A Novel UWB Imagng System Setup for Computer- Aded Breast Cancer Dagnoss Xang He, Ja L, Chenxng Wu Electrcal and Computer Engneerng Oakland Unversty, OU Rochester, I 48309, U.S.A xhe2@oakland.edu, l4@oakland.edu, cwu23@oakland.edu Abstract crowave magng usng UWB radar has been proved as a promsng technque for breast cancer dagnoss. However, most of the researches n ths doman are based on smulated envronment, due to the fact of complex system setup. Few groups have bult up an actual UWB magng system for real lfe experment. In ths paper we present a novel UWB magng system setup for breast cancer dagnoss. The system conssts of one horn antenna as the transmtter, and a 4-element lnear planar antenna array as the recever to collect the backscattered sgnal. Based on ths system setup, we propose a data-drven approach usng lnear classfer to detect tumor presence and further explore tumor characterstcs, wthout constructng the UWB mage. Ths approach bypasses the complcated mage reconstructon algorthm and ll-posed nverse problem. Wth the low complexty system setup, our data-drven approach gves promsng results n the breast cancer dagnostc experment. The experment shows that ths system setup can be adopted as a prototype for research project n computer-aded dagnoss of breast cancer. Keywords UWB magng system, breast cancer, computeraded dagnoss, lnear classfer I. INTRODUCTION In recent years, many researchers have nvestgated the method of employng UWB radar to detect breast cancer. Successful works have been proposed usng numercal breast model n FDTD smulatons [1-2]. Expermental works usng breast phantom have also been reported n [3-6]. The key to a successful detecton s based on the fact that the breast tumors have delectrc propertes that are greatly dfferent from those of healthy breast tssues. The UWB sgnals are transmtted nto the breast, and the backscattered sgnals are recorded and processed to fnd out the tumor s presence. However, the UWB mcrowave magng technque does not provde hgh enough resoluton to reconstruct the fne spatal features of a detected leson n the breast. Confocal crowave Imagng (CI) technque [1] apples delay-and-sum (DAS) beamformng to reconstruct the mage. Ths technque has the advantage of smplcty n mage reconstructon process, but t offers lmted performance n terms of mage resoluton and clutter rejecton. crowave Imagng va Space-Tme (IST) algorthm [4] has addressed these lmtatons and outperformed DAS usng data-ndependent algorthm. Other technques nclude Tssue Sensng Adaptve Radar (TSAR) magng [5], and tme reversal method n mcrowave magng. For all these methods, the general dea s to process the backscattered sgnal through a specfc mage reconstructon algorthm to obtan knowledge of the breast tumor. UWB mage reconstructon s an nverse problem whch s ll-posed and challengng due to the nonlnear nature of the problem tself. Thus t requres complex system setup and sophstcated algorthm to solve such problem. Alternatvely, we could bypass the mage reconstructon process and use data-drven approach to obtan nformaton about the breast tumor drectly,.e. nformaton regardng the presence of the tumor. Sardar and shra [7] proposed a method called Applcaton Specfc Instrument (ASIN) framework for the dagnoss of breast cancer. They appled radal bass functon based neural network to perform a pattern recognton task to determne whether the tumor s present or not. If the answer s postve, they further estmate the sze of the tumor through regresson machne, agan, wth the use of RBF based neural network. But they ddn t gve a complete soluton, as ther results were only based on smulated data and not valdated wth real lfe expermental data. Alshehr et al [8] bult a UWB magng system to detect tumor n breast phantom. The system conssts of commercal UWB transcevers and neural network based pattern recognton software. However, such commercal UWB transcever s hard to confgure to adapt to varous expermental envronments, thus not practcal for laboratory use. oreover, f we could locate the proper features related to the target sgnature tumor characterstcs, we do not need a sophstcated machne learnng tool neural network to solve the breast cancer dagnoss problem. A lnear classfer s capable for the work. We wll demonstrate ths dea n our experment n secton IV. Once the presence of tumor s determned, we could gather more nformaton related to the tumor s characterstcs. Davs et al [9] presented a data-drven approach for breast tumor characterzaton based on UWB backscattered sgnals. Lnear classfer s appled to dstngush tumors wth dfferent shape and sze. Ther results are encouragng but only tested n smulaton. We desgn an experment to explore the tumor characterstcs wth our system setup. The prelmnary result verfes the feasblty of such data-drven approach. In ths paper we frst propose a novel UWB magng system setup for breast cancer dagnoss. The transmtter conssts of a trgger, a pulse generator and a sngle horn antenna. On the recever sde, the backscattered sgnals are collected by a lnear planar antenna array, sampled by data acquston unt wth hgh samplng rate (40 GHz), and post-processed n a computer for dagnostc purpose. Ths system setup s smple and cost 978-1-4799-5396-7/14/$31.00 2014 IEEE 260
effectve. It could be confgured quckly to adapt to any UWB radar based breast cancer research project or served as a prototype of UWB magng system for computer-aded breast cancer dagnoss. Based on our system setup, we nvestgate how the datadrven approach could be adapted n breast cancer dagnostc experment usng homogeneous breast phantom. We descrbe a lnear classfer based scheme n breast cancer dagnoss. The dagnoss operates n two stages. In the frst stage, t detects whether a tumor s present or not nsde the breast phantom. If t s determned that the tumor s present, the second stage s performed to explore the tumor characterstcs. Ths paper s organzed as follows. The system setup s presented n the next secton, followed by the dagnoss procedure n secton III. The expermental results are presented n secton IV. The concluson and future work dscusson are gven n the last secton. II. UWB IAGING SYSTE SETUP The proposed UWB magng system setup conssts of three man parts: breast tssue and tumor model, UWB transcever and computer aded dagnostc program. We descrbe the frst two parts n ths secton and the thrd part n secton III. A. Breast Tssue and Tumor Phantom Constructon In the study of breast phantom, the most mportant metrcs s the delectrc dfference between the normal breast tssues and the malgnant tumor. Prevous studes n delectrc spectroscopy suggest that the delectrc propertes contrast between malgnant and normal breast tssue s greater than 2:1 n the mcrowave frequency range [4]. In our expermental setup, soybean ol wth delectrc constant ε r = 3.0 s chosen as the normal breast tssue smulant due to ts smlarty n delectrc propertes to low-water-content fatty tssue [4]. The soybean ol s contaned n a tube made of acyclc (ε r =1.9 ), whch smulates the skn layer. Alumnum block wrapped by alumnum fol (ε r = 9.3 ) s hanged nsde the ol to represent the tumor. B. UWB Transcever Desgn The UWB antenna used for transmttng the pulse s a large double-rdged horn antenna, as shown n Fgure 1 and 2. It has low return loss n the frequency range of 0.8 to 18 GHz and has an average gan of 12 db whch s hgh enough to guarantee the sgnal to be able to penetrate the breast model [10]. Fgure 2: Double-edge horn antenna mechancal drawng [10] At the recever sde, we desgn a checker-shape planar antenna as shown n Fgure 3 and 4. It s small enough to be easly attached to the surface of the breast model. The antenna s return loss s under -10dB n the frequency range of 3 to 15 GHz, whch meets the expermental requrement [11]. Fgure 3: Checker-shape planar antenna [11] Fgure 4: Checker-shape planar antenna mechancal drawng [11] A 4-element lnear array based on the check-shape planar antenna s developed to capture the backscattered sgnals, as shown n Fgure 5. Ths lnear array setup has smple arrangement, and offers better temporal-spatal resoluton than a sngle antenna. Fgure 5: Lnear antenna array A pulse wth 70 ps wdth s created usng the PSPL 3600 pulse generator, as shown n Fgure 6. Ths UWB pulse s used to excte the breast model. Fgure 1: Double-edge horn antenna [10] Fgure 6: Pulse for exctaton 261
The backscattered sgnals collected by the recever antenna array are acqured by a 4-channel hgh samplng rate osclloscope-tektronx TDS6154C, as shown n Fgure 7. physcal characterstcs, ncludng shape and sze. The general procedure s llustrated n Fgure 10. Fgure 7: The backscattered sgnals dsplay on the screen Overall, the UWB transcever setup and block dagram are shown n Fgure 8 and 9, respectvely. It conssts of the followng: Trgger: Wavetek, Arbtrary Waveform Generator Pulse generator: PSPL 3600, wth 70 ps duraton Transmtter: sngle bg horn antenna Recever: 4 element checkered-shaped antenna array Data acquston: Aglent Osclloscope, wth 40 GSa/s Fgure 8: UWB transcever setup Fgure 9: The UWB transcever block dagram III. DIAGNOSTIC PROCEDURE The computer-aded dagnoss of the breast cancer operates n two stages. In the frst stage, a bnary classfer s constructed to detect whether the tumor s present or absent wthn the breast tssue. Once the answer s postve, the second stage s ntalzed to analyze the tumor s shape and sze. ult-way classfers are developed to classfy the tumor wth dfferent Fgure 10: Computer-aded dagnoss procedure The dagnoss s done through dentfyng whch class a backscattered sgnal belongs to base on the observaton data. To acheve ths goal, we need to extract features from the data that are good dscrmnants among the C classes, where C s the total number of class. In our experment, C equals 2 n the frst stage, equals 3 n the second stage. The feature extracton { X, y } =, where s based on a set of labeled tranng data 1 th X s the nput vector contanng the measured data for the target realzaton and y {1, 2,..., C} specfes whch class th the target belongs to. The classfer s desgned as a three-step procedure as follows: Frst, we construct the tranng set. The tme doman samples collected by the lnear antenna array served as the nput vector X n the tranng set. The target value y s defned as the label of classes. Second, we perform prncple component analyss (PCA) to reduce the feature space dmenson, and select the domnant features for the classfcaton purpose. Ths s acheved by applyng sngular value decomposton (SVD) to project the sgnals to low dmensons. Specfcally, f we have N samples, each sample s a tme doman receved sgnals wth N entres. We form a matrx, carry out the SVD, N T A R A= U Σ V, and truncate t to dmenson of 2. Then U Σ wth k = 2 are 2D vectors wth N the components k k data entres. These 2D ponts can be plot on a plane to vsualze the data. Thrd, once the good features are selected, we could buld a lnear classfer to fulfll the classfcaton work. The commonly used lnear classfer, such as K nearest neghbor (KNN), lnear dscrmnant analyss (LDA) or Naïve Bayes could be mplemented n our classfer and they all acheve good classfcaton result. IV. EXPERIENTAL RESULTS In ths secton, we present the expermental results of the two stages breast cancer dagnoss. 262
A. Tumor Detecton The frst stage s to detect whether or not a tumor s present wthn the breast tssue. Ths s a bnary classfcaton problem. The breast phantoms wth and wthout tumor are both llumnated by the transmtter. Backscattered sgnals are collected by the recever antenna array. As the dstances between the antenna and the breast phantom are known beforehand, we are able to do a smple tme gatng to preserve the expected tumor response. As shown n Fgure 11, for a sngle backscattered sgnal, the tumor response s located at the secondary peak of the entre tme samples. Fgure 11: Tumor response n a sngle backscattered sgnal The total response of the phantom ncludes skn response and tumor response. To remove the skn response from the sgnal, we performed tme gatng usng knowledge of the dstance between antenna and phantom. After performng tme gatng, we construct the nput vector for the tranng set. Fgure 12 shows backscattered data vectors wth and wthout tumor. Each vector contans 4 concatenated tme-seres sgnal correspondng to the 4-elements antenna array. From ths ndvdual observaton, we could expect dfferent sgnal sgnature n terms of tumor presence. Ths leads to correct pattern recognton, n other words, correct tumor detecton. Fgure 13. Dscrmnant space vsualzaton for tumor detecton A total of 20 readngs are taken for breast phantom wth tumor and another 20 readngs are taken for breast phantom wthout tumor. K-fold valdaton wth K = 10 s appled for tranng and testng of data usng the lnear classfer. Fgure 14 shows the bnary classfer performance as a functon of SNR. LDI classfer s selected as the lnear classfer. The Addtve Whte Gaussan Nose (AWGN) s added to the backscattered sgnal manually. Ths fgure demonstrates the robustness of the lnear classfer, as the sclassfcaton Rate (CR) drops to below 1% when SNR = 9 db. Fgure 14: CR-SNR curve Thus, ths prelmnary experment proves the feasblty of applyng lnear classfer n the detecton of breast tumor. Once we have determned the presence of tumor wthn the breast tssue, we could move to the second stage to estmate the characterstcs of the tumor. Fgure 12: Space-tme samples The nput waveform forms the orgnal feature space for the classfer. The next step s to perform dmenson reducton usng PCA, to fnd the domnant feature of the target sgnature. Fgure 13 shows an example of the 2-D dscrmnant space from the PCA-based dmenson reducton for tumor detecton. We could see the tranng ponts form two clusters correspondng to the two classes, wth and wthout tumor, respectvely. The two test ponts le wthn two dfferent clusters suggestng the correct classfcaton result. B. Tumor Shape and Sze Classfcaton In a clncal envronment, to determne whether a tumor s bengn or malgnant, t s mportant to understand the characterstcs of the tumor. Wthout mage reconstructon, lnear classfer lacks capablty of provdng exact shape and sze nformaton of the tumor. To extract as much nformaton as we can, we develop mult-way classfers to dstngush tumor wth dfferent shape and sze. In order to expermentally test f the classfer could correctly characterze tumor wth dfferent shape and sze, alumnum blocks are used to smulate the tumor, as the alumnum s easy to carve. In order to better examne the tumor characterstcs, samples are collected n 4 drectons by the 4- element antenna array. In total we have 16 spatal samples, whch are concatenated nto one feature vector. Each spatal 263
sample contans 30 tme doman samplng ponts. The tme samples are extracted from a tme gatng wndow contanng the expected target backscattered sgnal. As the tme doman backscattered sgnal contans the target sgnatures, we are expected to be able to classfy the physcal characterstcs of the target usng mult-way classfer. The general procedure s the same as the bnary classfer dscussed n the frst stage. In the tumor sze and shape classfcaton experment, we test spheres wth dfferent dameters and dfferent shapes wth smlar volumes, respectvely. Fgure 15 vsualzes the sze and shape dscrmnatons of tumor. V. CONCLUSION & FUTURE WORK We have establshed an expermental UWB magng system for breast tumor detecton and characterzaton. Ths setup has low system complexty, whch s easy to realze n the laboratory envronment and could serve as a computer-aded dagnoss prototype n breast cancer research project. Usng ths smple system setup, the detecton result based on lnear classfer s promsng, whch successfully ndcates the presence of the tumor. We further explore the possblty of usng the lnear classfer to dstngush tumor wth dfferent sze and shape, the prelmnary expermental result shows the feasblty of ths dea. We wll keep workng on ths prototype system and the expected future works nclude: development of heterogeneous breast phantom and constructon of better transcever system for backscattered sgnal collecton. Fgure 15: Dscrmnant space vsualzaton for sze/shape classfcaton Overall, the expermental results are promsng. It llustrates the feasblty and advantage of usng lnear classfer n the computer-aded dagnoss of breast cancer based on UWB magng system. Due to the dffculty n constructng a relable heteogeneous breast phantom, we have just consdered homogeneous breast phantom. Also, we have assumed stuaton wth only a sngle tumor. So the experments and results presented are stll n the prelmnary stage. These restrctons are expected to be removed n future experments. REFERENCES [1] H. Lm, N. Nhung, E. L, N. Thang. Confocal mcrowave magng for breast cancer detecton: delay-multply-and-sum mage reconstructon algorthm n IEEE Transacton on Bomedcal Engneerng, vol. 55, no. 6, 1697-1704 (2008) [2] E. Fear, X. L, S. Hagness,. Stuchly. Confocal mcrowave magng for breast tumor detecton: localzaton of tumors n three dmensons n IEEE Transactons on Bomedcal Engneerng, vol. 49, no. 8, 812-822 (2002) [3] E. Fear, J. Stll,. Stuchly. Expermental feasblty study of confocal mcrowave magng for breast tumor detecton n IEEE Transactons on crowave Theory and Technques, vol. 51, no. 3, 887-897 (2003) [4] X. L, S. Davs, S. Hagness, D. Wede, B. Veen. crowave magng va space-tme beam formng: expermental nvestgaton of tumor detecton n multlayer breast phantoms n IEEE Trans. crowave. Theory Technques, vol. 52, no. 8, 1856-1865, (2004) [5]. Sll, E. Fear. Tssue sensng adaptve radar for breast cancer detectonexpermental nvestgaton of smple tumor models n IEEE Transactons on crowave Theory and Technques, vol. 53, no. 11, 3312 3319 (2005) [6]. Klemm, I. Craddock, J. Leendertz, A. Preece and R. Benjamn. Radar- Based Breast Cancer Detecton Usng a Hemsphercal Antenna Array- Expermental Results n IEEE Transactons on Antennas and Propagaton, vol. 57, no. 6, June 2009 [7] A. shra and S. Sardar. Applcaton specfc nstrumentaton and ts feasblty for uwb sensor based breast cancer dagnoss, n IEEE Conference on Power, Control & Embedded Systems, Nov 2010, pp. 1-3 [8] S. Alshehr, S. Khatum, Z. Awang. A UWB Imagng System to Detect Early Breast Cancer n Heterogeneous Breast Phantom, n Internatonal Conference on Electrcal, Control and Computer Engneerng, Jun 2011 [9] S. Davs, B. Van Veen, S. Hagness. Breast Tumor Characterzaton Based on Ultra-wdeband crowave Backscatter, n IEEE Transactons on Bomedcal Engneerng, Vol. 55, NO. 1, Jan 2008 [10] SAS-571 Double rdge horn antenna gude. A.H.Systems, Inc. http://www.ahsystems.com/ [11]. oosazadeh, C. Ghobad,. Doust. Small monopole antenna wth checkered-shaped patch for UWB applcaton n IEEE Antenna and Wreless Propagaton Letters, vol. 9, 2010 264