A feasibility study of novel ultrasonic tissue characterization for prostate-cancer diagnosis: 2D spectrum analysis of in vivo data with histology as gold standard

A feasibility study of novel ultrasonic tissue characterization for prostate-cancer diagnosis: 2D spectrum analysis of in vivo data with histology as gold standard
of 8
All materials on our website are shared by users. If you have any questions about copyright issues, please report us to resolve them. We are always happy to assist you.
Related Documents
  A feasibility study of novel ultrasonic tissue characterizationfor prostate-cancer diagnosis: 2D spectrum analysisof  in vivo   data with histology as gold standard Tian Liu a   Department of Radiation Oncology, Emory University School of Medicine, Atlanta, Georgia 30322 Mahesh M. Mansukhani  Department of Pathology, Columbia University, New York, New York 10032 Mitchell C. Benson  Department of Urology, Columbia University, New York, New York 10032 Ronald Ennis  Department of Radiation Oncology, St. Luke’s-Roosevelt Hospital, New York, New York 10019 Emi Yoshida, Peter B. Schiff, Pengpeng Zhang, Jun Zhou, and Gerald J. Kutcher  Department of Radiation Oncology, Columbia University, New York, New York 10032  Received 28 January 2009; revised 3 June 2009; accepted for publication 10 June 2009;published 1 July 2009  This study demonstrates the feasibility of using a novel 2D spectrum ultrasonic tissue character-ization   UTC   technique for prostate-cancer diagnosis. Normalized 2D spectra are computed byperforming Fourier transforms along the range   beam   and the cross-range directions of the digitalradio-frequency echo data, then dividing by a reference spectrum. This 2D spectrum method pro-vides axial and lateral information of tissue microstructures, an improvement over the current 1Dspectrum analysis which only provides axial information. A pilot study was conducted on fourprostate-cancer patients who underwent radical prostatectomies. Cancerous and noncancerous re-gions of interest, identified through histology, were compared using four 2D spectral parameters:peak value and 3 dB width of the radially integrated spectral power   RISP  , slope and intercept of the angularly integrated spectral power   AISP  . For noncancerous and cancerous prostatic tissues,respectively, our investigation yielded 23  1 and 26  1 dB for peak value of RISP, 7.8  0.5° and7.6  0.6° for 3 dB of RISP, −2.1  0.2 and −2.7  0.4 dB / MHz for slope of AISP, and 92  5 and112  6 dB for intercept of AISP. Preliminary results indicated that 2D spectral UTC has thepotential for identifying tumor-bearing regions within the prostate gland. ©  2009 American Asso-ciation of Physicists in Medicine .   DOI: 10.1118/1.3166360  Key words: prostate cancer, ultrasound tissue characterization, 2D spectrum analysis, spectral pa-rameters I. INTRODUCTION Prostate cancer is the most common cancer among Americanmen, second only to skin cancer. The American Cancer So-ciety projected 186 320 new diagnoses of prostate cancer inthe United States during 2008. 1 Since conventional imagingmodalities such as x-ray, ultrasound, CT, and MRI are unableto reliably distinguish cancerous prostatic tissues from nor-mal tissues, the majority of prostate cancers are diagnosed bytransrectal ultrasound-guided prostate biopsies. However, be-cause conventional ultrasound cannot pinpoint the locationof the malignant tumors within the prostate gland, theseultrasound-guided biopsies currently have an estimated sen-sitivity of 50%. 2 The ability to reliably image cancer-bearingregions within the prostate could increase the success rates of prostate-cancer detection as well as treatment.In the past decade, major advances have been made inprostate-cancer diagnosis that use advanced imaging tech-niques such as magnetic resonance spectroscopy   MRS   andultrasonic tissue characterization   UTC  . MRS provides bio-chemical   metabolic   information of the prostate gland anddetermines cancerous tissue by identifying regions with anelevated choline/citrate ratio, an indicator of cancer. 3–9 UTCprovides physical information of the prostate gland and de-termines cancerous tissue by evaluating the tissue vascula-ture as well as their attributes such as size, shape, andcompressibility. 10–13 Due to the safe, cost-effective, and real-time nature of ultrasound imaging, novel ultrasonic tech-niques are widely pursued to assess the malignancy of pros-tatic tissues. The following is a brief summary of variousultrasonic approaches in prostate cancer detection.  i   Elasticity approach : Ultrasound elastography   strainimaging   provides information about tissue elasticity  stiffness  . Konig  et al.  incorporated tissue elasticity asa factor in malignant-tissue detection of the prostategland, by tissue typing with a trainable classificationsystem. 14 Souchon  et al.  developed an imaging systemfor prostate elastography  in vivo , using a transrectalultrasound probe to guide a high-intensity focused ul- 3504 3504Med. Phys. 36  „ 8 … , August 2009 0094-2405/2009/36 „ 8 …  /3504/8/$25.00 © 2009 Am. Assoc. Phys. Med.  trasound therapy. 15 Hoyt recently reviewed the use of tissue elasticity properties as biomarkers for prostatecancer. 16 Zhang  et al.  described the viscoelastic prop-erties of the human prostate and correlated them to theinherent elastic contrast produced by cancer. 17  ii   Power Doppler method  : The power Doppler is a 3Dvascular-imaging technique that reveals abnormalblood flow. Histological studies have demonstrated in-creased microvessel density associated with prostatecancer as well as a positive correlation between mi-crovessel density and aggressiveness of disease. Sau-vain  et al.  reported that the power Doppler improvesdetection of abnormal vascular density in isoechoic tu-mors of the prostate. 18 In addition, the power Dopplercan determine the risk of extracapsular involvement,an important prognostic factor, by the presence of ves-sels perforating the capsule. 18  iii   1D spectrum analysis method  : The most widely inves-tigated UTC technique in prostate-cancer detection, todate, uses the 1D spectrum analysis method. The 1Dspectrum method analyzes the radio-frequency   rf   data received at the clinical ultrasound probe and pro-vides quantitative measures of tissue microstructurephysical properties. Feleppa  et al.   Riverside ResearchInstitute   pioneered utilization of 1D spectrum analysisin the early 1990s. In practice, nonlinear classificationmethods were applied to the spectral parameter values  slope, intercept, and midband value computed fromthe linear regression of the 1D spectral curve   alongwith PSA values in order to differentiate cancerousfrom noncancerous tissues. 19–21 With a database of over 200 patients, the 1D spectrum technique srci-nally yielded an area under the receiver-operating-characteristic   ROC   curve of 0.80  0.05. 20 More re-cently, this group reported a ROC curve area of 0.844  0.018. 13 Similarly, Scheipers  et al.  described amultifeature tissue characterization including 1D spec-tral and texture parameters to detect prostate cancer. 22 In a study of 100 patients, Scheipers’ method yielded aROC curve area of 0.86 when distinguishing hyper-echoic and hypoechoic tumors from normal tissue, anda ROC curve area of 0.84 when distinguishing iso-echoic tumors from healthy tissue. This performance isas good as, if not superior to, current prostate-cancerimaging by MRI/MRS with a mean area under theROC curve ranging from 0.69 to 0.76. 23 We have developed a novel 2D spectrum technique thatextends the analytic concept from one dimension to twodimensions. 24,25 The theoretical model for our calibrated 2Dspectrum analysis method was based on wave propagation ina globally homogenous medium with randomly distributedlocal inhomogeneities and is describable by a 3D Gaussiantissue model. 24,25 2D spectra are employed by performingtwo fast-Fourier transform   FFT   operations on windowed rf data obtained from a region of interest   ROI  . The rf dataalong each line were first transformed with respect to time  the range   and the resulting complex spectrum was trans-formed with respect to the cross range. This allows bothaxial and lateral evaluations of physical properties of pros-tatic tissues, an improvement over the 1D method that onlyprovides axial evaluation. Our 2D method supplies more in-formation regarding tissue microstructures   e.g., morphol-ogy  , thereby offering more diagnostically relevant informa-tion and potentially yielding a greater ROC curve area.The central premise enabling 2D UTC is that canceroustissues have different structural features than those of non-cancerous tissues. For example, cancerous tissues are harderthan normal prostatic tissues. 26 2D spectra are sensitive tophysical properties of tissue microstructures, such as size,shape, and spatial acoustic impedance fluctuation. Even formore challenging cases, such as tumors, which often posescattering from randomb   unpredictable   spatial distributionsof small scatterers, 2D spectrum analysis has the potential todifferentiate malignant from benign prostatic tissues. Thus,the measurements of the 2D ultrasonic spectra could prove tobe promising tools for physicians noninvasive prostate-cancer diagnosis.The purpose of this study was to demonstrate the clinicalfeasibility of, and lay the groundwork for, using our 2D spec-trum technique in prostate-cancer diagnosis. This paper pre-sents the following:   1   a summary of the theory underlying2D spectrum analysis;   2   the data acquisition system em-ployed in 2D spectrum analysis;   3   data processing proce-dures of this 2D method in prostate-cancer diagnosis;   4  histology evaluation, which serves as the gold standard forROI selection in clinical tissue typing;   5   preliminary resultsof our  in vivo  prostate-cancer study. II. SCATTERING THEORY UNDERLYING 2DSPECTRUM ANALYSIS The theory underlying our 2D spectrum analysis methodhas been discussed in previous reports but is summarized inthis section. 24,25,29 2D spectrum analyses are performed onthe digital rf echo signals gated over a ROI. Radio-frequencydata within the ROI undergo the 2D Fourier algorithm alongboth the range    X  , beam propagation   direction and the crossrange   Y  , scan   direction. Figure 1 is a diagram illustratingthe scanning geometry. Cross‐Range (Y) BeamROI  Prostate        R     a     n     g     e        (       X        ) F IG . 1. Diagram showing a transducer scanning the prostate, where X is therange direction   beam   and Y is the cross-range direction. 3505 Liu  et al. : 2D spectrum ultrasonic tissue typing of prostate 3505Medical Physics, Vol. 36, No. 8, August 2009  Equation   1   is the general equation of the calibrated 2Dpower spectrum for tissues exhibiting wide-sense stationar-ity. The equation demonstrates that 2D power spectra of rf signals are determined by three autocorrelation functions  ACFs  :  R T  ,  R  D , and  R G . The calibrated 2D power spectrumis a function of temporal frequency  f   and cross-range spatialfrequency    , S  2D   f  ,     =    R T     x  ,   y ,   z   R  D     ,   z    R G    x  ,   y  e  j 2 k    x  +  j     y e −  j       d    xd    yd    zd     ,   1  where      y −  y t  ,       y  −  y t   ,     =   y −   y t   −  y t   ,  R T   is the tis-sue acoustic impedance ACF,  R  D  is a spatial ACF describingthe two-way directivity function of the beam,  R  D     ,   z   =   D 2    ,  z   D 2     +    ,  z  +   z  d    dz ,   2  and  R G  is a spatial ACF of the 2D gating function,  R G    x  ,   y   =  g 1   x   g 1   x   +   x   g 2   y  g 2   y  +   y  dxdy .  3  The second degree of freedom in frequency domain allowsthe exposition of spatial anisotropy and orientation of 3Dtissue scatterers, previously unavailable in the 1D spectrummethod.We have solved this equation with a Gaussian approxima-tion for the tissue ACF    R T   . The closed form solution isexpressed as S   k  ,     =0.24  L  X   L Y  CV  s Q 2 k  3     a /  R   1 + 0.22  ka R   1.17  2  1 / 2  exp  −   k    1.17  2 −14     1.17  2 −10.88     Rka   2  ,   4  where  L  X   and  L Y   are ROI lengths along the range and cross-range directions, respectively,  a  and  R  are the aperture andfocal length of the transducer,  C   is the effective volumetricscatterer concentration,  Q  specifies the relative acoustic im-pedance difference between the scatterers   with acoustic im-pedance  Z    and the surrounding medium   with averageacoustic impedance  Z  0  ,  V  s  specifies the volume of an aver-age scatterer, and    ,    , and     represent effective scatterersizes along the  X  ,  Y  , and  Z   directions, respectively. The the-oretical analysis demonstrates that 2D power spectra are re-lated to tissue properties such as scatterer size and relativeacoustic impedance, transducer properties such as apertureand focal length, system properties such as bandwidth andcenter frequency, and processing properties such as ROI win-dow sizes.Although a Gaussian ACF is used in our theoretical modelfor randomly distributed scatterers, our 2D spectrum methodis not limited to any particular form of ACF. In addition, our2D spectrum would exhibit distinctive features for determin-istic structures. For example, if the structure has well-definededges, the spectrum will have a peak width that is inverselyproportional to the thickness of the structure. III. MATERIALS AND METHODS In this feasibility study, we scanned four prostate cancerpatients with transrectal ultrasound under an institutional re-view board-approved protocol at Columbia University Medi-cal Center. The patient selection criterion was prostate-cancer patients undergoing radical prostatectomy who didnot receive preoperative hormone or radiation therapy. Thepatient clinical characteristics are shown in Table I.The prostate 2D spectrum analysis study consists of thefollowing three elements.  a   Acquisition of rf data of the prostate using a clinicalultrasound scanner.  b   Ultrasonic data processing through 2D spectrum analy-sis.  c   Validation of the ultrasonic spectral method with patho-logical findings. III.A. Ultrasonic imaging and RF data acquisition Figure 2  a   is a schematic of our ultrasound imaging sys-tem. As previously described by Feleppa  et al. , 12,21 the dataacquisition system has three main components: a clinical ul-trasound machine, a custom interface module, and a com-puter. A B&K Medical System   Marlborough, MA   scanner  model 3535   was employed to transmit and receive ultra-sonic signals. The biplane transrectal probe   model 8558  had a nominal center frequency of 7.5 MHz and the convexarray transducer had a sector angle of 110°. The rf data  32-bits   was acquired using a 40 MHz sampling frequency.Each scan consisted of 192 scan lines and each scan lineconsisted of 3000 scan points. For all ultrasound scans, weused fixed scanner settings: 70% overall gain, a single focalzone at 0.5 cm, and the calibrated maximum setting timegain control   TGC  . In the future, physicians will be able toadjust the TGC level at their discretion and we will subse-quently adjust the spectral normalization to match the corre-sponding setting. T ABLE  I. Patient and tumor characteristics.Patient Age PSAProstatevolume  cm 3  Gleasonscore  1   61 4.8 18.8 7  2   54 5.0 24.2 7  3   60 10.7 51.8 7  4   56 3.0 26.3 7 3506 Liu  et al. : 2D spectrum ultrasonic tissue typing of prostate 3506Medical Physics, Vol. 36, No. 8, August 2009  During ultrasound imaging, each patient was scannedwith a transrectal ultrasound in the lithotomy position priorto radical prostatectomy. Ultrasound scans were performedalong the axial plane of the prostate in 2 mm incrementsfrom the base to the apex. An AccuSeed 3D stepper wasutilized   Fig. 2  b   to hold the probe and achieve a 2 mmstep size. 3D ultrasound rf data, backscattered from the pros-tate, was acquired as a set of closely spaced scan planes overthe full volume of the gland, and was digitally stored forfuture processing. 3D data acquisition enables 2D analysis of any ROI within the prostate gland and was further used toensure accurate correlation with 3D histology images forcancerous/noncancerous ROI identification. For a typicalprostate   4–6 cm from base to apex  , 20–30 scan planeswere required to cover the full gland volume. The duration of an ultrasound examination was between 5 and 10 min. III.B. 2D spectrum analysis method We developed  MATLAB  software for the analysis of thedigital rf ultrasound signals backscattered from the prostategland. 2D spectrum analysis consisted of the following steps. Step 1: B-model display . 2D spectrum analysis began witha digital computation of B-mode images. An  ULTRASONIX ®program was used to derive the envelope of stored rf echosignals. The enveloped data were then displayed in a sectorformat that matched the B-mode image generated by theclinical scanner   Fig. 3  . This B-mode image served to iden-tify overall anatomic relationships and acted as the templateupon which the ROI was specified. Step 2: ROI selection . In this pilot study, physiciansmanually specified the cancerous and noncancerous ROIswithin the transducer’s focal zone   prostate peripheral zone  based on the corresponding regions from the histology can-cer map. The prostate peripheral zone is the posterior portionof the prostate gland, closest to the rectum, where prostatecancers commonly develop. ROIs were positioned and super-imposed on the B-mode image, shown in Fig. 3. The ROI specified the region in which spectrum analysis was applied.Due to the small size of the ROI, we ignored the scan linedistance variation of the sector scan through the ROI. Thespectrum analysis was applied directly to the rf echo data  not the enveloped data displayed in the B-mode images  .There are tradeoffs between the resolution and accuracy inchoosing the ROI size. 27 For example, a larger ROI results inmore statistically accurate parameter estimates, yet simulta-neously results in decreased resolution. In consideration of this tradoff, we employed a ROI size of approximately4  4 mm. We chose to use a ROI size that is larger   20wavelengths at center frequency   than that used in 1D UTCstudies to ensure a less biased variance. 2D spectral param-eters have yet to be evaluated for resolution versus variance;however, the underlying statistical principles are consistentwith those developed by Lizzi. 27 Step 3: Calibrated spectrum . A calibration procedure wasperformed to remove the effects of the transducer and systemelectronic modules. The calibration spectrum was the 1Dpower spectrum of the echo signal received from an opticallyflat glass place, placed perpendicular to the beam axis in thefocal zone of the transducer   with a flat TGC  . The calibra-tion power spectrum   in dB   was later subtracted from the2D tissue spectrum along the range direction to compensatefor the temporal-frequency transfer function of the beam andreceiver. The temporal-frequency range used for 2D spec-trum analysis was chosen to ensure an adequate signal-to- RF SIGNALSTIMING AND CONTROLSIGNALS SCANNER COMPUTERSYSTEM PROBE (a) (b) RF SIGNALS   TIMING AND CONTROLSIGNALS SCANNER COMPUTERSYSTEM PROBE (a) RF SIGNALS   TIMING AND CONTROLSIGNALS SCANNER COMPUTERSYSTEM PROBE (a) (b)(b) F IG . 2.   a   Schematic of the ultrasound data acquisition system and   b   photograph of an AccuSeed stepper. 1 cm F IG . 3. Conventional B-mode image of the prostate. 3507 Liu  et al. : 2D spectrum ultrasonic tissue typing of prostate 3507Medical Physics, Vol. 36, No. 8, August 2009  noise ratio. The frequency bandwidth was determined by en-suring the lowest-level echo signal from the prostateparenchyma was at least 6 dB above the noise level of thetransducer. 28 In this study, no normalization/calibration wasperformed along the cross-range direction to compensate forthe beam-transfer function. Step 4: Spectrum generation . To compute the 2D spec-trum, the 2D rf signals inside the ROI were multiplied by a2D Hamming function with window lengths  L  X   and  L Y   rep-resenting the axial   range   and lateral   cross-range   direc-tions, respectively. A 1D FFT with respect to the axial direc-tion was applied for each scan line. The complex 1Dspectrum along each scan line was divided by the 1D cali-bration spectrum. 11 The resulting normalized complex 1Dspectrum was then Fourier transformed along the lateral di-rection to obtain a 2D spectrum. The spectral magnitude wassquared to compute a single realization of the 2D powerspectrum, shown in Fig. 4. The resulting 2D spectrum was specified in terms of spatial frequencies  k   k  =2    f  / c   alongthe range direction and       =2   / c   along the cross-rangedirection   where  c  is the speed of sound  . In this study, afrequency bandwidth range of 5–8.5 MHz was used foranalysis of all patients. The cross-range spatial frequencyranged from −100 to 100 cm −1 . The intensity of the 2D spec-trum was displayed in color, with red and blue indicatinghigh intensity and low intensity, respectively   Fig. 4  . All 2Dpower spectra reported were calculated in decibel units   dB  as  S  2D   k  ,    10 log  S  2D  k  ,    . Step 5: Feature extraction . To quantitatively measure andclassify the physical properties of tissue, we defined twospectral functions   Fig. 5   and two 2D spectral parametersfrom each spectral function. The two spectral functions, ra-dially integrated spectral power   RISP   and angularly, inte-grated spectral power   AISP  , are computed from the 2Dpower spectrum   in dB  . 29 RISP is defined as an integrationof spectral power along each radial line in the 2D spectra, asa function of its angle     measured from the  k  -axis. The inte-gration extends over the transducer’s bandwidth and the ex-pression for the RISP isRISP      =  k  1 k  2  S  2D   k  ,    dk   k  1 k  2 dk  ,   5     =  k   · tan     ,   6  where  k  1  and  k  2  are the minimum and maximum spatial fre-quencies in the bandwidth and     specifies the angle of theline of integration. Physically, RISP is the distribution of spectral power density for each radial line in the frequencyspace. AISP is defined as an integration of spectral powerover an arc at a specific spatial frequency. AISP measures thespectral power distribution on each arc in the frequencyspace. The arc is the part of a circle of fixed frequency  K  within the frequency range of interest. The expression forAISP isAISP  K    =    1   2  S  2D   k  ,    Kd       1   2 Kd    ,   7  K   =   k  2 +    2 ,   8  where    1  and    2  are the minimum and maximum angleswithin the frequency range of interest and  K   specifies thefrequency of the arc. The four spectral parameters,   1   peak value of RISP,   2   3 dB width of RISP,   3   slope ofAISP, and  4   intercept of AISP, were used to quantitatively evaluate     r    o    s    s  -    a    n    g    p    a     t    a    r    e    q    u    e    n    c    y    m    m    r    o    s    s  -    a    n    g    p    a     t    a    r    e    q    u    e    n    c    y    m    m    r    o    s    s  -    a    n    g    p    a     t    a    r    e    q    u    e    n    c    y    m    m    r    o    s    s  -    a    n    g    p    a     t    a    r    e    q    u    e    n    c    y    m    m 05 6 7 8 9-10010080-8060-6040-4020-2005 6 7 8 9-10010080-8060-6040-4020-2005 6 7 8 9-10010080-8060-6040-4020-2005 6 7 8 9-10010080-8060-6040-4020-200 - 25- 50     r    o    s    s  -    a    n    g    p    a     t    a    r    e    q    u    e    n    c    y    m    m    r    o    s    s  -    a    n    g    p    a     t    a    r    e    q    u    e    n    c    y    m    m    r    o    s    s  -    a    n    g    p    a     t    a    r    e    q    u    e    n    c    y    m    m        C    r    o    s    s  -     R    a    n    g     S    p    a     t     i    a     l     F    r    e    q    u    e    n    c    y     (     1     /    m    m     ) 05 6 7 8 9-10010080-8060-6040-4020-2005 6 7 8 9-10010080-8060-6040-4020-2005 6 7 8 9-10010080-8060-6040-4020-2005 6 7 8 9-10010080-8060-6040-4020-200 - 25- 50 TemporalFrequency(MHz)        C    r    o    s    s  ‐    r    a    n    g    e     S    p    a     t     i    a     l     F    r    e    q    u    e    n    c    y     (     1     /    m    m     ) F IG . 4. Calibrated 2D power spectrum of the ROI of prostate. ‐10 ‐5 0 5 10 Angle(degree) 2624222018        R       I       S       P       (       d       B       ) 5 6 7 8 Frequency(MHz) 605550457040       A      I      S      P      (      d      B      ) 65 (a)(b) F IG . 5. 2D spectra curves of the prostate tissue analyzed:   a   RISP and   b  AISP. 3508 Liu  et al. : 2D spectrum ultrasonic tissue typing of prostate 3508Medical Physics, Vol. 36, No. 8, August 2009
Similar documents
View more...
Related Search
We Need Your Support
Thank you for visiting our website and your interest in our free products and services. We are nonprofit website to share and download documents. To the running of this website, we need your help to support us.

Thanks to everyone for your continued support.

No, Thanks