameter of 1mm. The summation is over sixteen receiving points. herein are performed within 48 hours of the completion of the creation of the phantoms. During the measurements, the phantoms rest on the bags of ice water. In our studies here, we show the results of using two breast phantoms; a normal breast phantom and a cancer breast phantom. The cancer breast phantom contains the donated tumor tissue. As it is shown in Figure 3.29, the tumor tissue is placed inside the glandular tissue. First, the fatty tissue solution is created and poured inside a container which is covered by aluminium sheet. Using aluminium sheet makes it easier to later separate the phantom from the container. A beaker filled with ice water is placed in the container while pouring the fatty tissue solution, so that it preserves the required space for glandular tissue. 4 Cumulative Ex Normalized Phase Difference 40 Cumulative Ey Normalized Phase Difference 60 Tumor Diameter=1cm Height=5mm Location1 Magnitude 1 Cumulative Ez Normalized Phase Difference Cumulative |Ex| Normalized Difference 30 Cumulative |Ey| Normalized Difference 116 20 10 With dome Without dome 0 0 40 1 2 3 Summation Threshold With dome Cumulative |Ez| Normalized Difference 20 0 0 Without dome 10 20 30 Summation Threshold 8 6 4 Without dome 2 0 0 With dome 0.2 0.4 0.6 0.8 Summation Threshold 3 Tumor Diameter=1cm Height=5mm Location1 Phase 2 1 0 0 Without dome With dome 0.2 0.4 0.6 Summation Threshold 0.8 3 2 With dome 1 Without dome 0 0 0.5 1 1.5 Summation Threshold 2 4 3 2 With dome 1 0 0 Without dome 0.5 1 1.5 Summation Threshold 2 i i Figure 3.24. |End | and arg(End ) for the superficial cylindrical tumor. The summation is over sixteen receiving points. After keeping the container inside the refrigerator for almost an hour, the fatty tissue becomes solid. Next, the ice water is replaced with hot water to make it easier to remove the beaker. Then, the glandular tissue solution is poured in the empty space left behind by removing the beaker. The tumor is added during this step. At the final step, the outer surface of the phantom is covered with skin solution. All tissue solutions are gelatine type materials and building the complete phantom requires careful temperature control. In order to remember which phantom contains the tumor, some dye powder is spread at one point of the surface 10 5 0 0 Without dome With dome 1 2 3 Summation Threshold 4 20 15 10 With dome 5 0 0 Without dome 2 4 6 Summation Threshold 8 6 4 Without dome 2 With dome 0 0 0.2 0.4 0.6 0.8 Summation Threshold 1 Cumulative Ex Normalized Phase Difference 15 Cumulative Ey Normalized Phase Difference Tumor Diameter=1cm Height=5mm Location2 Magnitude Cumulative Ez Normalized Phase Difference 20 Cumulative |Ez| Normalized Difference Cumulative |Ey| Normalized Difference Cumulative |Ex| Normalized Difference 117 4 Tumor Diameter=1cm Height=5mm Location2 Phase 3 2 With dome 1 0 0 Without dome 0.5 1 1.5 Summation Threshold 2 3 2 With dome 1 0 0 Without dome 0.5 1 1.5 Summation Threshold 2 4 3 2 Without dome 1 0 0 With dome 0.5 1 1.5 Summation Threshold i i Figure 3.25. |End | and arg(End ) for the deep cylindrical tumor. The summation is over sixteen receiving points. of the phantom which has the tumor inside (see Figure 3.29). The phantoms have a base diameter of 12 cm and height of 4.5 cm. The tumor tissue is a relatively thin piece, the thickness of which is not more than 3 mm; however, it has an irregular spread shape with the largest dimension of 1 cm. Measurement setup is shown in Figure 3.30. Flexible RF cables are used to connect the ENA arms to the SMA connectors on the radiator. It is noteworthy to mentions that while 5 0 0 With dome Without dome 0.5 1 1.5 2 Summation Threshold 40 30 With dome 20 10 0 0 Without dome 5 10 15 20 Summation Threshold 25 15 10 Without dome 5 0 0 With dome 1 2 3 4 Summation Threshold 5 Cumulative Ex Normalized Phase Difference 10 Cumulative Ey Normalized Phase Difference 15 Gland−shape Tumor Close to a Superficial Gland Magnitude Cumulative Ez Normalized Phase Difference Cumulative |Ez| Normalized Difference Cumulative |Ey| Normalized Difference Cumulative |Ex| Normalized Difference 118 4 Gland−shape Tumor Close to a Superficial Gland Phase 3 2 With dome 1 0 0 Without dome 0.5 1 1.5 Summation Threshold 2 4 3 2 With dome 1 0 0 Without dome 0.5 1 1.5 Summation Threshold 2 2 1.5 1 With dome 0.5 0 0 Without dome 0.2 0.4 0.6 Summation Threshold 0.8 i i Figure 3.26. |End | and arg(End ) for the superficial gland-shaped tumor. The summation is over sixteen receiving points. measuring the parameters for two ports, the other fourteen SMA connectors are left as open circuits in this experiment. 3.4.2 Results and analysis Equations (1.7)-(1.10) shown in chapter one show the measured data is analyzed. The main goal of the experiment is to find the signatures in the microwave radiation parameters due to the tumor. The signatures are detected as the result of comparing normal and cancerous tissues. Equations (1.7)-(1.10), in fact, show the cumulative normalized difference between 500 Tumor Diameter=1cm Location1 Magnitude 400 300 200 100 0 0 With dome Without dome 50 100 Summation Threshold 150 500 400 300 With dome 200 100 0 0 Without dome 100 200 300 Summation Threshold 100 50 With dome Without dome 0 0 5 10 Summation Threshold 15 Cumulative Ez Cumulative Ey Cumulative Ex Normalized Phase Difference Normalized Phase Difference Normalized Phase Difference Cumulative |Ez| Normalized Difference Cumulative |Ey| Normalized Difference Cumulative |Ex| Normalized Difference 119 50 Tumor Diameter=1cm Location1 Phase 40 30 Without dome 20 10 0 0 With dome 0.5 1 1.5 Summation Threshold 2 20 15 10 5 0 0 With dome Without dome 0.5 1 1.5 Summation Threshold 2 30 20 Without dome 10 With dome 0 0 0.5 1 1.5 Summation Threshold 2 i i Figure 3.27. |End | and arg(End ) for the superficial spherical tumor with diameter of 10mm. The summation is over 101 receiving points. normal and cancer tissues in terms of magnitude and phase for the reflection and transmittance coefficients at a single frequency. where the operator ’arg’ is the phase operator. The magnitudes of the S parameters are in dB and the phases are in radian. In order to remove the undesired spikes due to zero phase values, all the phases are summed with 2pi. The antennas are originally designed to resonate at 1.2 GHz; however, as expected, the experimental setup, the environmental factors, the radiator creation procedure and the presence of the phantom lead to a slightly different resonating frequency. Therefore, the analysis presented herein is performed in the new optimum resonance frequency. Figure 3.31 shows 200 With dome 100 0 0 Without dome 50 100 150 Summation Threshold Cumulative |Ez| Normalized Difference 100 50 0 0 With dome Without dome 10 20 30 40 Summation Threshold 80 60 40 20 0 0 With dome Without dome 5 10 Summation Threshold 15 Cumulative Ex Normalized Phase Difference 300 Cumulative Ey Normalized Phase Difference 150 Tumor Diameter=1cm Height=5mm Location1 Magnitude Cumulative Ez Normalized Phase Difference Cumulative |Ex| Normalized Difference 400 Cumulative |Ey| Normalized Difference 120 30 20 Tumor Diameter=1cm Height=5mm Location1 Phase Without dome 10 With dome 0 0 0.5 1 1.5 Summation Threshold 2 15 10 5 0 0 With dome Without dome 0.5 1 1.5 Summation Threshold 2 30 20 10 With dome Without dome 0 0 0.5 1 1.5 Summation Threshold 2 i i Figure 3.28. |End | and arg(End ) for the superficial cylindrical tumor. The summation is over 101 receiving points. the |S11 | for the sixteen antennas at the presence of the normal and cancer phantoms (two curves per antenna). It is observed that the optimum frequency is 1.354 GHz instead of 1.2 GHz. The effect of this deviation on the results is negligible since the resonance frequency is still inside the frequency range which is recommended for MBMD measurements. Figures 1.13 and 1.14, in chapter one, demonstrated the cumulative normalized difference between normal and cancer tissues in terms of magnitude for the reflection and transmittance coefficients in the measurements. Figures 3.32 and 3.33 show the phase results. Using ENA settings, it is possible to gather the data with minimum steps of 42.5 MHz; however, in order 121 Glandular Tissue Fatty Tissue First layer Ice Tumor Sign Fatty Tissue, Second layer Figure 3.29. The creation process of the two breast phantoms: one presenting normal tissue and the other one having a tumor tissue inside it. to show the significant advantage of multi-frequency measurement over a single-frequency one, the results for multi-frequency studies using frequency steps of 42.5, 85, and 127.5 MHz are shown in all four figures. The results for a single-frequency measurement at 1.354 GHz are shown, as well. In multi-frequency analysis, the cumulative differences achieved at all considered measurement frequencies are added up. It is evident that the tumor signatures are noticeable in all the figures. The phase signatures in Figures 3.32 and 3.33 keep increasing with frequency in a step-shape manner with almost equal jumps at each step. That can be explained by the intrinsic periodical behavior of the phase. 3.4.3 Comparing with a competing MBMD design In this section, the measurement results for the proposed device is compared with the measurements reported by another group working on MBMD in the Duke University (Stang, 122 2008). It is obvious that, in general, conducting a fair comparison of the studies conducted by two separate groups with distinct set of constraints is not easy; however, here, an attempt has been made to give the reader some idea of these two devices. It is critical to first mention the differences between two studies. They are: • The tumor considered in the Duke University study is a spherical one with the diameter of 9 mm placed inside a phantom with a base diameter of 9.5 cm and a height of 5.5 cm. In our study, however, the tumor was a real breast tumor and had a flat irregular shape with a maximum thickness of 3 mm and the largest dimension smaller than 1 cm. Overall, one can conclude that our study is using a smaller tumor, which is indeed more difficult to detect by palpation or by MBMD. In our study, the tumor is placed in a phantom with a base diameter of 12 cm and a height of 4.5cm. • Both tumors were placed in glandular tissue; however, in our study the tumor is placed in a deeper location (2 cm deeper). • Due to the antenna designs, this study is performed at 1.2GHz and the Duke University one is done at 2.7GHz. • The Duke University study targets a bulky clinical device while this study targets a self examination (small form factor) tool; therefore, they are more open to adding complexity. A key difference between the two design is the number of antennas; the Duke University device has 36 bowtie antennas, whereas this design has 16 bent dipole antennas (See Figure 3.34). As it is shown in Figure 3.35, the device introduced here shows signatures which are at least 15 dB stronger than those shown by the Duke University group when comparing the average values. It is also superior by at least 24 dB when comparing the peak values. 123 It is, hence, worthy of acclaim that, since the tumor scenario considered here presents a more challenging detection problem as compared to that offered by the Duke University study, the above results suggest that the proposed device provides much stronger signatures overall as compared to those offered by the Duke University platform when identical situations are encountered. 3.4.4 Calibration The user interface dome structure should be attached to the user’s body tightly (it will ultimately have a bra-shaped design), so that any unwanted effect from the operator misplacement is removed. However, to take care of inevitable errors caused by an inexperienced operator, the device needs to be calibrated before starting any rounds of measurements. Calibration will remove the impact of minute misplacement effects or similar environmental phenomena. As an example, Figure 3.36 shows the signatures created by a 7 mm movement of the tissue as compared to the signatures created by a deep tumor which is 5 mm in radius. The normalized scattering parameters are calculated (see 1.9). Figure 3.37 shows the HFSS models of the aforementioned movement and tumor. HFSS data is analyzed in Matlab to create the figures. This is a simple example of misplacement which is modeled in order to show how calibration works. In this study, first the transmittance magnitude signatures caused by tumor are extracted. Then, removing the tumor, the effect (signature) caused by a 7 mm movement (misplacement) is calculated. The signature created when these two phenomena happen simultaneously are examined as well. Finally, the answer to this question is investigated: Are we able to extract merely the tumor signature if we have had the chance to know the movement signature by means of calibration? Here, it is shown that the answer is ”Yes”. As it is clear in the figures, the tumor signatures are retrievable after removing the movement signatures. Figure 3.36 only focuses on the magnitude of the transmittance coefficient when antenna 1 is transmitting at 4 selected frequencies of 0.8GHz, 1.2GHz, 1.6GHz 124 and 2.4GHz. In fact, there is a large amount of data to analyze, since reflection coefficients, phase values, and the results for the cases when the other fifteen antennas are radiating are not shown here for the sake of conciseness. The presented figures, however, clearly show the principle idea. In order to calibrate the device for eliminating the movement effects, the transmittance coefficients between the adjacent antennas are used. The reason for choosing this technique is that, the amount of energy received by the antenna positioned right next to the transmitting antenna stays almost the same at the presence and absence of the tumor. Due to the reflection from body surface and creation of surface waves, the wave mainly travels around, and not through, the tissue in order to get to the adjacent antenna. Also, this adjacent-antenna energy transfer is affected by the entire tissue placement due to the change in the reflection coefficient. This fact has been shown in Figure 3.38 for the scenario shown in Figure 3.37. Consequently, adjacent-antenna transmittance coefficients are appropriate parameters for device calibration. The simulation observations validate this claim specially at or close to the resonance frequency of the antennas. 125 Figure 3.30. The measurement setup using an ENA. 126 Figure 3.31. The measured |S11 | of the sixteen antennas in two cases of having the normal phantom under the test and having the cancer phantom under the test. 2000 1800 1600 Star: Frequency step of 42.5MHz Circle: Frequency Step of 85MHz Cross: Frequency Step of 127.5MHz Square: Single frequency of 1.354GHz arg(S11nd) 1400 1200 1000 800 600 400 200 0 0 0.5 1 1.5 2 2.5 Frequency (up to which the results are added) (Hz) 3 9 x 10 11 Figure 3.32. Cumulative phase difference of reflection coefficient (arg(Snd )) for singlefrequency and multi-frequency studies. 127 3500 3000 Star: Frequency step of 42.5MHz Circle: Frequency Step of 85MHz Cross: Frequency Step of 127.5MHz Square: Single frequency of 1.354GHz arg(S21nd) 2500 2000 1500 1000 500 0 0 0.5 1 1.5 2 2.5 Frequency (up to which the results are added) (Hz) 3 9 x 10 21 Figure 3.33. Cumulative phase difference of transmittance coefficient (arg(Snd )) for singlefrequency and multi-frequency studies. 128 Figure 3.34. The MBMD design from Duke University and their phantom. 129 Figure 3.35. Comparing the phantom measurement results of this study with that of the Duke University. 130 Figure 3.36. Comparing signatures achieved from tumor with movement indications at four different frequencies. Figure 3.37. Digital breast phantom modeling misplacement. 131 Figure 3.38. Tumor signature and movement effect on adjacent antennas. CHAPTER 4 CONCLUSIONS AND FUTURE RESEARCH DIRECTIONS Microwave-base medical diagnosis is in its early stages of development and many unknowns and ambiguities reside in this field of research. The author has studied both hardware and software aspects of the problem and the preliminary model of a self examination breast monitoring tool was introduced and tested on phantoms. Yet, there is long way to having a clinically approved device. The following items show the possible future research studies in this field: - Since it is desirable to configure the final design in the form of a wearable device, an extensive study is required to develop textile radiator structures. This study needs to fully analyze and compare different applicable materials for embedding the radiator as well as various conformal antenna designs. It should also cover the impacts of human body on the device and suggest solutions to avoid damages (perspiration is an example). Simulation models are necessary to be developed for full characterization. Measurement trials will follow the simulations for validation. Since the device is desired to be a wearable one (similar to sport bras), it will need to be fabricated and tested in different sizes. - A comprehensive study is required on designing the measurement unit. The ultimate goal is to have a screening device that is easily readable not only by a physician, but also by the user herself. There have been new technologies which can be of significant utility in this regard. Copper Mountain Technologies, for instance, has recently introduced a miniature Vector Network Analyzer unit. Designing a Graphical User Interface (GUI) will be the next step. Throughout this part of study, although focusing on user-friendliness, an effort should be done to reduce the impact of the operator or environment on final results. 132 133 - Different types of radiation modalities, namely, continuous-wave, pulsed, and continuousmodulated wave, are needed to be compared both at system and performance levels. There is no such study in the literature at this times. Yet, the impact is predicted to be remarkable. - Another suggested study is the optimization of the radiation power and exposure duration with respect to the required detection accuracy. The results of such a study will have nontrivial impacts on the final configuration of the design of a wearable device. - One very important step is to investigate the normal changes of the breast tissue from time to time, in order to train the device not to detect them as abnormalities. The changes in the breast tissue throughout menstrual cycles are required to be well defined. Also, the changes due to losing/gaining weight should be taken into account. - Although it was shown that PSO, in its various versions, can successfully handle the desired problem, there are a handful of algorithms in the literature which can be used for this end; however, discovering which ones are the best options demands an extensive comparative study. This step can eventually make a huge impact on the detection speed and resource consumption. - Since the initial studies cannot involve human subjects, the other important part of this project is to study and develop body tissue phantoms that mimic human body close enough to support the primary experiments. Existing studies in the literature are based on simplified models. It is noteworthy that any imperfection in this part of the project is indeed an impediment to the reliability and precision of the final product. Chemistry and biology skills come into the picture for this section. - An extensive statistical study is required once a prototype is fabricated in order to collect feedback from groups of future users regarding possible improvements in the design. It is also important to study how women feel about such a technology and what they need to be able to operate the device comfortably. REFERENCES Abu-Al-Nadi, D. I. A., Ismail, T. H., Al-Tous, H., Mismar, M. J., (2012). Design of Linear Phased Array for Interference Suppression Using Array Polynomial Method and Particle Swarm Optimization. Wireless Personal Communications, 63(2), pp.501-513. Afshinmanesh, F., Marandi, A., Shahabadi, M. (2008). Design of a Single-Feed Dual-Band Dual-Polarized Printed Microstrip Antenna Using a Boolean Particle Swarm Optimization, WIEEE Transactions on Antennas and Propagation, 56(7), pp.1845-1852. Arunachalam, K., Udpa, L., and Udpa, S. S. (2008). A computational investigation of microwave breast imaging using deformable reflector, IEEE Trans. Biomedical Engineering, 55(2), pp.554-562. Balanis, C. A., (2008). Modern Antenna Handbook . Hoboken, NJ: Wiley-Interscience. Banks, A., Vincent, J., Anyakoha, C. (2007). A review of particle swarm optimization. Part I: background and development. Natural Computing, Volume 6, Issue 4, pp.467-484. Bayraktar, Z., Werner, P. L., Werner, D. H. (2006). The design of miniature three-element stochastic Yagi-Uda arrays using particle swarm optimization. IEEE Antennas and Wireless Propagation Letters, 5(1), 22-26. 134 135 Benedetti, M., Azaro, R., Franceschini, D., and Massa, A. (2006). PSO-based real-time control of planar uniform circular arrays. IEEE Antennas and Wireless Propagation Letters, 5(1), 545-548. Bhattacharya, R., Bhattacharyya, T.K., Garg, R. (2012). Position Mutated Hierarchical Particle Swarm Optimization and its Application in Synthesis of Unequally Spaced Antenna Arrays. IEEE Transactions on Antennas and Propagation, , vol.60, no.7, pp.3174-3181. Boeringer, D. W., and Werner, D. H. (2004). Particle swarm optimization versus genetic algorithms for phased array synthesis. IEEE Transactions on Antenna and Propagation, 52(3), 771-779. Cai, X., Cui, Y., Tan, Y. (2009). Predicted modified PSO with time-varying accelerator coefficients. International Journal of Bio-Inspired Computation. Camci, F. (2008). Analysis of Velocity Calculation Methods in Binary PSO on Maintenance Scheduling, the First International Conference on the Applications of Digital Information and Web Technologies, Ostrava, Czech Republic. Carlisle, A., and Dozler, G. (2002). Tracking Changing Extrema with Adaptive Particle Swarm Optimizer, the 5t h Biannual World Automation Congress, Orlando, USA. Chamaani, S., Abrishamian, M. S., and Mirtaheri, S. A. (2010). Time-domain design of UWB vivaldi antenna array using multiobjective particle swarm optimization. IEEE Antennas and Wireless Propagation Letters, 9, 666-669. 136 Chen W. N., Zhang, J., Chung, H. S. H., Zhong, W. L., Wu, W. G., and Shi Y. H. (2010, April). A novel set-based particle swarm optimization method for discrete optimization problems. IEEE Transactions on Evolutionary Computation, 14(2), 278-300. Cho, J., Yoon, J., Cho, S., Kwon, K., Lim, S., Kim, D., Lee, E. S., Kim, C. H., Choi, J. W., Cheon, C., and Kwon, Y. (2006). In-vivo measurements of the dielectric properties of breast carcinoma xenografted on nude mice. International Journal of Cancer, 119(3), 593-598. Clegg, J., Robinson, M.P. (2010). A genetic algorithm used to fit Debye functions to the dielectric properties of tissues. IEEE Congress on Evolutionary Computation (CEC), pp.1823. Clerc, M., and Kennedy, J. (2002). The particle swarm-Explosion, stability and convergence in a multidimensional complex space. IEEE Transactions on Evolutionary Computations, 6(1), 58-73. Cui, S., and Weile, D. (2005). Application of parallel particle swarm optimization scheme to the design of electromagnetic absorbers. IEEE Transactions on Antennas and Propagation, 53(11), 3616-3624. Das, S., Abraham, A., Konar, A. (2008). Particle Swarm Optimization and Diferential Evolution Algorithms: Technical Analysis, Applications and Hybridization Perspectives. SpringerVerlag Berlin Heidelberg. Delbary, F., Brignone, M., Bozza, G., Aramini, R., Piana, M. (2010). A Visualization Method for Breast Cancer Detection Using Microwaves. SIAM J. Appl. Math., 70(7), pp.2509-2533. 137 Delille, J. P., Slanetz, P. J., Yeh, E. D., Kopans, D. B., Garrido, E. (2002). Breast cancer: regional blood flow and blood volume measured with magnetic susceptibility-based MR imaging-initial results. Monthly Journal of Radiology by the Radiological Society of North America Inc., radiol.2232010428, 558-565. Eberhart, R., and Shi, Y. (2001). Particle Swarm Optimization: Developments, Applications, and Resources, the Congress on Evolutionary Computation, Seoul, Korea. Epitropakis, M. G., Plagianakos, V. P., Vrahatis, M. N. (2012). Evolving cognitive and social experience in Particle Swarm Optimization through Differential Evolution: A hybrid approach. SciVerse Information Sciences, Volume 216, pp.50-92. Fear, E.C., Li, X., Hagness, S.C. and Stuchly, M.A. (2002). Confocal microwave imaging for breast cancer detection: localization of tumors in three dimensions, IEEE Trans. Biomedical Engineering, vol.49, no.8, pp.812-822. Fear, E.C., Meaney, P.M., and Stuchly, M.A. (2003). Microwaves for breast cancer detection?, IEEE Potentials, 22(1), pp. 12-18. Gamagami, P., Silverstein, M.J., and Waisman, J.R. (1997). Infra-Red Imaging in Breast Cancer, the proceedings of the 19t h Annual International Conference of the IEEE Engineering in Medicine and Biology Society, Chicago, USA. Gandhi, K. R., Karnan, M., and Kannan, S. (2010). Classification Rule Construction Using Particle Swarm Optimization Algorithm for Breast Cancer Data Sets, the International Conference on Signal Acquisition and Processing, Singapore, Singapore. 138 Gao, H., and Xu, W. (2011). A new particle swarm algorithm and its globally convergent modifications. IEEE Transactions on Systems, Man, and Cybernetics, Part B: Cybernetics, 41(5), 1334-1351. Garca-Nieto, J., Alba, E. (2011). Restart particle swarm optimization with velocity modulation: a scalability test. Soft Computing, Volume 15, Issue 11, pp. 2221-2232. Genovesi, S., Monorchio, A., Mittra, R., and Manara, G. (2007). A Sub-boundary approach for enhanced particle swarm optimization and its application to the design of artificial magnetic conductors, IEEE Trans. Antennas and Propagation, 55(3), pp.766-770. Gies, D. and Rahmat-Samii, Y. (2003). Particle swarm optimization for reconfigurable phasedifferentiated array design. Microwave and Optical Technology Letters, 38(3), 172-175. Gilmore, C., and Lo Vetri, J. (2008). Enhancement of microwve tomography through the use of electrically conducting enclosures, IOP Science, Inverse Problems, 24(3) , pp.1-21. Golnabi, A.H., Meaney, P. M., Epstein, N. R., and Paulsen, K. D. (2011). Microwave Imaging for Breast Cancer Detection: Advances in Three- Dimensional Image Reconstruction, the Annual International Conference of the IEEE on Engineering in Medicine and Biology (EMBC), Boston, MA. Hagness, S. C., Taflove, A., and Bridges, J. E. (1998). Two-dimensional FDTD analysis of a pulsed microwave confocal system for breast cancer detection: fixed-focus and antenna-array sensors, IEEE Trans. Biomedical Engineering, 45(12), pp.1470-1479. 139 Hall P., and Hao, Y. (2006). Antennas and Propagation for Body-Centric Wireless Communications. Norwood, MA: Artec House, INC. Hand, J. H. (2008). Modeling the interaction of electromagnetic fields (10 MHz-10 GHz) with the human body: methods and applications. Physics in Medicine and Biology. Hassan, A., and El-Shenawee M. (2011). Review of electromagnetic techniques for breast cancer detection. IEEE Reviews in Biomedical Engineering, 4, 103-114. He, N., Xu, D., and Huang, L. (2009). The application of particle swarm optimization to passive and hybrid active power filter design. IEEE Transactions on Industrial Electronics, 56(8), 2841-2851. Ho, S. Y., Lin, H. S., Liauh, W. H., and Ho, S. J. (2008). OPSO: Orthogonal Particle Swarm Optimization and its application to task assignment problems. IEEE Transactions on Systems, Man and Cybernetics, Part A: Systems and Humans, 38(2), 288-298. Hoorfar, A. (2007). Evolutionary programming in electromagnetic optimization: A Review. IEEE Transactions on Antennas and Propagation, 55(3), 523-537. Hwang, K. C. (2009). Design and optimization of a broadband waveguide magic-T using a stepped conducting cone. IEEE Microwave and Wireless Components Letters, 19(9), 539-541. Haynes, M., Stang, J., and Moghaddam, M. (2012). Microwave breast imaging system prototype with integrated numerical characterization, Hindawi Publishing Corporation, Int. J. 140 Biomedical Imaging, 2012(706365). Hahn, C., and Noghanian, S. (2012). Heterogeneous breast phantom development for microwave imaging uisg regression models, Hindawi Publishing Corporation, Int. J. Biomedical Imaging. Hwang, K. C. (2009). Design and optimization of a broadband waveguide magic-T using a stepped conducting cone, IEEE Microwave and Wireless Components Letters, 19(9), pp.539541. Hoorfar, A. (2007). Evolutionary programming in electromagnetic optimization: A Review. IEEE Trans. on Antennas and Propagation, 55(3), pp.523-537. bibitemRc31a Ibrahim, I., Yusof, Z. M., Nawawi, S. W., Rahim, M. A. A., Khalil, K., Ahmad, H., Ibrahim, Z. (2012). A Novel Multi-state Particle Swarm Optimization for Discrete Combinatorial Optimization Problems. Fourth International Conference on Computational Intelligence, Modelling and Simulation (CIMSiM), pp.25-27. Imaginis, (2011). Breast Anatomy and Physiology; Accessible on http://www.imaginis.com/breasthealth-non-cancerous/breast-anatomy-and-physiology. Irishina, N., Moscoso, M., and Dorn, O. (2009). Microwave imaging for early breast cancer detection using a shape-based strategy, IEEE Trans. Biomedical Engineering, 56(4), pp.11431153. 141 Ismail, T. H., and Hamici, Z. M. (2010). Array pattern synthesis using digital phase control by quantized particle swarm optimization, IEEE Transactions on Antennas and Propagation, 58(6), 2142-2145. Janson, S., Middendorf, M. (2006). A hierarchical particle swarm optimizer for noisy and dynamic environments. Genetic Programming and Evolvable Machines, Volume 7, Issue 4, pp.329-354. Jensen, P. D., Rubak, T., Mohr, J. J., Zhurbenko, V. (2011). Nonlinear 3-D microwave imaging for breast-cancer screening: Log, phase and log-phase formulation. Antennas and Propagation Conference (LAPC), pp.14-15. Jin, N., and Rahmat-Samii, Y. (2007). Advances in particle swarm optimization for antenna designs: real-number, binary, single-objective and multiobjective implementations, IEEE Transactions on Antennas and Propagation, 55(3), 556-567. Jin, N., and Rahmat-Samii, Y. (2005). Parallel particle swarm optimization and finite difference time-domain (PSO/FDTD) algorithm for multiband and wide-band patch antenna designs. IEEE Transactions on Antennas and Propagation, 53(11), 3459-3468. Jin, N., and Rahmat-Samii, Y. (2008). Particle swarm optimization for multi-band handset antenna designs: A hybrid real-binary implementation, IEEE International Symposium of Antennas and Propagation, Taipei, Taiwan. 142 Jin, N., and Rahmat-Samii, Y. (2010). Jin, Nanbo, and Yahya Rahmat-Samii. ”Hybrid Real-Binary Particle Swarm Optimization (HPSO) in Engineering Electromagnetics. IEEE Transactions on Antennas and Propagation. Kennedy J. and Eberhart, R. (1995). Particle Swarm Optimization. Paper presented at IEEE International Conference of Neural Networks, Piscataway, NJ. Khan, U.A., Al-Moayed, N., Nguyen, N., Korolev, K.A., Afsar, M.N., and Naber, S.P. (2007). Broadband dielectric characterization of tumorous and nontumorous breast tissues. IEEE Transactions on Microwave Theory and Techniques, 55(12), 2887-2893. Klemm, M., Leendertz, J.A., Gibbins, D., Craddock, I.J., Preece, A., and Benjamin, R. (2010). Microwave radar-based differential breast cancer imaging: imaging in homogeneous breast phantoms and low contrast scenarios, IEEE Trans. Antennas and Propagation, 58(7), pp.2337-2344. Ko, C. N., Chang, Y. P., and Wu, C. J. (2009). A PSO method with nonlinear time-varying evolution for optimal design of harmonic filters. IEEE Transactions on Power Systems, 24(1), 437-444. Li, D., Meaney, P.M., Paulsen, K.D. (2003). Conformal microwave imaging for breast cancer detection, IEEE Trans. Microwave Theory and Techniques, 51(4), pp.1179-1186. Liu, H., Abraham, A. (2009). Chaos and Swarm Intelligence. r-Verlag Berlin Heidelberg, Intel. Computing Based on Chaos, SCI 184, pp. 197-212. 143 Liu, W. (2005). Design of multiband CPW-fed monopole antenna using a particle swarm optimization approach. IEEE Transactions on Antennas and Propagation, 53(10), 3273-3279. Lazebnik, M., Changfang, Z., Palmer, G.M., Harter, J., Sewall, S., Ramanujam, N., and Hagness, S.C. (2008). Electromagnetic spectroscopy of normal breast tissue specimens obtained from reduction surgeries: Comparison of optical and microwave properties. IEEE Transactions on Biomedical Engineering,55(10), 2444-2451. Lazebnik, M., Popovic, D., McCartney, L., Watkins, C. B., Lindstorm, M. J., Harter, J., Sewall, S., Ogilvie, T., Magliocco, A., Breslin, T. M., Temple, W., Mew, D., Booske, J. H., Okoniewski, M., and Hagness, S. C. (2007). A large-scale study of the ultrawideband microwave dielectric properties of normal, benign and malignant breast tissues obtained from cancer surgeries. Physics in Medicine and Biology Journal, IOP publishing Ltd., 52(20), 6093-6115. Lazebnik, M., Okoniewski, M., Booke, J. H., and Hagness S. C. (2007). Highly accurate debye models for normal and malignant breast tissue dielectric properties at microwave frequencies, IEEE Microwave and Wireless Components Letters, 17(12). Luan, F., Choi, J. H., and Jung, H. K. (2012). A Particle Swarm Optimization algorithm with novel expected fitness evaluation for robust optimization problems. IEEE Transactions on Magnetics, 48(2), 331-334. Mahmoud, K. R., El-Adawy, M. I., Ibrahem, S. M. M., Bansal R., Zainud-Deen, S. H. (2008). MPSO-MOM: A Hybrid Modified Particle Swarm Optimization and Method of Moment Algo- 144 rithm for Smart Antenna Synthesis, Electromagnetics, Volume 28, Issue 6, 2008, pp. 411-426. Matekovits, L., Mussetta, M., Pirinoli, P., Selleri, S., and Zich, R. (2005). Improved PSO Algorithms for Electromagnetic Optimization, IEEE Antennas and Propagation Society International Symposium, Washington, DC. Menhas, M. I., Fei, M., Wang, L., Fu, X. (2011). A Novel Hybrid Binary PSO Algorithm. Advances in Swarm Intelligence Lecture Notes in Computer Science, Volume 6728, pp. 93100. Mikki S., and Kishk, A. (2005). Investigation of the Quantum Particle Swarm Optimization Technique for Electromagnetic Applications, at IEEE Antennas and Propagation Society International Symposium, Washington, DC. Modiri, A. and Kiasaleh, K. (2010). Efficient Design of Microstrip Antennas Using Modified PSO Algorithm,the 14t h Biennial IEEE Conference on Electromagnetic Field Computation (CEFC), Chicago, IL. Modiri, A. and Kiasaleh, K. (2011a). Efficient design of microstrip antennas for SDR applications using modified PSO algorithm. IEEE Transactions on Magnetics, 47(5), 1278-1281. Modiri, A. and Kiasaleh, K. (2011b). Modification of real-number and binary PSO algorithms for accelerated convergence. IEEE Transactions on Antennas and Propagation, 59(1), 214224. 145 Modiri, A. and Kiasaleh, K. (2011c). Permittivity Estimation for Breast Cancer Detection Using Particle Swarm Optimization Algorithm, the 33rd Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), Boston, MA. Modiri, A. and Kiasaleh, K. (2011d). Real Time Reconfiguration of Wearable Antennas, IEEE Topical Conference on Biomedical Wireless Technologies, Networks and Sensing Systems, Phoenix, AZ, USA. Modiri, A. and Kiasaleh, K. (2012). A Novel Discrete Particle Swarm Optimization Algorithm for Estimating Dielectric Constants of Tissue, 34t h Annual International Conference of the IEEE Engineering in Medicine and Biology Society ,EMBC2012, San Diego. Modiri, A., Kiasaleh, K., and Chandrahas, S.(2013a). Characterizing a Proposed SixteenElement Array Antenna Designed for Microwave Imaging of Breast CancerNovel Microwave Radiator Structure for Non-invasive Breast Cancer Detection, IEEE Topical Conf. Wireless Sensors and Sensor Networks (WiSNet), pp.61-63. Modiri, A., Kiasaleh, K. (2013b). Experimental Results for a Novel Microwave Radiator Structure Targeting Non-invasive Breast Cancer Detection. Accepted to be presented in IEEE Global Humanitarian Technology Conference, GHTC 2013. Modiri, A., Kiasaleh, K. (2013c). Non-invasive Microwave Breast Cancer Detection - A Comparison Study. Accepted to be presented in IEEE Global Humanitarian Technology Conference, GHTC 2013. 146 Mojabi, P., LoVetri, J. (2010). Eigenfunction contrast source inversion for circular metallic enclosures, Inverse Problems, 26(025011), pp. 1-23. Montes de Oca, M.A., Stutzle, T., Birattari, M., and Dorigo, M. (2009). Frankenstein’s PSO: A composite Particle Swarm Optimization algorithm. IEEE Transactions on Evolutionary Computation, 13(5), 1120-1132. Nakano, S., Ishigame, A., and Yasuda, K. (2007). Particle Swarm Optimization Based on the Concept of Tabu Search,IEEE Congress on Evolutionary Computation (CEC), Singapore, Singapore. Nieuwstadt, L. M.C. (2011). Microwave Measurement System for Breast Cancer Imaging: An Experimental Prototype Towards Time-Domain Inverse Scattering; PhD Thesis, Department of Electrical Engineering, University of Michigan. Nikolova, N. K. (2011). Microwave imaging for breast cancer. IEEE Microwave Magazine, 12(7), 78-94. O’Halloran, M., Conceicao, R., Byrne, D., Glavin M., Jones, E. (2009). FDTD MODELING OF THE BREAST: A REVIEW. Progress In Electromagnetics Research B, Vol. 18, pp.1-24. Orfanidis, S. J. (2008). Electromagnetic waves and antennas. eBook, Retrieved January 3, 2012, from http://www.ece.rutgers.edu/ orfanidi/ewa/. 147 Ostadrahimi, M., Reopelle, R., Noghanian, S., Pistorius, S., Vahedi, A., and Safari, F. (2009). A heterogeneous breast phantom for microwave breast imaging, in Proc. the 31st Annual International Conf. IEEE Eng. in Medicine and Biology Society: Engineering the Future of Biomedicine, Minneapolis, Minn, USA, pp.2727-2730. Paulsen, K. D., Meaney, P. M., and Gilman, L. C.(2005). Alternative Breast Imaging, Four Model-Based Approaches, USA: Springer Science + Business Media Inc. Poli, R., (2009). Mean and variance of the sampling distribution of particle swarm optimizers during stagnation. IEEE Transactions on Evolutionary Computation, 13(4), 712-721. Ratnaweera, A., Halgamuge, S. K., and Watson, H. C. (2004). Self-organizing hierarchical particle swarm optimizer with time-varying acceleration coefficients. IEEE Transactions on Evolutionary Computation, 8(3), 240- 255. Robinson, J., and Rahmat-Samii, Y. (2004). Particle swarm optimization in electromagnetic. IEEE Transactions on Antennas and Propagation, 52(2), 397-407. Rubaek, T., Kim, O.S., and Meincke, P. (2009). Computational validation of a 3-D microwave imaging system for breast-cancer screening, IEEE Trans. Antennas and Propagation, 57(7), pp.2105-2115. Rubk, T., Meaney, P.M., Meincke, P., and Paulsen, K.D. (2007). Nonlinear microwave imaging for breast-cancer screening using gauss-newton’s method and the CGLS inversion 148 algorithm, IEEE Trans. Antennas and Propagation, 55(8), pp.2320-2331. Sabouni, A., Noghanian, S., and Pistorius, S.(2010). A global optimization technique for microwave imaging of the inhomogeneous and dispersive breast. Canadian Journal of Electrical and Computer Engineering,35(1), pp.15-24. Sedighizadeh, D., Masehian, E. (2009). Particle Swarm Optimization Methods, Taxonomy and Applications. Particle Swarm Optimization Methods, Taxonomy and Applications. Sharaf, A. M., and El-Gammal, A. A. A. (2009). A Discrete Particle Swarm Optimization Technique (DPSO) for Power Filter Design, 4th International Design and Test Workshop (IDT). Shi, Y., and Eberhart, R. (1999). Empirical Study of Particle Swarm Optimization, Congress on Evolutionary Computation, Washington, DC. Sill, J. M., and Fear, E. C. (2005). Tissue sensing adaptive radar for breast cancer detection: study of immersion liquids, Electronics Letters , 41(3), pp.113-115. Stang, J. P., (2008). A 3D Active Microwave imaging system for breast cancer screening, Ph.D. dissertation, Dept. Elect. and Computer Eng., Duke University, Durham, NC. Sun, J., Fang, W., and Xu, W. (2010). A quantum-behaved particle swarm optimization with diversity-guided mutation for the design of two-dimensional IIR digital filters. IEEE 149 Transactions on Circuits and Systems II: Express Briefs, 57(2), 141-145. Tao, Q., Chang, H. Y., Yi, Y., Gu, C. Q., and Li, W. J. (2010). An Analysis for Particle Trajectories of a Discrete Particle Swarm Optimization. 3r d IEEE International Conference on Computer Science and Information Technology (ICCSIT). Thakur, K. P., and Chan, K., Holmes, W. S., and Carter, G. (2002). An Inverse Technique to Evaluate Thickness and Permittivity Using Reflection of Plane Wave from Inhomogeneous Dielectrics, 59t h Automatic RF Techniques Group (ARFTG) Conference, Seattle, WA. Vural, R. A., Yildirim, T., Kadioglu, T., and Basargan, A. (2012). Performance evaluation of evolutionary algorithms for optimal filter design. IEEE Transactions on Evolutionary Computation, 16(1), 135-147. Wang, W., Lu, Y., Fu, J. S., and Xiong, Y. Z. (2005). Particle swarm optimization and finite-element based approach for microwave filter design. IEEE Transactions on Magnetics, 41(5), 1800- 1803. Winters, D.W.; Shea, J.D.; Kosmas, P.; Van Veen, B.D.; Hagness, S.C. (2005). ThreeDimensional Microwave Breast Imaging: Dispersive Dielectric Properties Estimation Using Patient-Specific Basis Functions. IEEE Transactions on Medical Imaging, vol.28, no.7, pp.969-981. Woten, D. A., and El-Shenawee, M. (2008). Broadband dual linear polarized antenna for statistical detection of breast cancer, IEEE Trans. Antennas and Propagation, 56(11), pp.3576- 150 3580. Xu, W. (2009). A hybrid particle swarm optimization approach with prior crossover differential evolution. Proceedings of the first ACM/SIGEVO Summit on Genetic and Evolutionary Computation - GEC. Yeung, C. W., Leung, F. H., Chan, K. Y., and Ling, S. H. (2009). An Integrated Approach of Particle Swarm Optimization and Support Vector Machine for Gene Signature Selection and Cancer Prediction, The International Joint Conference on Neural Network, Atlanta, Georgia. Zhao, X., Wenbin, W. (2010). An AntiCentroid-oriented Particle Swarm Algorithm for Numerical Optimization, Artificial Intelligence and Computational Intelligence Lecture Notes in Computer Science Volume 6320, pp. 302-309. Zhou, T., Meaney, P. M., Pallone, M. J., Geimer, S., Paulsen, K. D.(2010). Microwave tomographic imaging for osteoporosis screening: a pilot clinical study, Annual Int. Conf. on Engineering in Medicine and Biology Society (EMBC), pp.1218-1221. Zhurbenko, V. (2011). Challenges in the Design of Microwave Imaging Systems for Breast Cancer Detection. Advances in Electrical and Computer Engineering Date of publication, Volume 11, Issue 1, pp.91 - 96. Zhurbenko, V., Rubk, T., Krozer, V., and Meincke, P. (2010). Design and realisation of a microwave three-dimensional imaging system with application to breast-cancer detection, IET Microwaves, Antennas and Propagation, 4(12), pp.2200-2211. VITA Arezoo Modiri received her BS degree in 2001 from University of Tehran, Iran, and her MS degree in 2005 from Iran University of Science and Technology, both in Electrical Engineering. She entered the Electrical Engineering doctoral program at The University of Texas at Dallas in 2009. During her PhD which was focused in RF and Microwave, she had the opportunity to experience part time internships in The University of Texas Southwestern Medical Center, as well as AWR Corporation. Her research areas of interest are microwave diagnostic techniques, antennas, evolutionary algorithms, RF and microwave engineering and electromagnetics. Publications Journals: • A. Modiri and K. Kiasaleh, ”Radiation and Scattering Analysis of a Proposed Microwave Radiator Targeting Tissue Malignancy Detection, Submitted to IEEE transactions on microwave theory and techniques”, August 2013. • D. Cao, A. Modiri, G. Sureka, and K. Kiasaleh, ”DSP Implementation of the Particle Swarm and Genetic Algorithms for Real-Time Design of Thinned Array Antennas,” IEEE Antennas and Wireless Propagation Letters, vol.11, no., pp.1170-1173, 2012. • A. Modiri and K. Kiasaleh, ”Efficient Design of Microstrip Antennas for SDR Applications Using Modified PSO Algorithm”, IEEE Transactions on Magnetics, vol.47, no.5, pp.1278-1281, May 2011. • A. Modiri and K. Kiasaleh, ”Modification of Real-Number and Binary PSO Algorithms for Accelerated Convergence,” Antennas and Propagation, IEEE Transactions on , vol.59, no.1, pp.214-224, Jan. 2011. Conferences: • A. Modiri and K. Kiasaleh, ”Non-invasive Microwave Breast Cancer Detection - A Comparison Study”; Accepted to be presented in IEEE Global Humanitarian Technology Conference, GHTC 2013. • A. Modiri and K. Kiasaleh, ”Experimental Results for a Novel Microwave Radiator Structure Targeting Non-invasive Breast Cancer Detection”; Accepted to be presented in IEEE Global Humanitarian Technology Conference, GHTC 2013. • A. Modiri, K. Kiasaleh and S. Chandrahas, ”Characterizing a Proposed SixteenElement Array Antenna Designed for Microwave Imaging of Breast Cancer”, IEEE Topical Conference on Biomedical Wireless Technologies, Networks and Sensing Systems, BioWireleSS 2013, RWW 2013, Austin, TX, Jan. 20-23, 2013. • J. Singh, A. Modiri, and K. Kiasaleh, ”Novel UWB Hybrid Dipole Antenna with QuasiIsotropic Radiation Pattern”, IEEE Radio and Wireless Symposium, RWS 2013, RWW 2013, Austin, TX, Jan. 20-23, 2013. • A. Modiri and K. Kiasaleh, ”A Novel Discrete Particle Swarm Optimization Algorithm for Estimating Dielectric Constants of Tissue”, 34th Annual International Conference of the IEEE Engineering in Medicine and Biology Society ,EMBC2012, San Diego, 28 Aug.-1 Sep. 2012. • A. Modiri, K. Kiasaleh, ”Improvements on Active Microwave Radiometry for Breast Cancer Detection”, Metroplex Day, Annual Conference on Research Opportunities, Arlington, TX, 3 Feb. 2012. • A. Modiri and K. Kiasaleh, ”Permittivity Estimation for Breast Cancer Detection Using Particle Swarm Optimization Algorithm”, 33rd Annual International Conference of the IEEE Engineering in Medicine and Biology Society ,EMBC2011, Boston, 31 Aug.-3 Sep. 2011. • A. Modiri and K. Kiasaleh, ”Real Time Reconfiguration of Wearable Antennas”, IEEE Topical Conference on Biomedical Wireless Technologies, Networks and Sensing Systems, BioWireleSS 2011, RWW 2011, Phoenix, AZ, 16 - 20 Jan. 2011. • A. Modiri and K. Kiasaleh,”Efficient Design of Microstrip Antennas for SDR Applications Using Modified PSO Algorithm”, 14th Biennial IEEE Conference on Electromagnetic Field Computation (CEFC), Chicago, IL, 9-12 May 2010. Book Chapter: • Book Title: ”Swarm Intelligence for Electric and Electronic Engineering” Chapter Title: ”Particle Swarm Optimization Algorithm in Electromagnetics- Case Studies: Reconfigurable Radiators and Cancer Detection” Chapter Authors: A. Modiri, K.Kiasaleh Book Editors: Girolamo Fornarelli, Luciano Mescia published late in December 2012 in the United States of America by Engineering Science Reference (an imprint of IGI Global). Pending Patent: • K. Kiasaleh; A. Modiri; Systems And Methods For Detecting Breast Cancer, filed with the USPTO on 3/15/2013 and has been given the serial number 61/790,988. Technology Disclosure: • J. Singh; A. Modiri; K. Kiasaleh; Hybrid Dipole Antenna, Tech ID:13-022.

1/--страниц