Spectrum Tissue

Pattern Recognition Letters 24 (2003) 637658 www.elsevier.
com/locate/patrec
Ultrasonic spectrum analysis for tissue evaluation

Frederic L. Lizzi *, Ernest J. Feleppa, S. Kaisar Alam, Cheri X. Deng
Biomedical Engineering Laboratories, Riverside Research Institute, 156 William Street, New York, NY 10038, USA
Abstract Spectrum analysis procedures have been developed to improve upon the diagnostic capabilities aorded by conventional ultrasonic images. These procedures analyze the frequency content of broadband, coherent echo signals returned from the body. They include calibration procedures to remove system artifacts and thereby provide quantitative measurements of tissue backscatter. Several independent spectral parameters have been used to establish databases for various organs; several investigations have shown that these parameters can be used with statistical classiers to identify tissue type. Locally computed spectra have been used to generate sets of images displaying independent spectral parameters. Stained images have been derived by analyzing these parameter images with statistical classiers and using color to denote tissue type (e.g., cancer). This report describes spectrum analysis procedures, discusses how measured parameters are related to physical tissue properties, and summarizes results describing estimator precision. It also presents illustrative clinical results showing how such procedures are being adapted to address specic clinical problems for a number of organs. This report indicates where further developments are needed and suggests how these techniques may improve image segmentation for three-dimensional displays and volumetric assays. 2002 Elsevier Science B.V. All rights reserved.
Keywords: Ultrasonic imaging; Ultrasonic spectrum analysis; Prostate ultrasonography; Ocular ultrasonography; Ultrasonic parameter images
1. Introduction Over the past three decades, ultrasonic imaging has emerged as a standard diagnostic technique within a broad range of medical specialities (Kremkau, 1990). Pulse-echo ultrasound systems are routinely used to obtain cross-sectional images of the abdomen, heart, breast, prostate, and eye, and they have become the international standard
Corresponding author. Tel.: +1-212-502-1774; fax: +1-212502-1729. E-mail address: lizzi@rrinyc.org (F.L. Lizzi).
for imaging the fetus. The use of ultrasound is motivated by its proven clinical utility as well as several other factors including its safety record, real-time visualization, ease of use, and the availability of economic systems. Ultrasonography is likely to become even more useful because of ongoing developments (Goldberg et al., 1994; Sherar and Foster, 1989) that include advanced ultrasonic arrays and digital processing (for improved imaging), contrast agents (to enhance imaging and quantication of blood ow), probe miniaturization (for incorporation in catheters to examine blood-vessel disease), and very-high-frequency transducers (for improved spatial resolution).
0167-8655/03/$ - see front matter 2002 Elsevier Science B.V. All rights reserved. PII: S 0 1 6 7 - 8 6 5 5 ( 0 2 ) 0 0 1 7 2 - 1
638
F.L. Lizzi et al. / Pattern Recognition Letters 24 (2003) 637658
While conventional ultrasonic images (termed B-mode images) convey key diagnostic information, these images are degraded by phenomena that are not usually encountered with other imaging modalities, such as magnetic resonance and computed tomography, used in radiology. As described in this report, image degradations occur because of the coherent nature of ultrasound, the complex interaction of tissues and ultrasonic waves, and the fact that reections (echoes) are employed for imaging. These factors lead to random speckles, artifactual specular-reection dropouts, and spatially varying resolution in ultrasonic images. The extent of these degradations is aected by the instrument and the instrument settings which are used in examinations. The image degradations impede the direct application of conventional pattern recognition procedures to ultrasonic images. They also hinder attempts to quantify tissue features for objective diagnostic schemes. These considerations have prompted many investigations that seek to improve ultrasonic imaging and to provide a framework for quantitative tissue evaluation. This report describes frequencydomain approaches that have been developed to analyze ultrasonic echoes and to generate alternative types of ultrasonic images. Frequency-domain techniques oer fundamental advantages for addressing a number of constituent problems in ultrasonography. First, they permit a systems perspective that claries and separates the eects of system components and tissue properties on image features. Second, they aord a convenient means of incorporating well-established frequency-domain results describing ultrasonic beam propagation and tissue scattering. Third, averaged power spectra provide a cogent means of addressing the stochastic nature of tissue microstructure. Spectral techniques are designed to analyze coherent radio frequency (RF) echo signals, digitally acquired at the outputs of ultrasonic transducers, as opposed to the RF-signal envelope (video signals) that are displayed in conventional B-mode ultrasonic images (Lizzi et al., 1983; Feleppa et al., 1986). This is an important distinction because calibration and corrective procedures that can be applied to RF spectra are usually not ap-
plicable following the non-linear process of envelope detection. Video detection also obliterates subtle RF-signal features that can convey important information regarding tissue microstructure. The spectral techniques described in this report can evaluate two independent parameters that characterize ultrasonic scattering by tissues: one parameter provides a measure of overall scattering strength, while the other measures the frequency dependence of scattering. Under certain conditions (e.g., known or negligible acoustic attenuation), these spectral parameters can be used to estimate two independent physical properties of tissue constituents (related to their size and concentration) (Lizzi et al., 1987). Thus, unlike conventional ultrasonography, which provides a single qualitative image, spectral procedures can yield a pair of images depicting independent, quantitative tissue parameters. Spectral techniques have been clinically deployed in two complementary modes. The rst computes average spectral parameters within a demarcated spatial region: this mode is often employed in database studies that elucidate parameter values indicative of specic diseases. The second mode generates spectral parameter images that have been linked to quantitative clinical databases to assist disease detection and diagnosis (Feleppa et al., 1986). Spectral parameter images oer new opportunities for pattern recognition based on conjoint, independent parameters. They may become particularly valuable for automated boundary determination and segmentation of different tissue structures. These opportunities are just beginning to be explored and promise to become key elements for automated tissue biometry and three-dimensional (3-D) imaging. This report rst summarizes the operation of ultrasonic systems and describes image artifacts associated with dierent types of tissues. It then describes spectrum analysis and calibration procedures, and presents illustrative averaged spectra for dierent types of tissue. The report next describes how local spectral features are computed and displayed to form sets of cross-sectional spectral parameter images. Next, the report summarizes how spectral parameters are related to
639
physical scatterer properties including the eective sizes, concentrations, and mechanical properties of subresolution tissue constituents. The statistics of spectral parameters are described in terms of system and analysis parameters, and explicit relationships are presented regarding the trade-os between, e.g., spatial resolution and statistical variability in spectral parameter images. The last section of the report summarizes clinical spectral results for several organs and illustrates how spectral techniques can be adapted to meet particular medical needs. Because calibrated spectrum analysis provides quantitative outputs, images of spectral parameters can be compared to organ-specic databases that characterize sets of spectral parameters indicative of particular diseases. In examinations of the eye, where biopsies are precluded, database information is being used to identify and subclassify ocular tumors (Feleppa et al., 1986; Feleppa and Lizzi, 1993), using colored stains superimposed on two-dimensional (2-D) and 3-D images. Non-invasive treatment monitoring is also being implemented by staining tumor segments whose spectral parameters have changed due to microstructural alterations induced by radiotherapy or intense-ultrasonic therapy (Lizzi et al., 1997a). In prostate examinations, database information combining spectral parameters and blood levels of prostate specic antigen (PSA) are being used to detect prostate cancer (Feleppa et al., 1996, 1997); suspicious regions are being color coded to guide biopsy placement. In breast examinations, spectral parameters and morphological descriptors of tumor shapes are being used with the goal of dierentiating benign and malignant tumors to avoid the risks, expense, and anxiety associated with unneeded biopsies (Alam et al., 2000, 2002a,b). Cardiac examinations have employed a spectral parameter (integrated backscatter (IB)) to detect aberrant cyclic variations associated with myocardial infarction (ODonnell et al., 1981; Vered et al., 1987). Studies of the kidney have shown how spectral parameters and derived estimates can elucidate kidney microstructure and characterize kidney disease (Insana et al., 1991; Garra et al., 1994). Several studies have shown how spectral techniques may help diagnose local and diuse
liver disease (Oosterveld et al., 1991; King et al., 1985). Other promising results have been obtained for characterizing threats posed by vascular plaque (Lee et al., 1999) and for identifying cancerous metastases in lymph nodes (Tateishi et al., 1998). Recent in vitro studies have shown that spectral techniques may sense cell division and death (Kolios et al., 1999), thereby providing an important non-invasive potential for monitoring the ecacy of emerging tumor-therapy agents. In addition to tissue applications, the theoretical framework for spectrum analysis has been modied to treat ultrasonic contrast agents, so that spectra can be used to help improve ultrasonic evaluation of blood ow and tissue perfusion (Deng et al., 1998). This report cites our own research results to explain key points involved in spectral procedures and to provide a unied framework for describing the relations between theory, implementation, and clinical results. Using this framework, the report also cites relevant reports of the many other investigators who have made important contributions to developing frequency-domain approaches for ultrasonic examinations.
2. Conventional ultrasonic imaging A discussion of frequency-domain techniques requires a systems perspective identifying instrument and tissue components that inuence conventional ultrasonic imaging. Ultrasonic systems employ piezoelectric transducers that act as focused transmitters and receivers in a pulse-echo mode (Kremkau, 1990). Along a single look direction, the transducer is excited with a short voltage pulse and launches a brief ultrasonic pressure pulse. Each pulse comprises a series of alternating compressions and rarefactions that propagates through the body, via a coupling bath or gel. As the ultrasonic pulse propagates, its pressure amplitude is progressively diminished by absorption, and it is partially scattered by changes occurring in tissue density and/or acoustic propagation velocity. Tissues typically produce weak scattering so that most of the ultrasonic energy
640
continues propagating to deeper sites in the body. Portions of the backscattered pressure pulses propagate back to the transducer; the transducer coherently integrates these bipolar echo pulses over its receive aperture and generates a corresponding RF voltage pulse. As described above, the video envelope of the returned echo signals is used to generate image signals along a line corresponding to the transducer orientation. A cross-sectional image is synthesized by repeating the pulse-echo imaging operations along a series of scan directions oset by lateral beam displacements (linear scans) or angled beam displacements (sector scans). Simple systems utilize mechanically scanned transducers, typically with 12 cm radiating diameters, which are focused by acoustic lenses. Many modern systems employ piezoelectric arrays whose beams are scanned and focused under electronic control. The axial (depth) resolution of an ultrasonic system is equal to cT =2 where c is the tissues acoustic propagation velocity and T is the duration of the echo signal received from a single reector. (The factor 2 accounts for two-way travel.) T depends primarily on the bandwidth of the transducer, which is usually near 3040% of its resonant frequency fr . Lateral resolution is determined by the focused beamwidth, typically equal to 1.2F k=D where F is focal length, D is aperture diameter and k c=fr is the wavelength at the center frequency. It is useful to dene a resolution volume whose axial length is cT =2 and whose cross-sectional area is equal to the beamwidth; this volume encompasses the tissue volume contributing to received echoes at a single instant of time. These relationships show that both axial and lateral resolution improve as frequency increases. However, tissue attenuation increases with frequency, typically at a rate of 0.5 dB/MHz cm, limiting the useful depth of penetration at high frequencies. Because of this, organs that require deep penetration, such as the liver, are examined at relatively low center frequencies (e.g., 2 MHz) and relatively coarse resolution (near 1 mm) is achieved. At the other extreme, supercial sections of the eye can be examined using 50 MHz to achieve resolution ner than 50 lm. Most clinical
examinations employ center frequencies between 5 and 7.5 MHz providing a resolution near 0.3 mm. The following section discusses frequencydomain relationships governing echo signals from realistic tissue and system models. Corresponding time-domain expressions are complex, involving numerous spatially dependent convolutions and factors, such as attenuation and backscatter functions, that have only been well characterized over nite frequency bands. These considerations make it dicult to calibrate conventional images and to remove system-dependent artifacts. The summary descriptions presented above can help elucidate factors that aect ultrasonic images and complicate eorts at employing standard pattern recognition and image segmentation procedures. It is useful to consider two types of tissue structures, characterized as deterministic and stochastic. First, we consider simple deterministic or coherent reectors, which comprise smooth tissue surfaces that are broader than the beamwidth. Such surfaces may be found at the boundaries of organs, encapsulated tumors, and large blood vessels. In general, such surfaces are well imaged, but they constitute specular ultrasonic reectors and their echo amplitudes depend upon the angle of incidence of the ultrasound beam. When viewed obliquely, reections from these surfaces are not captured by the transducer, leading to specular drop-out. If the surfaces also exhibit a degree of roughness commensurate with the incident wavelength, then, various elements within the beamwidth can produce coherent interference phenomena, leading to a speckled brightness pattern as described below. Stochastic tissues are more complex and contain many closely spaced, independent scatterers within a resolution volume. This situation is frequently encountered in normal organs (liver, breast, prostate, etc.), in blood masses, and in many, if not most, types of tumors. The RF echo signals received from these tissues involve the coherent summation of pressure components from many constituent scatterers; the resultant RF signal is aected by the exact number and position of each scatterer in the systems resolution volume. The stochastic natures of scatterer density and location lead to a randomness in the RF signal
641
and result in a corresponding random modulation (speckling) in B-mode image brightness. Ultrasonic speckling is similar to that found with coherent laser illumination, and it can obscure ne tissue details, as discussed in (Abbott and Thurstone, 1979; Wagner et al., 1983). Speckling is evident in Fig. 1, which shows a Bmode image of the prostate obtained with a transrectal ultrasonic probe (located at the bottom of the image). The prostate is centrally located within the displayed 6-cm depth. Fig. 2 plots RF echo signals from adjacent scan lines within the rectangle that is superimposed on the prostate image. The statistics of echo signals and speckle from stochastic scatterers is of prime importance in signal and image analysis, and it has been studied by several investigators (Wagner et al., 1983; Tuthill et al., 1988). In simple situations where more than 5 or 10 independent scatterers are present in the resolution volume, the RF echo amplitudes exhibit a Gaussian probability density function (pdf) and corresponding video signals manifest a Rayleigh pdf. Speckling complicates boundary detection and segmentation operations. Several techniques to suppress speckle involve averaging B-mode images
Fig. 2. RF echo signals from prostate.
that exhibit independent speckle patterns. Independent speckle can be obtained by using B-mode images obtained at sequential time instants (where small transducer or tissue motion decorrelates speckle patterns) or images obtained using dierent frequency bands (since speckle is frequency dependent) (Magnin et al., 1982). Adaptive spatial ltering has also been employed (Bamber and Cook-Marten, 1987). We have developed a technique that lters received RF echo signals into M, ideally, non-overlapping frequency bands (Lizzi et al., 1986; Lizzi and Feleppa, 2000), the corresponding B-mode images (with independent speckle) are then averagedp to reduce the standard deviation of speckle by M . (As described in (Lizzi et al., 1986), pre-whitening is included to partially oset the concomitant loss in axial resolution.)
3. Ultrasonic spectrum analysis In the late 1960s, several investigators realized that the frequency dependence of tissue backscatter might convey useful information. Initially, such information was gathered by simply imaging tissues with transducers that had dierent center frequencies (Coleman et al., 1977). In the early 1970s, frequency characteristics were analyzed digitally or by applying RF echo signals to analog
Fig. 1. B-mode image of prostate.
642
spectrum analyzers (Namery and Lele, 1972; Lizzi et al., 1976). One innovative system generated color images depicting backscatter frequency characteristics by applying RF signals to a bank of three analog bandpass lters with dierent center frequencies: lter outputs were color coded and combined so that the resultant color indicated the spectral characteristics of echo signals (Purnell et al., 1975). In the latter 1970s, more quantitative approaches (Lizzi and Laviola, 1975) were designed to compute the calibrated average power spectrum obtained from a demarcated region of interest (ROI) in examined tissue. We will rst describe the ROI approach and subsequently developed spectral imaging techniques. In many clinical applications, the ROI approach is used rst in order to establish databases that delineate the spectral features of particular diseases; this information is then applied in spectral images to identify diseased tissues. We will then consider the theoretical relationships among derived spectral parameters and physical properties of tissue microstructure. Lastly, we will examine the statistics of spectral estimates and spectral parameter images. 3.1. Spectrum analysis procedures ROI spectrum analysis involves a sequence of operations applied to digitized RF echo signals digitally acquired directly at the transducer (Lizzi et al., 1983; Feleppa et al., 1986). In our investigations, acquired signals from an entire scan are rst used to compute and display a B-mode image, which serves to identify overall anatomic relationships. A mouse is then used to demarcate an ROI on the image, as shown for the prostate in Fig. 1. The ROI denes the range-segment length (L) and the number of adjacent scan lines to be used in the analysis. Note that the B-mode image serves only as a map to dene the ROI; the actual analysis is applied to acquired RF data. Along each bracketed scan line, stored RF echo data (as shown in Fig. 2) are multiplied by a Hamming window of length L, and a fast Fourier transform (FFT) algorithm calculates the RF echo spectrum. An average power spectrum is then computed as the mean of the squared spectral
magnitudes from the ensemble of bracketed scan lines in the ROI. A calibrated power spectrum is next calculated by dividing the echo power spectrum by the power spectrum of RF echoes from a planar calibration target; the target is placed in a water-bath and located in the transducers focal plane. For ROIs within the transducers focal zone, this calibration removes artifacts associated with the composite transfer function of the electronic transmitter/receiver and the transducer (Lizzi et al., 1983); in addition, spectra are corrected for the recorded system gain setting in order to provide a common spectral level for all tissue measurements. When ROIs are not located in the transducer focal zone or when arrays with dierent transmit and receive foci are employed, compensation for range-dependent diraction eects is required. This compensation uses data from targets (e.g., gel blocks) that contain diuse suspensions of small scatterers (e.g., glass beads). Diraction compensation is achieved by dividing tissue spectra by the diraction-target power spectrum measured at the same range as that of the ROI. Fig. 3 shows relevant prostate and calibration spectra for the ROI of Fig. 1. The single-element focused transducer used to obtain these data had a nominal 7-MHz center frequency, and RF data were acquired (8 bits) using a 50-MHz sampling frequency (Feleppa et al., 1997).
Fig. 3. RF tissue spectrum and calibration spectrum.
643
The gure shows plots of power spectra (dB) vs. frequency. The uncalibrated average tissue spectrum has a shape that is heavily inuenced by the system transfer function, as evidenced in the calibration (glass-plate) spectrum. In this illustration, useful signal-to-noise ratios occur within the 38MHz frequency band. Fig. 4 shows the calibrated power spectrum, obtained by dividing the RF tissue spectrum by the calibration spectrum. By removing the system transfer function, this process extends the bandwidth available for analysis. The calibrated spectrum is seen to exhibit a quasi-linear shape over its bandwidth, and the gure also plots the corresponding linear t calculated from linear regression analysis (Feleppa et al., 1986; Lizzi et al., 1987). Many investigators have described alternative approaches to the procedures described above. For example, Wear et al. (1994) have shown that ARMA spectral estimation procedures may reduce the required gated signal length and thereby improve spatial resolution for spectral techniques. Calibration procedures, to remove extraneous system factors, have been treated by Waag and Astheimer (1993), Insana et al. (1994) and Boote et al. (1988). In addition, a number of investigators, including Kuc and Schwartz (1979), have described related methods that compute acoustic attenuation coecients by analyzing spectra over a range of depths in tissue.
3.2. Representative tissue spectra Early studies showed that it was useful to interpret spectra in terms of deterministic and stochastic tissue structures, as described above for conventional ultrasonic images. Many deterministic structures produce a scalloped pattern of periodic spectral peaks. This shape is most evident in tissues (e.g., detached retinas) with two well-dened boundaries, as shown in (Lizzi et al., 1983). This spectral shape arises because the RF echo complex has the form q1 et q2 et 2w=c where et is the echo from a single surface, w is the thickness of the tissue, and q1 and q2 represent the reection coecients of the rst and second surfaces, respectively. If these reection coecients are equal, the power-spectral magnitude of the RF echo will be proportional to 2 j cos2pfw=cj and successive spectral peaks will occur at frequencies separated by an interval of c=2w. As noted above, stochastic tissue structures (e.g., many tumors) contain numerous small scatterers with randomly distributed positions. The ROI spectra for such tissues exhibits relatively smooth monotonic shapes without prominent peaks. As shown for the prostate, in Fig. 4, these spectra (in dB) often exhibit quasi-linear shapes over the limits of the measurement bandwidth so that linear regression analysis is useful in summarizing key spectral features. Mixed tissue segments combine deterministic and stochastic elements. For example, normal liver has been reported to contain stochastic elements that are situated within a quasi-periodic matrix with a periodicity near 1 mm (Fellingham and Sommer, 1984). In this case, ROI spectra show small periodic peaks which can be used to estimate the periodic spacing interval within the tissue. Several methods have been employed to estimate this spacing; these can involve power spectra (as described above for detached retinas), autocorrelation functions (ACFs) (computed as the inverse FFT of the power spectrum), and cepstra (computed as the inverse FFT of the logarithm of the power spectrum) (Lizzi et al., 1981). The spectral parameters described above have been used with ROI procedures to establish
Fig. 4. Calibrated power spectrum.
644
clinical databases for a number of organs, as described in subsequent sections. Database studies typically involve the measurement of ROI spectra and correlation of spectral parameters with tissue type following denitive diagnosis of the examined tissue. Denitive diagnosis is usually obtained from histological examination of biopsied specimens. Statistical procedures including linear discriminant analysis, cluster analysis, or neural networks (NNs) have been used to identify conjoint values of spectral parameters (e.g., slope and intercept) that are correlated with specic diseases. 3.3. Spectral parameter images Sets of quantitative descriptors have been dened to summarize the key features of tissue spectra so that comprehensive database studies can be conducted for tissue identication. These descriptors are also used for spectral images, which display the spatial patterns of tissue properties. For deterministic structures, tissue thickness or spatial periodicity is the most important parameter. For stochastic tissues, usually of most interest, linear regression analysis (Fig. 4) has been applied to spectra expressed in dB in order to determine two basic parameters: spectral intercept (dB; extrapolation to zero frequency) and spectral slope (dB/MHz) (Lizzi et al., 1987, 1997a; Feleppa et al., 1996). An additional parameter which is often employed is the midband t (dB), dened as the
value of the regression line at the center frequency fc of the spectral band. The midband t is numerically equal to the average value of the spectrum over the measurement bandwidth. Thus, the midband t is directly related to a widely employed parameter termed integrated backscatter or IB (ODonnell et al., 1979) which is also calculated from the average value of spectral levels (in dB). (While only two linear regression parameters are independent, it is useful to consider all three as illustrated in following sections.) Spectral parameter images display local values of spectral slope, intercept, or midband t in tissue cross-sections (Feleppa et al., 1997). The images are formed using a sliding Hamming window to progressively analyze RF data along each scan line. At each window site, all of the spectral, calibration, and linear-regression procedures described above are employed to compute unaveraged calibrated spectra and corresponding local values of spectral parameters. These values are then encoded in gray scale or color, and a set of images is displayed using a common cross-sectional format. Fig. 5 shows midband t (Fig. 5a) and intercept (Fig. 5b) images of the prostate. Spectral parameter images of, e.g., slope and midband t are often analyzed in conjunction with database classiers from ROI spectral studies. The images are analyzed on a pixel-by-pixel basis to determine whether spectral properties are consistent with those of a particular disease. Stained
Fig. 5. Spectral parameter images of prostate: (a) midband t; (b) intercept.
645
Fig. 6. Stained image highlighting suspected prostate cancer.
images, using color-coded presentations, are then synthesized to depict classier outputs indicating tissue type (Lizzi et al., 1986; Feleppa et al., 1997). Fig. 6 presents a stained image of the prostate, which was generated by analyzing the images of Fig. 5 with a NN classier trained using an ROI database, as described in a subsequent section. Areas of suspected cancer are presented as red overlays, which appear as bright areas in this black-and-white reproduction. Staining procedures may be considered a form of image segmentation that spatially separates normal and abnormal regions. Because they convey independent information, pairs of spectral parameter images should also be useful in more general segmentation and boundary detection procedures that delineate dierent organs and tissue surfaces. These applications have not yet been systematically explored, although relevant theoretical treatments of parameter statistics are now available, as described subsequently. 3.4. Theoretical relationships between spectral parameters and tissue microstructure Empirical clinical studies, described below, have demonstrated that linear-regression spectral parameters are correlated with tissue type in several organs. We have developed a theoretical model in order to elucidate this correlation (Lizzi et al.,
1983, 1987). The analysis treats the relationships between physical properties of tissue microstructure and corresponding values of spectral slope, intercept, and midband t. Our model treats weak scattering (Born approximation) from tissue segments in the focal zone of a typical (weakly focused) ultrasonic transducer, where quasi-plane wave conditions exist. The stochastic tissue microstructure is specied in terms of an ACF that describes the spatial disposition of acoustic impedance within the examined tissue. (Acoustic impedance is the product of tissue density and acoustic propagation velocity.) The model assumes that tissue microstructure is statistically homogeneous in the examined region and that its statistics exhibit wide-sense stationarity. A number of acoustic-impedance ACFs have been examined with emphasis on isotropic 3-D Gaussian functions (Lizzi et al., 1987; Lizzi, 1998). This ACF applies to a continuous spatial variation of acoustic impedance, which is appropriate for most tissues. However, it has proven useful to interpret theoretical results in terms of a spatial distribution of eective, discrete scatterers characterized by three physical parameters: d, scatterer diameter (determined by the width or correlation dimension of the tissue ACF); C, the spatial concentration of scatterers; and Q, the fractional dierence between the acoustic impedance of scatterers and that of the surround. (The peak, zerolag, value of the tissue ACF is proportional to CQ2 .) Mathematically, the model shows that calibrated power spectra are determined by the tissue ACF, the ACF of the beam-pattern cross-section, and the ACF of the gating function (taken to be a Hamming function). We will rst consider results in the absence of intervening ultrasonic attenuation. As shown in (Lizzi et al., 1987), when d is much smaller than the wavelengths in the spectral frequency band, spectral amplitudes increase dramatically with increasing frequency. This situation corresponds to Rayleigh scattering where spectra vary as the fourth power of frequency. For larger scatterers, spectra increase more slowly with frequency; for scatterer sizes commensurate with or larger than constituent wavelengths, spectra decrease as a function of frequency.
646
Fig. 7. Plots of spectral slope and intercept vs. scatterer size for typical prostate examinations.
Linear regression was applied to theoretical spectra in order to relate these analytical results to the summary spectral parameters used in clinical studies. Fig. 7 plots the resulting spectral slope as a function of scatterer size for conditions relevant to prostate examinations (5.75-MHz center frequency and 4.5-MHz bandwidth). For small scatterers (d < 0:05 mm), spectral slope is relatively constant at a value corresponding to f 4 Rayleigh scattering; the slope progressively decreases from this value as scatterer size increases. Fig. 7 also plots the spectral intercept value for these scatterer sizes with CQ2 set to unity. For small scatterers, intercept increases rapidly with size, since power spectral amplitudes increase as d 6 for Rayleigh scatterers. A recent report (Lizzi, 1998) presents closedform general expressions that relate spectral intercept, slope, and midband t to d and CQ2 . These expressions involve a number of parameters which depend on the transducer dimensions, center frequency, and bandwidth. They permit d and CQ2 to be evaluated when ultrasonic attenuation is negligible in intervening tissue (interposed between the transducer and examination site). In general, bandwidth is particularly important for estimation of small scatterer sizes and, as shown below, it is also a critical factor in the precision of spectral parameter measurements. Insana et al. have described a useful alternate theoretical scattering model, which analyzes spectral shapes in terms of form factors associated with
the dimensions and morphology of tissue constituents (Insana and Brown, 1993). Intervening attenuation can be neglected in supercial organs (e.g., in 10-MHz ocular examinations), but it becomes a factor in examinations of deeper organs. In tissue, attenuation (dB) usually increases in an approximately linear fashion with frequency, and attenuation coecients are specied in dB/cm MHz. To consider the eects of attenuation, we treat an intervening tissue of depth X and average attenuation coecient a. We denote the ideal (a 0) value of spectral intercept as I, ideal spectral slope as m, and ideal midband t as M. The values of these parameters that are measured in the presence of intervening attenuation are indicated with a prime. The measured spectrum (dB) is equal to the ideal spectrum (dB) minus 2aXf , where the factor of 2 accounts for two-way beam transit. From this relation it follows (Lizzi et al., 1997a) that: I0 I m 0 m 2aX M 0 M 2aXfc 1a 1b 1c
where fc is the center frequency of the spectrum. These results demonstrate that spectral intercept values are not aected by intervening attenuation. Slope and midband t are aected in ways that can be easily accounted for, provided a and X
647
are known. Many clinical studies correct m0 and M 0 based on tabulated values of a for the organ being studied together with measurements of X from B-mode images (Feleppa et al., 1997). These corrections are used in both database studies and parameter images to remove the eects associated with tissue depth. The corrections also permit d and CQ2 to be estimated from spectral data using the relations in (Lizzi, 1998). 3.5. Statistics of spectral estimates and images The statistics of spectral estimates are important considerations in planning ROI spectral procedures and in analyzing spectral parameter images. The variance levels associated with ROI spectral estimates aect the degree of condence that can be realized in diagnostic applications. The variance associated with spectral images directly aects image quality. In both cases, trade-os exist between spatial resolution and statistical stability. We have analyzed the statistical quantities that are related to the spectral parameters employed in our investigations (Lizzi and Laviola, 1975; Lizzi et al., 1997b,c). Other investigators have also described relevant statistical models (Huisman and Thijssen, 1996; Chaturvedi and Insana, 1996). Our analysis of spectral statistics applies to the same situation that was used in discussing the statistics of conventional ultrasonic images (Lizzi et al., 1997b,c). We consider statistically homogeneous stochastic tissue segments whose RF signal amplitudes exhibit Gaussian statistics. The calibrated RF spectrum is computed with a Hamming window of length L along a set of adjacent independent scan lines. Along each line, the spectrum extends over a total bandwidth B, and it includes n spectral resolution cells, each of which has a subbandwidth b. The value of b is determined by the 3 dB width of the gating functions Fourier transform. For a Hamming window of temporal duration TH , b 1:33=TH where the corresponding spatial extent of the gate is L cTH =2. The value of c is typically near 1.5 mm/ls, so that b 1=L where b is specied in MHz and L in mm. Thus, the number of spectral resolution cells is n BL where B is that total spectral bandwidth (MHz) being analyzed.
The conditions stated above are the same as those treated in spectrum analysis of Gaussian white noise. As described in (Lizzi et al., 1997c), the n spectral resolution cells constitute independent random variables whose magnitude exhibits a Rayleigh pdf, and the corresponding power spectral values exhibit v2 (negative exponential) statistics. We now consider the statisticsmost importantly the standard deviationsthat apply to spectral slope, intercept, and midband t. As discussed in (Lizzi et al., 1997c), the pdfs of these parameters are aected by the non-linear conversion to dB and the linear-regression operators. The results are somewhat dierent for ROI and parameter-image values. For ROI analysis, power spectra (with v2 statistics) from N independent scan lines are averaged, and the result is converted to dB; the resulting n independent average-spectra cells are then analyzed with linear-regression techniques. We denote the mean values of ROI estimates for midband t, slope, and intercept as M , m, and I , respectively. If N is greater than about 10, the standard deviations of midband t (rM ), slope (rm ) and intercept (rI ) are: k rM p 2a n p k 12 n 1=2 2b rm B n2 1 2 2 1=2 rI r2 2c M fc rm p where k 4:34/ N dB. For spectral parameter images, individual (unaveraged) power spectra with v2 statistics are converted to dB and analyzed with linear-regression techniques. We designate means and standard deviations using a prime together with the corresponding ROI symbols. The means of midband t, slope, and intercept, respectively, are: M M 2:5 dB; I I 2:5 dB
0 0
m0 m; 3
The standard deviations for each unaveraged spectral parameter have the same forms as Eqs. (2a), (2b) and (2c) with k 5:6 dB.
648
Eq. (3) shows that the mean values of midband t and intercept for parameter images are 2.5 dB lower than ROI values; this is due to the fact that ROI averages are computed prior to dB conversion while parameter-image averaging occurs after conversion to dB. Mean slope values are the same for both approaches. The standard-deviation results for ROI and parameter images have similar forms and demonstrate the importance of the factor n BL in determining estimator precision. The spectral bandwidth B is particularly important for slope estimation because, for typical cases (n2 1), the standard deviations for slope are inversely proportional to B1:5 . Overall, estimates of midband t exhibit smaller standard deviations than other parameters. The pdfs for parameter images are derived in (Lizzi et al., 1997c). Midband t results are plotted for various values of n BL in Fig. 8 where the abscissa is midband t (dB) minus its mean value. The gure shows several interesting characteristics. The pdfs for small values of n are asymmetric; as n increases, the curves become more symmetric, approaching Gaussian curves for values of n near 10. Furthermore, the shape of the pdf curve does not depend on the mean midband-t value. Slope pdf results are symmetric about their mean value and approach Gaussian shapes for values of n near 5. Intercept pdf results depend upon slope and midband t pdfs together with the values of fc and B. The theoretical pdf results have been conrmed by deriving corresponding histograms of clinical
spectral parameter images. Homogeneous tissue regions have been examined in several organs (Lizzi et al., 1997c), and excellent agreement has been found with theoretical results. As expected, departures from theory occur when heterogeneous tissue regions are examined, violating the assumption of homogeneity. We are now examining whether these departures can be used to form quantitative measures of tissue heterogeneity. The statistical results described above have proven valuable for evaluating the merits of various ultrasonic systems (in terms of B and fc ) and for designing appropriate ROI and sliding-window parameters (L and N). The results explicitly describe the trade-os that must be made between statistical stability and spatial resolution. Specically, Eqs. (2a), (2b) and (2c) shows how the standard deviations of each parameter estimator is related to spatial resolution, which is determined by the window length L and the number of averaged scan-line segments N. The pdf results for spectral parameter images are relevant to tissue segmentation and boundary detection. These applications could also benet from the fact that two independent parameters are imaged. Additionally, segmentation should be expedited since the shapes of the pdf curves do not vary with the mean value of each parameter. Studies in these applications are still in their early stages and are largely motivated by the need for automated 3-D imaging and biometry, as described in subsequent sections.
4. Illustrative clinical results The spectrum analysis procedures described in preceding sections have been applied to clinical data obtained from a number of organs. The procedures have been adapted to address specic needs germane to the organ being studied. The objectives include the detection, diagnosis, and staging of both focal and diuse diseases. Information from spectral procedures has also been used for treatment planning and treatment monitoring in several organs. Most clinical applications of spectrum analysis involve two phases. First, database studies, using
Fig. 8. Probability density functions for midband t.
649
ROI procedures, are performed with statistical classiers to identify conjoint values of spectral parameters that are associated with specic diseases, as identied in subsequent biopsy procedures. In studies of supercial tissues (e.g., the eye), intervening attenuation can often be neglected so that measured values of spectral slope and midband t can be used directly in databases. In other organs, spectral intercept, which is not aected by attenuation, can be used together with slope or midband t after they are be compensated for estimated attenuation, as described in a preceding section. Several specialized parameters do not require attenuation compensation: these include measurements of tissue periodicity (from analysis of spectral peaks) or assays of heterogeneity (using standard deviations of spectral parameters). In addition, attenuation does not aect measurements of cyclic temporal changes in cardiac parameters. Some database studies have included auxiliary quantities, such as the measured blood level of PSA, to complement spectral data. The second phase in clinical applications typically involves the synthesis of spectral parameter images. The particular parameters to be imaged and the inclusion of attenuation compensation are planned using database results for the organ under study. Classication procedures developed in database studies can be used to process sets of parameter images and delineate tissue regions likely to contain specic diseases. Some clinical studies have incorporated theoretical results to generate images of eective scatterer size and CQ2 . The following sections illustrate how these procedures have been applied to a number of organs. 4.1. Ocular examinations Ultrasonic spectrum analysis was initially applied in ophthalmology in order to identify tumors developing in the posterior section of the eye (Feleppa et al., 1986). These studies were conducted using focused transducers with center frequencies of 10 MHz. More recently, spectrum analysis has been employed using focused PVDF transducers with center frequencies near 50 MHz (Silverman et al., 1995). These very-high-frequency
examinations have been used to evaluate small (1 2 mm) tumors in the anterior ocular segment, to examine traumatic injury, and to study microstructural changes associated with glaucoma. Clinical database studies (10 MHz) showed that ocular tumors could be classied using spectral slope and intercept; classication was improved to a modest extent by also including the residual uncertainty of the linear t. Classication was accomplished using linear discriminant analysis incorporating two discriminant functions (linear combinations of these three spectral parameters) (Feleppa and Lizzi, 1993). Discriminant maps delineated regions associated with malignant melanomas, metastatic carcinomas, and choroidal hemangiomas (Coleman et al., 1983). Melanomas could be further subclassied into two groups (types B and E) that were correlated with conventional histologic subclassications. These classication capabilities are clinically important because biopsies cannot be conducted within the eye and because these tumors can require dierent treatments, ranging from continued observation to enucleation (surgical removal of the globe). Ocular tumor spectra were also analyzed using the theoretical model described above (Lizzi et al., 1983; Feleppa et al., 1986, 1988). Fig. 9 shows how the dierent tumor types characteristically exhibit dierent values of eective scatterer size and CQ2 (termed acoustic concentration). Melanomas exhibit the smallest scatterer sizes, thought to be associated with small vessels as well as aggregations of specialized cells (melanocytes) that ingest the dense melanin pigment from these tumors. Metastatic carcinomas exhibit larger scatterer sizes, possibly associated with aggregations of cells within the tumor stroma. Hemangiomas exhibit large scatterer sizes, associated with internal blood-lled cavities. The results in Fig. 9 demonstrate the importance of measuring two complementary features: size or CQ2 values by themselves would not produce reliable classication of these tumors. Long-term clinical studies of melanoma patients showed that spectral parameters and derived estimates of scatterer properties can directly benet treatment planning (Coleman et al., 1990, 1991). These ROI studies found that such
650
Fig. 9. Scatter plot of CQ2 (acoustic concentration) and scatterer size for ocular tumors.
parameters in combination with tumor location in the globe can be used as predictors of survival times following alternative procedures (enucleation and radiation therapy). Additional clinical studies have shown that spectral estimates of scatterer properties can help identify small anomalous blood vessels that are associated with increased lethality potential in melanomas (Silverman et al., 1997). Animal and clinical ROI studies have also shown that tissue changes induced by both radiotherapy and ultrasonic hyperthermia cause significant elevations in intercept and CQ2 (Silverman et al., 1986). This nding is now being used to develop non-invasive treatment monitoring procedures. The ocular ROI studies described in the preceding paragraphs have been used to develop a number of special types of spectral parameter images. Ocular-tumor databases have been employed to synthesize stained images that use color to encode tumor type (Lizzi et al., 1986). These images are generated from a pair of images of spectral slope and intercept, respectively. Slope and intercept values at corresponding locations in each image were compared with database results to determine whether their conjoint values were
indicative of a specic tumor type. The stained image was then generated: locations whose parameters were indicative of malignant melanomas were displayed in, e.g., red; metastatic carcinoma was indicated in blue; the anechoic vitreous humor was indicated with black; other tissues were indicated with green (see cover illustration of Optics and Photonics News, Vol. 5 (1), 1994). The stained image of Fig. 10 displays a segment of a posterior tumor (2-mm thickness) whose shape conforms to the curved posterior sclera. In this gray-scale reproduction, metastatic carcinoma is represented in bright white, normal tissue in gray, and vitreous humor in black. The predominant color (bright white) of this tumor correctly classied it as a metastatic carcinoma. A second type of parameter image has been developed to delineate ocular-tumor regions whose microstructure has been altered by radiotherapy or ultrasonic heating. These images were developed based on ROI studies showing that such alterations produce substantial changes in intercept and
Fig. 10. Stained image highlighting metastatic carcinoma in rear segment of eye.
651
CQ2 . Parallel-plane scan data has been used to display color-coded intercept in 3-D tissue blocks for this purpose. As shown in (Lizzi et al., 1997a), post-treatment results can be sliced and compared with corresponding pre-treatment results to identify responsive tissue segments. An alternative dierential stain approach (Feleppa and Lizzi, 1993) can be employed to identify these segments. In this approach, pre-treatment slope and intercept values are computed from multiplane spectral parameter images and their joint distribution is computed. Following treatment, scan planes are again examined, and regions whose conjoint slope and intercept values lie outside of the pre-treatment range are highlighted. This method has identied internal tissue responses occurring before changes in tumor volume, and it is being used in long-term patient follow-ups. A third type of parameter image displays local computations of eective scatterer size and CQ2 . Fig. 11 shows clinical results obtained with a veryhigh-frequency ultrasound (VHFU) transducer (40-MHz center frequency), which aords a large (50 MHz) spectral bandwidth (Lizzi et al., 1992). The gure shows scatterer size and CQ2 within the region of a blood mass (hyphema, labelled H) in the anterior chamber, posterior to the cornea (labelled C), which has a 0.5-mm thickness. Regions where blood organization (clotting) is occurring demonstrate larger sizes and higher CQ2 levels, as
Fig. 12. 3-D excised image showing anterior ocular melanoma.
seen in the center of the hyphema. Such information may help in planning optimal treatments to restore visual acuity. VHFU spectral parameter images are starting to be employed with 3-D images of the anterior eye (Silverman et al., 1995). Fig. 12 shows a clinical example in which a simulated excision has been made through sections of the cornea (dark gray) and sclera (white) to reveal an underlying ciliarybody malignant melanoma. Midband-t images (from parallel scan planes) are especially useful in these applications because large VHFU bandwidths provide very stable (low rM ) gray-scale images that facilitate image segmentation. Thus far, computer-assisted segmentation has been employed to identify ocular structures and assign relevant optical properties (color, transparency, etc.) for 3-D rendering. An operator brackets each structure with coarse lines, and a threshold search algorithm is used for ne boundary delineation. Eorts are underway to totally automate this process, so that such detailed 3-D images can be used in applications that include diagnosis as well as realistic tumor-treatment simulations to help design optimal ultrasonic therapy beams (Lizzi et al., 2001). 4.2. Prostate examinations Ultrasonic spectrum analysis is being investigated for diagnosing and staging prostate cancer (Feleppa et al., 1996, 2001). As in ocular-tumor
Fig. 11. Images of eective size and CQ for anterior-chamber hyphema in eye.
652
studies, a clinical database has been established using ROI approaches to identify spectral parameters that can distinguish cancerous from noncancerous prostatic tissue. However, the spectral procedures have been modied to address specic clinical problems associated with prostate examinations. Conventional ultrasound, using transrectal probes, is commonly used for biopsy guidance in patients with elevated serum levels of PSA. Unfortunately, cancerous regions do not exhibit distinctive features in prostate B-mode images, and biopsy placement in the prostate is essentially random. Consequently, many prostate cancers (at least 30%) are missed by the, typically, six biopsy samples obtained from subjects. Thus, a rst goal has been to determine whether stained spectral images could be employed to highlight prostate regions likely to contain cancer. If successful, clinicians could biopsy these suspicious sites, reducing false-negative rates and permitting prompt treatment. To investigate this application, ROI spectral studies have been conducted using a B&K Medical Systems Model 3535 which employs a mechanical sector-scan transrectal probe (Feleppa et al., 1996, 1997). RF data are acquired from an entire scan plane immediately preceding the insertion of the biopsy needle. Calibrated power spectra are subsequently computed using an ROI that encompasses the biopsy site, which is spatially registered
Fig. 13. ROC curves for prostate-cancer classication.
with the ultrasonic image. Database correlations with subsequent biopsy reports employ the spectral intercept and midband t as well as the patients PSA level. Classication studies have employed nearestneighbor and NN techniques, using jack-knife statistical procedures (Feleppa et al., 2000, 2001). Results are documented in terms of receiver operator characteristics (ROC). As shown in Fig. 13, ROC curves plot the true positive fraction, TPF (correctly identied cancers) vs. the false positive fraction, FPF (non-cancers incorrectly classied as cancer); these fractions are plotted as the classication threshold is varied. For an ideal classier, the true positive fraction would always be unity, and the area under the ROC curve would equal one. For a totally random classier, the ROC curve would lie along the diagonal of the axes, and the ROC area would be 0.5. The results plotted in Fig. 13 were obtained from 1019 biopsies. The upper curve was obtained using spectral procedures with a NN classier. Several dierent NN models were evaluated; the best performance was obtained with a multi-layer perceptron model using o-the-shelf NConnect software. The inputs to the NN were spectral, intercept, and midband values along with PSA level; the gold standard for training and evaluating the classier was biopsy histology. The lower curve was derived from numerical levels of suspicion (LOS) that were independently assigned by clinicians based on conventional B-mode images and ancillary clinical data such as PSA levels. (The LOS increases from 1 to 5 with increasing suspicion of cancer.) The plotted results show that the spectral classication, which has an ROC area of 0:85 0:05, is signicantly superior to the clinicians classication, which has an area of 0:66 0:03. As in ocular applications, these ROI database results were used to plan stained images using spectral parameter images of intercept and midband t (see Figs. 5 and 6). Paired local values of these parameters were analyzed together with the patients PSA level. (PSA provides no spatial information, but it has been found to alter spectral parameter values associated with prostate cancer.) In order to expedite image synthesis, the NN, de-
653
veloped from the ROI studies, was used to establish a 3-D array containing a score indicative of the likelihood of cancer (LOC) as a function of intercept, midband t, and PSA. This array is now employed as a lookup table to evaluate LOC at each image site. Sites whose LOC exceeds a selected threshold appear as a red overlay on a Bmode or midband-t image. 3-D stained images have been generated by applying the above procedures to RF data acquired in sets of parallel planes. First, the color overlay is applied to highlight high LOC regions in each scan plane. Second, the prostate capsule is demarcated manually and the extraprostatic regions are masked. Finally, the masked, colorencoded images are depicted as surfaces (the capsular and tumor-foci surfaces) by assembling the set of 2-D images into a 3-D rendering. 3-D renderings are generated using available software such as VTK, AVS, or Noesys. Fig. 14 shows a black-and-white reproduction of an interactive 3D rendering in which the smooth prostate capsule bounds several foci of suspected cancer. A subsequent biopsy veried that cancer was present in the right apex of the gland, consistent with Fig. 14. Such images have substantial potential for treatment planning; for example, they could be used to select optimal sites for the radioactive seeds implanted for cancer therapy. Based on ocular-tumor
results, these images may also delineate regions responding or not responding to therapy, permitting treatment to be progressively adapted in each patient. 4.3. Breast examinations Spectral procedures have been employed to examine breast lesions in order to more reliably identify those tumors that are benign and, thus, do not require biopsies. The goal of these studies is to avoid the risks, expenses, and patient anxiety associated with unnecessary negative biopsies, which currently account for 7090% of the one million annual biopsies in the US (Stavros et al., 1995). Progress towards this goal has already been made using subjective interpretation of lesion features in conventional ultrasonic images (Stavros et al., 1995; Advanced Technology Laboratories (ATL), 1994). Both boundary and internal features are evaluated. Images of malignant (cancerous) lesions typically exhibit an irregular shape (often with protruding speculations), low echogenicity with a heterogeneous texture, and a posterior shadow arising from a relatively high attenuation coecient within the lesion. In contrast, benign lesions typically exhibit smooth, welldened boundaries and higher echogenicity with a homogeneous texture. While these diagnostic criteria can help in selecting cases for biopsies, they are inherently subjective and dependent upon physician skill as well as instrument characteristics. We conducted an initial study to explore whether spectral procedures can be used to quantify several features corresponding to the subjective features now being employed (Alam et al., 2000), our ultimate goal is to replace current assessments with objective, user-independent determinations. In order to investigate this concept, we modied our approach to utilizing spectrum analysis and included a set of lesion-boundary features, as described below. In our studies, digital RF data were acquired from patients with mammographically visible breast lesions, scheduled for subsequent biopsy. The data were acquired using a Spectrasonics Imaging Inc. (Wayne, PA) acquisition module
Fig. 14. 3-D image showing capsule of prostate and foci of suspected cancer.
654
interfaced with an ATL (Bothell, WA) Ultramark 9 scanner. An L10-5 linear array (7.5-MHz center frequency) was employed at a single transmit focal length. Data were sampled (14 bits) at a 20 MHz rate, and time gain control (TGC) data were acquired so that RF data could be corrected for TGC settings prior to analysis. Calibration data were acquired from planar and diuse targets and calibrated spectra were computed over the 59 MHz bandwidth with L 2:5 mm. The initial step in data analysis was the generation of B-mode images from captured RF data. As shown in Fig. 15, the boundary of the lesion was manually traced on these images, and additional rectangular analysis regions were placed about the lesion. Parameter images of spectral intercept and midband t were generated and stored for analysis using these demarcated regions. To evaluate the interior lesion properties, we dened quantitative measures corresponding to the subjective features used by clinicians. Echogenicity was dened as the average intercept value within the traced lesion boundary: intercept was selected since it does not require attenuation compensation. Heterogeneity was dened as the standard deviation of midband t within the lesion, after attenuation compensation (1 dB/ MHz cm): this parameter was selected because the estimator for midband t has high precision, and
the standard deviation for homogeneous scatterers 1=2 is simply 5:6=BL (dB). Shadowing measurements estimated the lesions attenuation coefcient by computing average midband-t values in comparable shadowed and unshadowed analysis regions posterior to the lesion; these measurements were normalized by lesion thickness and provide an attenuation coecient relative to that of surrounding normal tissue if the analyzed posterior segments contain similar structures. We also determined several morphometric features, which were computed from the traced lesion boundary. These included the aspect ratio (vertical-to-horizontal ratio), a fractal dimension characterizing boundary roughness, and convexity. Convexity is dened as the ratio of the perimeter of the convex polygon bounding the lesion contour to the perimeter of the lesion; this parameter decreases with increasing spiculation and has a maximum value of unity. These features were computed for 119 patients with biopsy-proven lesions. Nine spectral and morphometric features were employed to dierentiate malignant from benign lesions using jackknife procedures with nearest-neighbor classiers. While the optimal feature set remains to be identied, results were very encouraging producing an ROC area of 0:87 0:04. This breast study indicates how spectral procedures may be used to quantify subjective criteria now used by clinicians. It also shows how spectra, which measure ultrasonic scattering properties, may be complemented by quantitative boundary features, which are aected by tumor growth patterns. These studies are being expanded to improve upon the denition of each feature and to identify optimally ecient sets of parameters for objective classication. More recent analysis using rened procedures yielded an ROC area of 0:9164 0:0346 (Alam et al., 2002a,b). 4.4. Cardiac examinations An extensive series of cardiac investigations (ODonnell et al., 1979, 1981; Vered et al., 1987; Perez et al., 1994) has been conducted by Miller and associates. While these investigations have examined various spectral parameters (e.g., fre-
Fig. 15. Analysis regions for breast lesion L.
655
quency dependence of backscatter), they have concentrated on the use of IB. As noted above, IB measures the area under measured spectra (in dB) and is directly related to spectral midband t. IB values have been shown to be aected by cardiac ischemia, angulation of cardiac muscle bers, and degree of muscle contraction during the cardiac cycle. IB values, like midband-t values, are aected by intervening attenuation, so that absolute IB values are dicult to employ in standard (closed chest) examinations of the heart. However, the above investigations have shown that temporal variations in IB may be used to evaluate the status of the myocardium. In particular, the amplitude and temporal pattern of IB variations during the cardiac cycle is aected by ischemia. Overall, clinical examinations measuring IB may be of signicant use in echocardiographic examinations of infarction and reperfusion of the myocardium. Real-time cross-sectional images of IB values could be especially useful in these examinations (Perez et al., 1994). 4.5. Additional frequency-domain investigations Signicant frequency-domain results have been obtained in other investigations. Insana, Hall and coworkers have employed a form-factor model to analyze scattering, and they have generated images of IB and scatterer size in the kidney (Insana et al., 1991; Garra et al., 1994). Their animal studies have shown how spectra can be analyzed to separately study kidney microstructure in terms of spherical glomeruli and cylindrical tubules. This approach may be extremely useful in diagnosing kidney disease and elucidating its development in vivo. Several investigators (Oosterveld et al., 1991; King et al., 1985) have shown that sets of spectral parameters can provide diagnosis of diuse and focal diseases of the liver. Vascular plaque has also been characterized by Sigel and coworkers (Lee et al., 1999) and by Spencer et al. (1997). The frequency dependency of backscatter from blood has been extensively examined by Kuo and Shung (1994). Recent in vitro studies (Kolios et al., 1999) have also shown that cellular level phenomena (e.g., apoptosis in drug-induced cell
death) cause measurable changes in spectral parameters, supporting the use of spectral assays for treatment monitoring. Spectral studies of ultrasonic contrast agents have been conducted in our laboratories (Deng et al., 1998, 2000). These small (e.g., 3 lm) particles often consist of encapsulated gas bubbles which are injected in the bloodstream to enhance the ultrasonic detection of blood in the body. We have adapted our theoretical framework to include scattering from these agents and shown that, at suciently high frequencies, spectral intercept can provide quantitative estimates of their in vivo concentrations and radii. Procedures incorporating this nding are being investigated for quantitative evaluations of ow and perfusion in the body.
5. Summary and conclusion Frequency-domain analysis of ultrasonic echoes has been examined by a substantial number of investigators. Spectral techniques oer the potential for new improvements in medical imaging, but their full potential for conveying relevant information and for use in tissue segmentation have yet to be fully realized. There now exists a coherent framework for understanding how spectral features are related to tissue microstructure and for evaluating the statistics of spectral estimators in terms of system characteristics and processing parameters. Several topics warrant further investigation. Most theoretical models assume a specic isotropic function to characterize tissue microstructure. Investigations using acoustic microscopy may help improve these models and could support initial studies of 2-D and 3-D spectra, which incorporate coherent scanning of tissue structures. Calibration procedures for complex multifocus arrays require further investigation as do the spectral eects of non-linear propagation and beam aberrations caused by uctuations in propagation velocity. Additional clinical studies are required to validate the utility of spectral procedures in specic applications, and long-term studies are needed to identify the incremental benets of treatment
656
F.L. Lizzi et al. / Pattern Recognition Letters 24 (2003) 637658 multi-feature analysis procedure. In: Insana, M.F., Walker, W. (Eds.), Medical Imaging 2002: Ultrasonic Imaging and Signal Processing. Society of Photo-Optical Instrumentation Engineers, Vol. 4687, Bellingham, WA, pp. 296303. Alam, S.K., Lizzi, F.L., Feleppa, E.J., Liu, T., Kalisz, A., 2002b. Multi-feature analysis for automated breast lesion classication from ultrasonic data. In: Moxon, K., ElSherif, D., Kanakasabai, S. (Eds.), Proceedings of the IEEE 28th Northeast Bioengineering Conference. Institute of Electrical and Electronics Engineers, pp. 287288. Bamber, J.C., Cook-Marten, G., 1987. Texture analysis and speckle reduction in medical echography. SPIE Intl. Symp. Pattern Recog. Acous. Imaging 768, 120127. Boote, E.J., Zagzebski, J.A., Madsen, E.L., Hall, T.J., 1988. Instrument independent acoustic backscatter imaging. Ultrason. Imaging 10, 121139. Chaturvedi, P., Insana, M.F., 1996. Error bounds on ultrasound scatterer size estimates. J. Acoust. Soc. Am. 100, 392399. Coleman, D.J., Lizzi, F.L., Jack, R.L., 1977. Ultrasonography of the Eye and Orbit, Chapter V. Lea & Febiger, Philadelphia. Coleman, D.J., Lizzi, F.L., Silverman, R.H., Rondeau, M.J., Smith, M.E., Torpey, J.H., 1983. Acoustic biopsy as a means for characterization of intraocular tumors. In: Henkind, P. (Ed.), ACTA: XXIV International Congress of Ophthalmology, Lippincott, Philadelphia, pp. 115118. Coleman, D.J., Silverman, R.H., Rondeau, M.J., Coleman, J.A., Rosberger, D., Ellsworth, R.M., Lizzi, F.L., 1991. Ultrasonic tissue characterization of uveal melanoma and prediction of patient survival after enucleation and brachytherapy. Am. J. Ophthalmol. 112, 682688. Coleman, D.J., Silverman, R.H., Rondeau, M.J., Lizzi, F.L., McLean, I.W., Jakobiec, F.J., 1990. Correlation of acoustic tissue typing of malignant melanoma and histopathologic features as a predictor of death. Am. J. Ophthalmol. 110, 380388. Deng, C.X., Lizzi, F.L., Kalisz, A., Rosado, A., Silverman, R.H., Coleman, D.J., 2000. Study of ultrasonic contrast agents using a dual-frequency band technique. Ultrasound Med. Biol. 26 (5), 819831. Deng, C.X., Lizzi, F.L., Silverman, R.H., Ursea, R., Coleman, D.J., 1998. Imaging and spectrum analysis of contrast agents in the in vivo rabbit eye using very high frequency ultrasound. Ultrasound Med. Biol. 24 (3), 383394. Feleppa, E.J., Lizzi, F.L., 1993. Ophthalmological tissue characterization by scattering. In: Shung, K., Theime, G. (Eds.), Ultrasonic Scattering in Biological Tissues. CRC Press, Boca Raton, pp. 393408. Feleppa, E.J., Fair, W.R., Liu, T., Kalisz, A., Balaji, K.C., Porter, C.R., Tsai, H., Reuter, V., Gnadt, W., Miltner, M.J., 2000. Three-dimensional ultrasound analyses of the prostate. Mol. Urol. 4 (3), 133141. Feleppa, E.J., Fair, W.R., Kalisz, A., Sokil-Melgar, J.B., Lizzi, F.L., Rosado, A., Shao, M.C., Liu, T., Wang, Y., Cookson, M., Reuter, V., 1996. Typing of prostate tissue by ultrasonic spectrum analysis. IEEE Trans. Ultrason. Ferroelect. Freq. Contr. 43 (4), 609619.
monitoring aorded by these techniques. The application of independent sets of spectral parameter images in segmentation and boundary detection clearly warrants further investigation. These topics are assuming increasing importance for automated volume computations, morphology descriptors, and 3-D imaging. Spectral imaging techniques are likely to assume more signicance in view of demands for precise tissue data imposed by the variety of emerging non-invasive and minimally invasive procedures for therapy. 3-D spectral images could be of extreme importance in planning the use of techniques including high-intensity focused ultrasound (ter Haar, 1995; Lizzi et al., 2001), and they could become a key tool for monitoring the acute and chronic eects of such procedures.
Acknowledgements Portions of this research were supported by NIH Grants EY01212, CA53561, and HL59302 and by US Army Medical Research and Materiel Command grant DAMD17-98-1-8331. We wish to gratefully acknowledge the dedicated collaboration of Drs. D.J. Coleman and R.H. Silverman at the Weill Medical College of Cornell University, Dr. W.R. Fair, formerly at the Memorial SloanKettering Cancer Center, Dr. C.R. Porter at the Washington, DC Veterans 30. Aairs Medical Center and their stas, as well as our colleagues at Riverside Research Institute. References
Abbott, J.G., Thurstone, F.L., 1979. Acoustic speckle: theory and experimental analysis. Ultrason. Imaging 1, 303324. Advanced Technology Laboratories, 1994. Summary of ATLSponsored Breast Ultrasound Study and Preliminary Findings. ATL publications. Alam, S.K., Lizzi, F.L., Feleppa, E.J., Liu, T., Kalisz, A., 2000. Ultrasonic multifeature analysis procedures for breast lesion classication. In: Shung, K.K., Insana, M. (Eds.), Medical Imaging 2000: Ultrasonic Imaging and Signal Processing. Society of Photo-Optical Instrumentation Engineers, Vol. 3982, Bellingham, pp. 196201. Alam, S.K., Lizzi, F.L., Feleppa, E.J., Liu, T., Kalisz, A., 2002a. Computer aided diagnosis of breast lesions using a
F.L. Lizzi et al. / Pattern Recognition Letters 24 (2003) 637658 Feleppa, E.J., Liu, T., Kalisz, A., Shao, M.C., Fair, W.R., Fleshner, N., Reuter, V., 1997. Ultrasonic spectral-parameter imaging of the prostate. Int. J. Imaging Sys. Tech. 8, 1125. Feleppa, E.J., Lizzi, F.L., Coleman, D.J., 1988. Ultrasonic analysis for ocular tumor characterization and therapy assessment. News Physiol. Sci. 3, 193197. Feleppa, E.J., Lizzi, F.L., Coleman, D.J., Yaremko, M.M., 1986. Diagnostic spectrum analysis in ophthalmology: a physical perspective. Ultrasound Med. Biol. 12 (8), 623631. Feleppa, E.J., Urban, S., Kalisz A., Lizzi, F.L., Fair, W.R., Porter, C.R., 2001. Advanced ultrasonic imaging of the prostate for guiding biopsies and for planning and monitoring therapy. In: Proc. of the 2000 IEEE Ultrasonics Symposium, Institute of Electrical and Electronics Engineers, pp. 13991403. Fellingham, L.L., Sommer, F.G., 1984. Ultrasonic characterization of tissue structure in the in vivo human liver and spleen. IEEE Trans. Sonics Ultras SU-31, 418428. Garra, B.S., Insana, M.F., Sesterhenn, I.A., Hall, T.J., Wagner, R.F., Rotellar, C., Winchester, J., Zeman, R.K., 1994. Quantitative ultrasonic detection of parenchymal structural change in diuse renal disease. Invest. Radiol. 29 (2), 134 140. Goldberg, B.B., Liu, J., Fosberg, F., 1994. Ultrasound contrast agents:a review. Ultrasound Med. Biol. 20, 319333. Huisman, H.J., Thijssen, J.M., 1996. Precision and accuracy of acoustospectrograhic parameters. Ultrasound Med. Biol. 22, 855871. Insana, M.F., Brown, D.G., 1993. Acoustic scattering theory applied to soft biological tissues. In: Shung, K.K., Thieme, G.A. (Eds.), Ultrasonic Scattering in Biological Tissues. CRC Press, Boca Raton, pp. 75124. Insana, M.F., Hall, T.J., Cook, L.T., 1994. Backscatter coecient estimation using array transducers. IEEE Trans. Ultrason. Ferroel. Freq. Contr. 41 (5), 714723. Insana, M.F., Hall, T.J., Fishback, J.L., 1991. Identifying acoustic scattering sources in normal renal parenchyma from the anistrophy in acoustic properties. Ultrasound Med. Biol. 17, 613626. King, D.L., Lizzi, F.L., Feleppa, E.J., Wai, P.M., Yaremko, M.M., Rorke, M.C., Herbst, J., 1985. Focal and diuse liver disease studied by quantitative microstructural sonography. Radiology 155, 457462. Kolios, M.C., Czarnota, G.J., Lee, M., Hunt, J.W., Sherar, M.D., 1999. Ultrasonic spectrum analysis of apoptotic cell populations. Fed. Am. Soc. Exp. Biol. J. 13 (7), A1435 A1435, Abstract. Kremkau, F.W., 1990. Doppler Ultrasound: Principles and Instruments. W.B. Saunders, Philadelphia. Kuc, R., Schwartz, M., 1979. Estimating the acoustic attenuation coecient slope for liver from reected ultrasound signals. IEEE Trans. Sonics Ultras. 26, 353362. Kuo, I.Y., Shung, K.K., 1994. High-frequency ultrasonic backscatter from erythrocyte suspensions. IEEE Trans. Biomed. Engr. 41, 2933.
657
Lee, D.J., Sigel, B., Swami, V.K., Justin, J.R., Gahtan, V., OBrien, S.P., Dwyer-Joyce, L., Feleppa, E.J., Roberts, A.B., Berkowitz, H.D., 1999. Determination of carotidplaque risk by ultrasonic tissue characterization. Ultrasound Med. Biol. 24 (9), 12911299. Lizzi, F.L., 1998. Ultrasonic scatterer property images of the eye and prostate. In: Proc. of the 1997 IEEE Ultrasonics Symposium, Institute of Electrical and Electronics Engineers, pp. 11091117. Lizzi, F.L., Feleppa, E.F., 2000. Image processing and preprocessing for medical ultrasound. In: Aanstoos, J.V. (Ed.), Proceedings of the 29th Applied Imagery Pattern Recognition Workshop (AIPR2000). IEEE Computer Society, Los Alamitos, CA, pp. 187192. Lizzi, F.L., Laviola, M.A., 1975. Power spectra measurements of ultrasonic backscatter from ocular tissue. In: Proc. of the IEEE Ultrasonics Symposium 1975. Institute of Electrical and Electronics Engineers, pp. 2932. Lizzi, F.L., Astor, M., Liu, T., Deng, C., Coleman, D.J., Silverman, R.H., 1997a. Ultrasonic spectrum analysis for tissue assays and therapy evaluation. Int. J. Imaging Sys. Tech. 8, 310. Lizzi, F.L., Astor, M., Feleppa, E.J., Shao, M., Kalisz, A., 1997c. Statistical framework for ultrasonic spectral parameter imaging. Ultrasound Med. Biol. 23 (9), 13711382. Lizzi, F.L., Deng, C.X., Alam, S.K., Silverman, R.H., Rondeau, M., Coleman, D.J., 2001.Ocular tumor treatments with focused ultrasound: eects of beam geometry, tissue morphology, and adjacent tissues. In: Proc. of the 2000 IEEE Ultrasonics Symposium. Institute of Electrical and Electronics Engineers, pp. 12991301. Lizzi, F.L., Feleppa, E.J., Astor, M., Kalisz, A., 1997b. Statistics of ultrasonic spectral parameters for prostate and liver examinations. IEEE Trans. Ultrason. Ferroelect. Freq. Contr. 44, 935942. Lizzi, F.L., Feleppa, E.J., Coleman, D.J., 1986. Ultrasonic ocular tissue characterization. In: Greenleaf, J.F. (Ed.), Characterization of Tissue with Ultrasound. CRC Press, Boca Raton, pp. 4160. Lizzi, F.L., Feleppa, E.J., Yaremko, N., King, D.L. Wai, P., 1981. Liver-tissue characterization by digital spectrum and cepstrum analysis. In: Proc. of the IEEE Ultrasonics Symposium, Institute of Electrical and Electronics Engineers, pp. 575578. Lizzi, F.L., Greenebaum, M., Feleppa, E.J., Elbaum, M., Coleman, D.J., 1983. Theoretical framework for spectrum analysis in ultrasonic tissue characterization. J. Acoust. Soc. Am. 73 (4), 13661373. Lizzi, F.L., Ostromogilsky, M., Feleppa, E., Rorke, M.C., Yaremko, M.M., 1987. Relationship of ultrasonic spectral parameters to features of tissue microstructure. IEEE Trans. Ultras. Ferroel. Freq. Contr. 34, 319329. Lizzi, F.L., Rorke, M.C., Sokil-Melgar, J.B., Kalisz, A., Driller, J., 1992. Interfacing very-high-frequency transducers to digital-acquisition scanning systems. Proc. Soc. PhotoOptical Instrum. Eng. (SPIE) 1844, 313321.
658
F.L. Lizzi et al. / Pattern Recognition Letters 24 (2003) 637658 with microcirculatory patterns in uveal melanomas. Ultrasound Med. Biol. 23 (4), 573581. Silverman, R.H., Rondeau, M.J., Lizzi, F.L., Coleman, D.J., 1995. Three-dimensional high-frequency ultrasonic parameter imaging of anterior segment pathology. Ophthalmology 102, 837843. Spencer, T., Ramo, M.P., Salter, D.M., Anderson, T., Kearney, P.P., Sutherland, G.R., Fox, K.A.A., McDicken, W.N., 1997. Characteristics of atherosclerotic plaque by spectral analysis of intravascular ultrasound: an in vitro methodology. Ultrasound Med Biol. 23 (2), 191203. Stavros, A.T., Thickman, D., Rapp, C.L., Dennis, M.A., Parker, S.H., Sisney, G.A., 1995. Solid breast nodules: use of sonography to distinguish between benign and malignant lesions. Radiology 196, 123134. Tateishi, T., Machi, J., Feleppa, E.J., Oishi, R.H., Jucha, J., Yanagihara, E., McCarthy, L.J., Noritomi, T., Shirouzu, K., 1998. In vitro diagnosis of axillary lymph node metastases in breast cancer by spectrum analysis of radio frequency echo signals. Ultrasound Med. Biol. 24 (8), 1151 1159. ter Haar, G., 1995. Ultrasound focal beam surgery. Ultrasound Med.Biol. 21, 10891100. Tuthill, T.A., Sperry, R.H., Parker, K.J., 1988. Deviations for Rayleigh statistics in ultrasonic speckle. Ultrason. Imaging 10, 8189. Vered, Z., Barzilai, B., Mohr, G.A., Thomas III, L.J., Genton, R., Sobel, B., Shoup, T.A., Melton, H.E., Miller, J.G., Perez, J.E., 1987. Quantitative ultrasonic tissue characterization with real-time integrated backscatter imaging in normal human subjects and in patients with dilated cardiomyopathy. Circulation 76, 10671073. Waag, R.C., Astheimer, R., 1993. Measurement System Eects in Ultrasonic Scattering Experiments. In: Shung, K.K., Theime, G. (Eds.), Ultrasonic Scattering and Biological Tissues. CRC Press, Boca Raton, pp. 251290. Wagner, R.F., Smith, S.W., Sandrik, J.M., Lopez, R., 1983. Statistics of speckle in ultrasound B-scans. IEEE Trans. Ultrasound 30, 156163. Wear, K., Wagner, R., Garra, B., 1994. High resolution ultrasonic backscatter coecient estimation based on autoregressive spectral estimation using Burgs algorithms. IEEE Trans. Med. Imaging 13, 500507.
Lizzi, F.L., St. Louis, L., Coleman, D.J., 1976. Applications of spectral analysis in medical ultrasonography. Ultrasonics 14 (2), 7780. Magnin, P.A., von Ramm, O.T., Thurstone, F.L., 1982. Frequency compounding for speckle contrast reduction in phased array images. Ultrason. Imaging 4 (3), 267 281. Namery, J., Lele, P.P., 1972. Ultrasonic detection of myocardial infarction in dog. In: Proc. of the IEEE Ultrasonics Symposium. Institute of Electrical and Electronics Engineers, pp. 491494. ODonnell, M., Bauwens, D., Mimbs, J.W., Miller, J.G., 1979. Broadband integrated backscatter: an approach to spatially localized tissue characterization in vivo. In: Proc. of the IEEE Ultrasonics Symposium. Institute of Electrical and Electronics Engineers, pp. 175178. ODonnell, M., Mimbs, J.W., Miller, J.G., 1981. Relationship between collagen and ultrasonic backscatter in myocardial tissue. J. Acoust. Soc. Am. 69 (2), 580588. Oosterveld, B.J., Thijssen, J.M., Hartman, P.C., Romijnm, R.L., Rosenbusch, G.J.E., 1991. Ultrasound attenuation and texture analysis of diuse liver disease: methods and preliminary results. Phys. Med. Biol. 36, 10391064. Perez, J.E., Miller, J.G., Holland, M.R., Wickline, S.A., Waggoner, A.D., Barzilai, B., Sobel, B.E., 1994. Ultrasonic tissue characterization: integrated backscatter imaging for detecting myocardial structural properties and on-line quantitation of cardiac function. Am. J. Card. Imaging 8 (2), 106112. Purnell, E.W., Sokollu, A., Holasek, E., Cappeart, W., 1975. Clinical spectracolor ultrasonography. J. Clin. Ultras. 3, 187189. Sherar, M.D., Foster, F.S., 1989. The design and fabrication of high frequency polyvinylidene uoride transducers. Ultrason. Imaging 11, 7594. Silverman, R.H., Coleman, D.J., Lizzi, F.L., Torpey, J.H., Driller, J., Iwamoto, T., Burgess, S.E.P., Rosado, A., 1986. Ultrasonic tissue characterization and histopathology in tumor xenografts following ultrasonically induced hyperthermia. Ultrasound Med. Biol. 12 (8), 639645. Silverman, R.H., Folberg, R., Boldt, H.C., Lloyd, H.O., Rondeau, M.J., Mehaey, M.G., Lizzi, F.L., Coleman, D.J., 1997. Correlation of ultrasound parameter imaging

Spectrum Tissue

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Spectrum Tissue

Uploaded by

Copyright:

Available Formats

Pattern Recognition Letters 24 (2003) 637658 www.elsevier.

Ultrasonic spectrum analysis for tissue evaluation

F.L. Lizzi et al. / Pattern Recognition Letters 24 (2003) 637658

F.L. Lizzi et al. / Pattern Recognition Letters 24 (2003) 637658

F.L. Lizzi et al. / Pattern Recognition Letters 24 (2003) 637658

F.L. Lizzi et al. / Pattern Recognition Letters 24 (2003) 637658

Fig. 2. RF echo signals from prostate.

Fig. 1. B-mode image of prostate.

F.L. Lizzi et al. / Pattern Recognition Letters 24 (2003) 637658

Fig. 3. RF tissue spectrum and calibration spectrum.

F.L. Lizzi et al. / Pattern Recognition Letters 24 (2003) 637658

Fig. 4. Calibrated power spectrum.

F.L. Lizzi et al. / Pattern Recognition Letters 24 (2003) 637658

Fig. 5. Spectral parameter images of prostate: (a) midband t; (b) intercept.

F.L. Lizzi et al. / Pattern Recognition Letters 24 (2003) 637658

Fig. 6. Stained image highlighting suspected prostate cancer.

F.L. Lizzi et al. / Pattern Recognition Letters 24 (2003) 637658

F.L. Lizzi et al. / Pattern Recognition Letters 24 (2003) 637658

F.L. Lizzi et al. / Pattern Recognition Letters 24 (2003) 637658

Fig. 8. Probability density functions for midband t.

F.L. Lizzi et al. / Pattern Recognition Letters 24 (2003) 637658

F.L. Lizzi et al. / Pattern Recognition Letters 24 (2003) 637658

F.L. Lizzi et al. / Pattern Recognition Letters 24 (2003) 637658

Fig. 12. 3-D excised image showing anterior ocular melanoma.

F.L. Lizzi et al. / Pattern Recognition Letters 24 (2003) 637658

Fig. 13. ROC curves for prostate-cancer classication.

F.L. Lizzi et al. / Pattern Recognition Letters 24 (2003) 637658

F.L. Lizzi et al. / Pattern Recognition Letters 24 (2003) 637658

Fig. 15. Analysis regions for breast lesion L.

F.L. Lizzi et al. / Pattern Recognition Letters 24 (2003) 637658

You might also like