Reconstruction strategy for echo planar spectroscopy and its application to partially undersampled imaging
2000; Wiley; Volume: 44; Issue: 3 Linguagem: Inglês
10.1002/1522-2594(200009)44
ISSN1522-2594
AutoresLars G. Hanson, Kjeld Schaumburg, Olaf B. Paulson,
Tópico(s)NMR spectroscopy and applications
ResumoMagnetic Resonance in MedicineVolume 44, Issue 3 p. 412-417 Full PaperFree Access Reconstruction strategy for echo planar spectroscopy and its application to partially undersampled imaging Lars G. Hanson, Corresponding Author Lars G. Hanson larsh@magnet.drcmr.dk Danish Research Center for Magnetic Resonance, Hvidovre Hospital, Copenhagen, Denmark CISMI, Department of Chemistry, University of Copenhagen, Copenhagen, DenmarkDanish Research Center for Magnetic Resonance, Hvidovre Hospital, afs. 340, Kettegård Allé 30, DK-2650 Hvidovre, Denmark===Search for more papers by this authorKjeld Schaumburg, Kjeld Schaumburg CISMI, Department of Chemistry, University of Copenhagen, Copenhagen, DenmarkSearch for more papers by this authorOlaf B. Paulson, Olaf B. Paulson Danish Research Center for Magnetic Resonance, Hvidovre Hospital, Copenhagen, DenmarkSearch for more papers by this author Lars G. Hanson, Corresponding Author Lars G. Hanson larsh@magnet.drcmr.dk Danish Research Center for Magnetic Resonance, Hvidovre Hospital, Copenhagen, Denmark CISMI, Department of Chemistry, University of Copenhagen, Copenhagen, DenmarkDanish Research Center for Magnetic Resonance, Hvidovre Hospital, afs. 340, Kettegård Allé 30, DK-2650 Hvidovre, Denmark===Search for more papers by this authorKjeld Schaumburg, Kjeld Schaumburg CISMI, Department of Chemistry, University of Copenhagen, Copenhagen, DenmarkSearch for more papers by this authorOlaf B. Paulson, Olaf B. Paulson Danish Research Center for Magnetic Resonance, Hvidovre Hospital, Copenhagen, DenmarkSearch for more papers by this author First published: 01 September 2000 https://doi.org/10.1002/1522-2594(200009)44:3 3.0.CO;2-PCitations: 16AboutSectionsPDF ToolsRequest permissionExport citationAdd to favoritesTrack citation ShareShare Give accessShare full text accessShare full-text accessPlease review our Terms and Conditions of Use and check box below to share full-text version of article.I have read and accept the Wiley Online Library Terms and Conditions of UseShareable LinkUse the link below to share a full-text version of this article with your friends and colleagues. Learn more.Copy URL Share a linkShare onFacebookTwitterLinked InRedditWechat Abstract The most commonly encountered form of echo planar spectroscopy involves oscillating gradients in one spatial dimension during readout. Data are consequently not sampled on a Cartesian grid. A fast gridding algorithm applicable to this particular situation is presented. The method is optimal, i.e., it performs as well as the full discrete Fourier transform for band limited signals while allowing for use of the fast Fourier transform. The method is demonstrated for reconstruction of data that are partially undersampled in the time domain. The advantages of undersampling are lower hardware requirements or fewer interleaves per acquisition. The method is of particular interest when large bandwidths are needed (e.g., for high field scanning) and for scanners with limited gradient performance. The unavoidable artifacts resulting from undersampling are demonstrated to be acceptable for spectroscopy with long echo times. Magn Reson Med 44:412–417, 2000. © 2000 Wiley-Liss, Inc. Echo planar spectroscopy (EPS), as proposed by Mansfield (1), offers rapidly acquired data for spatially resolved spectroscopic studies (2-4). Several variants have been implemented (5-11). The common feature is that data are acquired in the presence of an oscillating readout gradient, while differences occur in methods for water and lipid suppression, spatial selection, excitation, and gradient schemes. This paper addresses aspects of sequence design and reconstruction for a simple EPS scheme involving an oscillating gradient in just one direction, and phase encoding in the remaining spatial dimensions. Data are sampled non-equidistantly in (k,t)-space, as shown in Fig. 1, and artifacts must be traded off for reconstruction time and computational complexity. The simplest approach is to ignore spin evolution not caused by imaging gradients in the readout periods. This allows for fast reconstruction using the fast Fourier transform (FFT) algorithm, and it works particularly well with spectral oversampling (8, 12). Alternatively, reconstruction can be done using discrete Fourier transform (DFT) directly. The optimal performance of the DFT is obtained at the expense of reconstruction time (O(N4) for 2D-DFT of N2 points versus O(N2 log2 N) for 2D-FFT). Another approach is gridding, which represents a flexible compromise. Figure 1Open in figure viewerPowerPoint Sampling strategy in (k,t)-space, where k is the spatial frequency and t is time. The trajectory is shown for two periods of readout gradient oscillation. The time evolution of the maximum frequency, kmax, is effectively only probed once every gradient oscillation period 2T, while the zero frequency is sampled twice as often. We present a gridding algorithm tailored for the problem at hand. It takes advantage of the localization of the sample points on straight lines in (k,t)-space (13, 14). The merits of the method are speed and optimal performance, i.e., performance as good as that of the full DFT. The method is demonstrated for undersampled spectroscopy, in which the gradient oscillation period 2T (Fig. 1) exceeds the sample interval required for resolving all metabolite signals. Undersampling minimizes hardware and interleaving requirements, and provides a fast alternative to oversampling (1). The article is organized as follows: The spectroscopic sequence relying on undersampling is described. General properties of the reconstruction are discussed, as well as its application to adequately sampled and undersampled data. Finally, examples of in vitro and in vivo spectra are presented. METHODS Spectroscopic Sequence A multi-slice EPS sequence was implemented on a Siemens Vision 1.5 T whole-body scanner. The sequence is conceptually similar to implementations by other groups (5-8). The main difference is a low acquisition bandwidth and non-interleaved sampling. Each cycle consists of outer volume and water suppression, slice-selective excitation and refocusing, phase encoding followed by data-acquisition in the presence of an oscillating readout gradient. This is repeated for each slice and each line of k-space. The oscillating readout gradient is initiated one cycle ahead of actual sampling to reduce effects of transient eddy currents. Each half-cycle consists of 0.2 msec sinusoidal gradient inversion and a 3.2 msec readout period with constant gradient. One dimension in k-space is consequently traversed 64 times during a 217.6 msec readout period, giving a spectral range of 294 Hz (4.6 ppm at 1.5 Tesla). Post-processing Reconstruction A spatial shift appears when the data from a particular gradient echo are reconstructed without accounting for spin evolution not caused by imaging gradients in the readout periods. The sign of the shift depends on the polarity of the readout gradient that differs between even and odd echoes. The bandwidth per pixel was 312 Hz for the implemented sequence corresponding to a shift of half a pixel of the N-acetylaspartate (NAA) signal relative to water. A reconstruction was developed to avoid the corresponding spatial broadening. This discussion is limited to the relevant spatial (readout) and frequency dimensions (x,ν) and the corresponding reciprocal space (k,t) in which data are collected. These quantities are all taken to be dimensionless, so x and ν are given in units of the spatial and spectral resolution, respectively. The correction for spatial shifts of metabolites amounts to an interpolation of data in (k,t)-space to a grid. Such an interpolation can be done perfectly if the Nyquist criterion is fulfilled. Artifact-free reconstruction can therefore be performed to a bandwidth equal to the minimum sampling rate of the signal (1/(2T) in Fig. 1) if the signal is equally band limited. Reconstruction is conveniently performed on even and odd echoes separately, since samples are equidistantly collected along straight lines in (k,t)-space, as shown in Fig. 1. A correction for these lines not being parallel to the t = 0 line can be done using the Fourier shift theorem. The shift corresponding to the frequency ν is ν/b, where b is the acquisition bandwidth per pixel given by the inverse of the readout time RO, b = 1/RO. Corrected signal matrices even and odd can be constructed from the even and odd echo signal matrices, Seven(k,ν) and Sodd(k,ν) according to (1) (2) Two applications of these equations are of interest: water reference reconstruction and reconstruction of oversampled metabolite data. After shimming, protons in water precess at a frequency range of a few Hertz around the carrier frequency being adjusted to the water resonance. Half bandwidth reconstruction is therefore sufficient. Signals from unsuppressed lipids can be filtered out, and spatial reconstruction completing the half-bandwidth reconstruction is rapidly performed using FFT. The other, more interesting use of Eqs. [1] and [2] is for metabolite reconstruction, in which spectral oversampling is employed, and half bandwidth reconstruction is enough for resolving all metabolite signals. The appropriate frequency range in this case is [−1/(2T);0] since the metabolites are known to oscillate at negative frequencies relative to water. Any residual water signal must be removed before applying the correction to avoid large spatial shifts of positive low-frequency components. The hardware demands can be reduced without increasing the number of k-space interleaves if sampling is performed with the readout gradient oscillating at only half the Nyquist frequency. In this case, half bandwidth reconstruction provided by Eqs. [1] and [2] is insufficient. In order to do full spatial and spectral bandwidth reconstruction of the metabolite data, the two corrected datasets must be appropriately combined. A corrected signal matrix can be constructed using Danielson-Lanczos lemma (15). This is illustrated in Fig. 2 and is expressed by the equation (3) The signals Seven and Sodd are periodic in frequency ν with period 1/(2T). The exponential factor exp(i2πνT) (period 1/T) accounts for the time shifts of the odd echoes with respect to even echoes. Consequently (k,ν) provides spectral information over the full frequency range, 1/T. This corresponds to optimal gridding in the (k,t) domain, and the remaining steps of the reconstruction are common to those outlined for half bandwidth reconstruction. The reconstruction Eq. [3] is exact if the Nyquist criterion is fulfilled for all frequencies. Figure 2Open in figure viewerPowerPoint Schematic overview of the reconstruction algorithm illustrating the spectral-spatial signal densities resulting from an oscillating point source. a,b: A signal with a single frequency component ν0 originating from position x0 (⊗) gives rise to two spatially displaced peaks (○) in the frequency range [0; −1/T] after reconstruction of odd echo images due to the periodicity of the Fourier transform. The peaks are displaced in the opposite direction for the even echo (×). c: The proposed algorithm shifts both sets of peaks according to their frequency. The true peaks overlap at the original position, while the aliased ones are shifted too much to cancel. d: The shifts are too small to cause ghost cancellation for a point source oscillating at a frequency in the interval [−1/(2T); −1/T]. Water Referencing and Subtraction Water referencing was performed pixel-wise in the reconstructed data sets. The water-suppressed metabolite signal in each voxel was divided by the complex phase of the water signal (16, 17). Any residual water signal must be removed before Eq. [3] is applied: Signals oscillating at slightly higher frequencies than the carrier are spatially far displaced, since the correction is designed to correct metabolite signals that are known to oscillate at negative frequencies. Any residual water peak will suffer severe line shape distortion and spatial broadening. The residual water signal was removed by subtracting the projection onto a space spanned by five vectors, constructed as the water reference signal frequency shifted by a fraction of the water linewidth. The projection was performed using singular value decomposition (15). Experimental Verification Phantom and in vivo experiments were performed for validation of the reconstruction procedure, and for comparison with standard FFT (the conventional approach). Four-slice acquisition with echo and repetition times of 144 msec and 1.5 sec was used for all experiments. For calculating phantom metabolite maps, two water-suppressed measurements and a reference scan were acquired in three min (each of one-min duration). The phantom contained 10 mM solutions of substances spectroscopically similar to the metabolites detectable at long echo times. These were present in tubes embedded in a larger cylinder with an acetone solution (similar to 10 mM NAA). The in vivo data were acquired in 10 min from four axial slices through the brain of a healthy volunteer. Six one-min acquisitions (1 ml isotropic voxels) were obtained, along with six additional 30-sec acquisitions of the central 50% of k-space (full range in the readout direction, half range in the phase-encoding direction). This weighted k-space acquisition scheme improves the spatial impulse response with little sacrifice of efficiency (18, 19). The water reference data set was obtained in one min. RESULTS The spatial and spectral resolutions are closely linked, and two-dimensional point spread functions (PSFs (1)) were calculated. The results of theoretical evaluations of the broadening are illustrated in Fig. 3. A point source is located in the middle of a 16 voxel FOV interpolated to 32 voxels by zero-filling. The source is chosen to oscillate at the NAA frequency with a signal decay time T*2 = 65 msec. This frequency is 37 in units of the bandwidth per point for T = 3.4 msec. A main peak is seen at the metabolite frequency in Fig. 3a and b, as well as spectral ghosts displaced by half the spectral range (32 in these units). Profiles through the PSFs are given in Fig. 3c, d, and e, as well as comparisons to the non-broadened case (the normal sinc function). Figure 3Open in figure viewerPowerPoint Spectral-spatial PSFs for NAA using (a) standard and (b) corrected FFT reconstruction. Details regarding the calculation and the units are given in the results section. The main lobe is seen at the metabolite frequency (37 in units of the bandwidth per pixel), as well as spectral ghosts displaced by half the spectral range (32 in these units). b: Substantial spatial narrowing results from the correction. This graph is conceptually similar to Fig. 2c. The image intensities are illustrated in c,d,e showing uncorrected (dotted) and corrected profiles (solid) through the PSFs using a common vertical scale. Dashed curves (mostly overlapping with corrected profiles (solid) in all graphs but (e)) illustrate the PSF in the absence of artifacts, i.e., in the limit of rapid gradient oscillation. The correction restores the spatial (c) and spectral (d) profiles through the main lobe of the PSF almost entirely, while the spatial profile through the ghosts is nearly unaffected by the correction (e). Figure 4 demonstrates the spatial broadening and how it is suppressed by proper reconstruction. Large differences between images reconstructed from the first even and odd echoes appear close to edges perpendicular to the readout gradient when the carrier frequency is adjusted 200 Hz off-resonance. This was done to make water oscillate at a frequency similar to the offset of lactate relative to water. Figure 4Open in figure viewerPowerPoint Effects of correction for odd/even echo asymmetry. a: Localizer image of water-filled tube phantom with saturation slices covering test tubes containing vegetable oil. b: Corresponding zero-filled water reference image with the carrier adjusted to water. The carrier was subsequently adjusted 200 Hz up to make water oscillate as would lactate. c,d: The resulting absolute difference images between images reconstructed from the first even and odd echoes acquired with opposite gradient polarity. c: Shift artifacts are clearly visible near vertical edges for standard FFT reconstruction. d: The maximum intensity of the edge artifacts is reduced by a factor of four compared to c using Eq. [3], and the image intensity is scaled up accordingly. Metabolite maps resulting from standard FFT and from the corrected reconstruction algorithms of phantom data are shown in Fig. 5. The nature of spectral ghosts is illustrated in Fig. 6, which shows spectral images evaluated at frequencies where ghosts are expected. Figure 5Open in figure viewerPowerPoint In vitro comparison of reconstruction algorithms. Normalized absolute spectral images |S(x,y,ν)| are shown corresponding to NAA, Cho, Cr, and Lac frequencies (the low frequency component of the Lac doublet). The proposed reconstruction algorithm is found to provide sharper contours and better homogeneity than standard FFT reconstruction. The peak intensity in the small tubes are larger in a compared to b by 7%, 4%, and 24% for Cho, Cr, and Lac, respectively (averaged over the widths of the tubes in the phase encoding direction). This is consistent with the expected spatial frequency-dependent narrowing in the readout direction. Figure 6Open in figure viewerPowerPoint Absolute spectral ghost images corresponding to the metabolite images in Fig. 5a calculated as S(x,y,(ν+(2T)−1) modulo T−1, i.e., images at frequencies displaced by half the spectral bandwidth. Low intensity spectral ghosts are visible near vertical edges. These result from undersampling and are common for the two reconstruction techniques. Finally, an in vivo example of a metabolite map and a spectrum from a 1 ml white matter voxel are shown in Fig. 7. Figure 7Open in figure viewerPowerPoint Example spectrum from an isotropic 1 ml voxel located mainly in white matter, as indicated with a box on the inset water reference image (left). The right inset is the measured NAA map in the same slice. This four-slice study was performed in 10 min. DISCUSSION The reconstruction presented offers enhanced spatial resolution compared to standard FFT applied to the non-uniformly sampled data, as shown in Fig. 3. This was demonstrated in theory (Fig. 3) as well as in practice (Figs. 4 and 5). The residual edge artifacts in Fig. 4 can be attributed to the substances other than water in this spectroscopy phantom. Improved spatial homogeneity is found in Fig. 5. This could be a result of unforeseen spatial broadening, rather than a result of accounting for the (k,t)-space trajectory. The observed intensity increase for the narrow tubes is, however, consistent with the expected spatial narrowing, thus showing that the algorithm provides improved overall image quality—including better spatial resolution—as expected. Residual spatial broadening over a length of δx = δν/b proportional to the spectral width δν of the metabolite signal can be expected, considering the nature of the reconstruction illustrated in Fig. 2. The PSFs differ from metabolite to metabolite for uncorrected data due to the chemical shift dependent displacement. The advantage of the improved reconstruction is greater for Lac and NAA compared to Cho and Cr, as can be seen in Fig. 5. Undersampled imaging is a natural candidate for the method, as artifacts from the non-Cartesian sampling are worst for low-bandwidth acquisition. The resulting spectral ghosts appear at positions where there are spatial variations in the metabolite concentrations. The ghosts demonstrated theoretically in Fig. 3 are particularly visible in the spectral image evaluated at the Cr frequency plus half a spectral bandwidth, as shown in Fig. 6 (the SNR is too low to make the feeble ghosts fully visible for the other substances). The magnitudes of the ghosts are largely unaffected by the reconstruction procedure, as can be anticipated from Fig. 2. Ghosts are not necessarily a problem, as demonstrated in Figs. 5 and 6: They are shifted by half the spectral range, which is 4.6 ppm in the present study. This was chosen to avoid overlap of any of the primary metabolites detectable at long echo times (NAA (2.0 ppm), Cr (3.0 ppm), Cho (3.2 ppm), and Lac (1.3 ppm)), and their respective ghosts. A water ghost would appear in the middle of the spectrum (2.4 ppm), which is one reason water signals must be carefully suppressed or removed in post-processing prior to the correction. The residual artifacts were not found to be problematic for evaluation of peak intensities. Finally, inadequately suppressed lipid signals will result in ghosts that overlap with the Cho and Cr signals. Undersampling is problematic if other metabolites are of interest, e.g., for spectroscopy with short echo times. The representative brain spectrum in Fig. 7 illustrates the performance of the undersampling sequence in vivo. High spatial resolution spectra from multiple slices can be obtained rapidly (1 ml resolution, one min per acquisition/10 min total scanning). Finally, it must be noted that standard FFT does not provide an alternative to spatial filtering, even though the side bands are somewhat suppressed compared to the suggested algorithm, as seen in Fig. 3a and b. Parseval's theorem (15) implies that the noise amplitudes are equal for the standard FFT, and the proposed reconstruction and the normal SNR gain caused by filtering is therefore not provided by standard FFT, even though the resulting PSFs are somewhat similar. If spatial filtering is wanted, it is best applied to the corrected data (k,ν). CONCLUSION Inherent spin evolution during signal readout causes spatial broadening if not accounted for with gridding or DFT reconstruction. A fast and optimal gridding algorithm relying on FFT was demonstrated. Like other reconstruction methods, it does not remove ghosts resulting from spectral undersampling, but ghosts were shown not to cause problems for long echo time spectroscopy. The advantages of undersampling are short gradient switching time and low hardware requirements with respect to gradient and eddy current performance. This performance was obtained without interleaved acquisition, which prolongs acquisition times. The main potential of the undersampling strategy is in echo planar spectroscopy using clinical scanners, in which hardware performance is often limited, and in high field scanners, where fast gradient oscillation is needed for resolving all metabolite signals. Acknowledgements The authors thank Elfar Adalsteinsson and Dan Spielman for useful comments regarding the manuscript. REFERENCES 1 Mansfield P. Spatial mapping of the chemical shift in NMR. Magn Reson Med 1984; 1: 370– 386. Wiley Online LibraryCASPubMedWeb of Science®Google Scholar 2 Posse S, Dager SR, Richards TL, Yuan C, Ogg R, Artru AA, Muller-Gartner HW, Hayes C. In vivo measurement of regional brain metabolic response to hyperventilation using magnetic resonance: proton echo planar spectroscopic imaging (PEPSI). Magn Reson Med 1997; 37: 858– 865. Wiley Online LibraryCASPubMedWeb of Science®Google Scholar 3 Heide AC, Kraft GH, Slimp JC, Gardner JC, Posse S, Serafini S, Bowen JD, Richards TL. Cerebral N-acetylaspartate is low in patients with multiple sclerosis and abnormal visual evoked potentials. Am J Neuroradiol 1998; 19: 1047– 1054. CASPubMedWeb of Science®Google Scholar 4 Lim KO, Adalsteinsson E, Spielman D, Sullivan E V, Rosenbloom MJ, Pfefferbaum A. Proton magnetic resonance spectroscopic imaging of cortical gray and white matter in schizophrenia. Arch Gen Psychiatry 1998; 55: 346– 352. CrossrefCASPubMedWeb of Science®Google Scholar 5 Matsui S, Sekihara K, Kohno H. Spatially resolved NMR spectroscopy using phase-modulated spin-echo-trains. J Magn Reson 1986; 67: 476– 490. CrossrefCASWeb of Science®Google Scholar 6 Guilfoyle DN, Blamire A, Chapman B, Ordidge RJ, Mansfield P. PEEP—a rapid chemical-shift imaging method. Magn Reson Med 1989; 10: 282– 287. Wiley Online LibraryCASPubMedWeb of Science®Google Scholar 7 Webb P, Spielman D, Macovski A. A fast spectroscopic imaging method using a blipped phase encode gradient. Magn Reson Med 1989; 12: 306– 315. Wiley Online LibraryCASPubMedWeb of Science®Google Scholar 8 Posse S, DeCarli C, Le Bihan D. Three-dimensional echo-planar MR spectroscopic imaging at short echo times in the human brain. Radiology 1994; 192: 733– 738. CrossrefCASPubMedWeb of Science®Google Scholar 9 Adalsteinsson E, Irarrazabal P, Spielman DM, Macovski A. Three-dimensional spectroscopic imaging with time-varying gradients. Magn Reson Med 1995; 33: 461– 466. Wiley Online LibraryCASPubMedWeb of Science®Google Scholar 10 Hirata S, Bito Y, Yamamoto E. Expansion of the spectral bandwidth by spatial and chemical shift selective saturation in high-speed magnetic resonance spectroscopic imaging. Magn Reson Med 1996; 35: 611– 616. Wiley Online LibraryCASPubMedWeb of Science®Google Scholar 11 Adalsteinsson E, Irarrazabal P, Topp S, Meyer C, Macovski A, Spielman DM. Volumetric spectroscopic imaging with spiral-based k-space trajectories. Magn Reson Med 1998; 39: 889– 898. Wiley Online LibraryCASPubMedWeb of Science®Google Scholar 12 Posse S, Tedeschi G, Risinger R, Ogg R, Le Bihan D. High speed 1H spectroscopic imaging in human brain by echo planar spatial-spectral encoding. Magn Reson Med 1995; 33: 34– 40. Wiley Online LibraryCASPubMedWeb of Science®Google Scholar 13 Brown TR, Kincaid BM, Ugurbil K. NMR chemical shift imaging in three dimensions. Proc Natl Acad Sci USA 1982; 79: 3523– 3526. CrossrefCASPubMedWeb of Science®Google Scholar 14 Twieg DB. The k-trajectory formulation of the NMR imaging process with applications in analysis and synthesis of imaging methods. Med Phys 1983; 10: 610– 621. Wiley Online LibraryCASPubMedWeb of Science®Google Scholar 15 Press W, Teukolsky S, Vetterling W, Flannery B. Numerical recipes in C (the art of scientific computing). Cambridge: Cambridge University Press; 1988. p 60– 68, 408. Google Scholar 16 Spielman D, Webb P, Macovski A. Water referencing for spectroscopic imaging. Magn Reson Med 1989; 12: 38– 49. Wiley Online LibraryCASPubMedWeb of Science®Google Scholar 17 Klose U. In vivo proton spectroscopy in presence of eddy currents. Magn Reson Med 1990; 14: 26– 30. Wiley Online LibraryCASPubMedWeb of Science®Google Scholar 18 Mareci TH, Brooker HR. High-resolution magnetic resonance spectra from a sensitive region defined with pulsed field gradients. J Magn Reson 1984; 57: 157– 163. CrossrefCASWeb of Science®Google Scholar 19 Parker DL, Gullberg GT, Frederick PR. Gibbs artifact removal in magnetic resonance imaging. Med Phys 1987; 14: 640– 645. Wiley Online LibraryCASPubMedWeb of Science®Google Scholar Citing Literature Volume44, Issue3September 2000Pages 412-417 FiguresReferencesRelatedInformation
Referência(s)