A comparison of Raman chemical images produced by univariate and multivariate data processing—a simulation with an example from pharmaceutical practice Slobodan Šašić,*a,b Donald A. Clark,a John C. Mitchellb and Martin J. Snowdenb a Pfizer Global Research and Development, Ramsgate Road, Sandwich, UK CT13 9NJ b University of Greenwich, Medway Sciences, Chatham Maritime, UK ME4 4TB. E-mail: [email protected]; Fax: 144 (0)1304 656726; Tel: 144 (0)1304 643250 Received 30th June 2004, Accepted 31st August 2004 First published as an Advance Article on the web 7th October 2004 A direct comparison of univariate and multivariate data analysis has been performed to show the effect of spectral noise on the quality of chemical images derived from hyper-spectral data cubes. A data processing approach has been developed using a numerical model, based on spectra of common pharmaceutical excipients, and then applied to a real multi-layered solid dosage formulation. The results of this study demonstrate that the multivariate analysis, which in its simplest form only de-noises data using principal component analysis (PCA), produces significantly better quality chemical images than the univariate approach, even from data sets which appear visually poor. If pure component spectra are available, ordinary least squares (OLS) regression offers even better results. The ability to de-noise spectra using these approaches impacts on Raman experimental conditions and increases information content collected per unit time. Data acquisition time, which is a rate limiting step in the production of chemical images using Raman mapping and imaging techniques, is reduced by 60% and still produces multivariate chemical images of appropriate quality with which to study pharmaceutical formulations. DOI: 10.1039/b409879j Introduction Raman mapping/imaging is a fast progressing application of vibrational spectroscopy, although it was first described in 1975.1 It has been used in numerous applications that include the pharmaceutical and polymer industries,2–6 in investigation of biological/medical samples,7–12 archaeology,13 single molecule spectroscopy14,15 etc. An overview of instrumentation and applications relating to this area can be found in ref. 16. There are three kinds of laser light illumination (point and line for the mapping systems and so-called wide field illumination for the imaging systems) that excite the sample in a Raman mapping/imaging instrument.16 The instruments with the first two sources provide high quality, full-range spectra allowing different mathematical tools to be used in the process of producing images. The wide field imaging equipment should have much faster experiment times due to more specific experimental conditions. However, these instruments do not always provide spectra that are comparable in quality with those from the instruments equipped with standard dispersion optical elements. In addition, the main advantage of imaging instruments (rapid production of images) holds only if the scanning is carried out without collecting full spectra which leads us to the question of how chemical images are in fact formed. The simplest and still most convenient approach to producing chemical images is the so-called univariate approach. Here a chemical image is generated that represents the spatial distribution of a component of interest based on its intensity at a uniquely assignable Raman shift. The key requirement of univariate imaging is a precise spectral characterization of the sample before starting the imaging process, as the uniquely assignable wavenumbers should be known in advance. Another { In this study ‘multivariate imaging’ refers to using the entire spectra anywhere in the process of producing an image. option is multivariate imaging{ that employs full spectra and is tightly associated with chemometrics.17–21 There are several examples (microdamage in a bone,17 separation of waste on a conveyer belt,19 characterization of an emulsion21) that illustrate how, via the multivariate approach, one can image demanding samples for which there is very little background information (which makes univariate approach practically inapplicable). The motivation behind the work reported herein is to quantitatively compare performances of the univariate and multivariate imaging techniques. To the best of our knowledge, despite numerous publications, there are no studies that compare these two methods side-by-side because the experimentalist generally selects the method that best fits the overall conditions of the experiment. In addition, comparison of chemical images can rarely be found in the scientific literature because it is difficult to establish a reference or standard to compare with an experimental image (except for the comparative study of Raman imaging instrumentation22). Therefore, there is no literature data that numerically compare performances of the two approaches for producing images. Furthermore, the technical aspects of different Raman instruments need to be taken into consideration. The instruments that employ point and line sources produce spectra. The information content is such that the multivariate approach will always over-perform when compared to the univariate analysis and should be used whenever possible. However, the univariate approach fits well with wide-field illumination instruments and that is why we consider it worth pursuing. In this work we have produced and mapped a multicomponent model that contains components typically found in a pharmaceutical tablet, and then considered a multi-layered bead formulation as an example of a sample routinely analysed in our laboratory. The results and conclusions of this study compare the model and experimental data in terms of chemical image quality and Raman spectral signal to noise ratio. The This journal is ß The Royal Society of Chemistry 2004 Analyst, 2004, 129, 1001–1007 1001 study conclusions are discussed with respect to chemical imaging in general. Experimental Creation of the spectral model The spectra of five materials that are typically used in pharmaceutical products were used to build the spectral model. These were micro-crystalline cellulose (hereafter called avicel), explotab, di-calcium phosphate anhydrous (DCP), magnesium stearate (MgSt) and active pharmaceutical ingredient (API). The spatial distribution in the chemical images of these components was created by assigning to them the following contents: avicel 60%, DCP 31%, API 5%, explotab 3% and MgSt 1%. The spatial distributions were created such that a pixel in the image represents only one component. The model simulates data collection over a 30 6 30 pixel area so that 1% of MgSt means that 9 out of 900 pixels are assigned to this excipient. Four pixels representing MgSt grouped together represent the smallest cluster in all five chemical images. The pixels which correspond to the spatial position of the components (y0 for absence and y1 for presence) are multiplied with the pure component spectra. In this way, five 3D arrays or hyper-spectral data cubes (30 6 30 spatial 6530 spectral dimension) are obtained, one for each component. These pure component data cubes are then co-added to create the noiseless hyper-spectral model data cube. Different amounts of noise are added to the model data cube. As Raman noise is dependent on the intensity of the signal, multiplicative (heteroscedastic) noise is generated and added to the pure data in the following way. A normally distributed noise is created by a Matlab 6 (Mathworks, Natick, MA, USA) command, weighted by the pure data, multiplied by the different constants, and then added to the same data. This resulted in the presence of 1, 10, 25 and 35% of relative standard deviation (rsd) of noise for each peak in the five corresponding data cubes, respectively. The rsd is calculated from the series of spectra as the ratio between the standard deviation and the mean at a peak position. Fig. 1 shows an example of the spectra with 25% and 35% of noise. Note that the variations at the peaks are exaggerated with respect to the spectra normally encountered in practice, while the noise at the baseline appears negligible in comparison with noise on the peaks. All the spectra are meannormalized (each spectrum is divided by its mean intensity). Collection of the spectra Raman images were collected on a Renishaw Ramascope System 1000 using Wire V.1.3 software. The spectra were obtained by exciting the sample with a laser line at 782 nm. The sample was viewed and Raman data collected through a 620 objective. The data collection was set to cover the 200– 1800 cm21 range. Usually (and if possible) one selects the spectral interval that does not involve movements of the grating. For this instrument this covers a range of ca. 500 cm21. However, in this case it was necessary to collect over the whole fingerprint region in order not to miss potential univariate points and this meant much longer data collection time due to the grating scanning. The spectral resolution was reduced after binning from the original y1 cm21 to y3 cm21. The spatial resolution in the X dimension was 20 mm, set by the experiment stepsize while in the Y dimension (along the excitation line) the spatial resolution was determined by the relative CCD pixel size. The spectra were not spatially binned prior to the analysis. Sample preparation The beads were halved with a scalpel to produce a flat surface for chemical mapping. Only one half of the bead was mapped. For the purpose of this work it was only important to recognize the beads’ layers and collect the spectra from them. The composition of the bead is not discussed. Data analysis The chemical images of the model and the bead were created from the data cubes in three ways: (1) by following the intensities at the wavenumbers which are specific for each component (univariate approach), (2) by de-noising the data via principal component analysis (PCA)23,24 and then using the same univariate analysis or (3) by using the pure component spectra via ordinary least square (OLS) regression24 (this option was not applied to the bead data because the pure component spectra of the bead were not available). The images from the model with varying noise were binarized and compared with the binarized true images of the species. Only pixels with the intensities above half of the maximum intensity in an image of a component are displayed in the binarized images. As mapping of the bead was a real experiment, no true data were available for the bead images so that the binarized univariate images of the bead’s components were compared with the ones pre-processed by PCA. Results Model images Fig. 1 A noiseless spectrum from the model and the two corresponding spectra with the relative standard deviations of noise of 25 and 35% respectively (same noise, different amplitude). 1002 Analyst, 2004, 129, 1001–1007 The wavenumbers for each component that are used for univariate imaging are shown in Fig. 2. The univariate point associated with avicel is weakly selective as it is not at a peak position and has a very weak intensity. Comparison between the true images of the components and those obtained in a univariate fashion is shown in Table 1. The term ‘difference’ in Table 1 stands for the sum of absolute values obtained via subtraction of the binarized images of the model components from the corresponding true ones. Evidently, 1% of noise does not represent a problem as the binarized images of the model closely match the true ones. For 10% of noise in the model data a mismatch of 5–10% in the position of white pixels (‘1’ in a binarized image) is detected. For 25% of noise, however, the discrepancy is evident and ranges between 20 and 40% in the positions of the white pixels. Fig. 3 reveals that the image of avicel barely matches the true one. Combining images of the components in order to make a composite image of the investigated sample is not possible with 25% of noise. This is because the here defined pixel Fig. 2 The arrows mark the univariate points for the components of the model; (–) explotab, (– –) API, (–#–) MgSt, (?–?) avicel and (–.–) DCP. Table 1 Avicel API MgSt Explotab DCP a Difference in the number of white pixels (see the text for explanation) in binarized images of the model as compared with the true data Rsd of noise 1% 10% 25% 25% 35% True Method Univariate 6 (1%) 0 0 0 5 (2%) Univariate 25 (5%) 2 (6%) 0 1 (4%) 35 (12%) Univariate 212 (39%) 13 (36%) 3 (33%) 4 (17%) 118 (40%) Multivariate 50 (9%) 0 0 0 14 (5%) Multivariate 35 (6%) 3 (8%) 0 1 (4%) 26 (8%) 548 36 9 23 294 ‘Multivariate’ refers to noise elimination by PCA and subsequent imaging via univariate points. thresholding criterion is vulnerable to considerable noise and may lead to poor and unreliable composite images. 25% noise is not an issue for the multivariate approach even for avicel the spectrum of which is highly overlapped. The simplest form of this approach is to use a routine such as PCA to de-noise the data and then to produce images via the univariate points. The results obtained in this way can only be better if additional multivariate tools are employed. The eigenvalue analysis straightforwardly recognizes 5 components, and after truncating the data so that only signal is retained, a significant improvement in the quality of images is accomplished (Fig. 3). The discrepancies are now detected only for the two most abundant species (Table 1). Good results in multivariate imaging in the same way as described above are obtained with 35% of noise, and with the use of pure component spectra (OLS regression) even with 50% of noise. However, the model with so much noise does not convincingly correlate with reality so that these results are not further commented on here. Finally, an average error per all the pixels can be calculated for each component. Using OLS regression on the spectral model, the true data are quantitatively reconstructed, while for univariate and PCA-univariate approaches the images had to be re-scaled in order to match the original [0,1] interval. For 25% of noise, which is considered critical here, the average error for the univariate approach varies from 22% for DCP to 32% for MgSt. For the multivariate approaches the errors are much smaller and vary between 5 and 11% for the PCA approach, and between 4 and 9% for the OLS regression. These numbers refer to the grey scale images and as such they do not fully take into account the spatial distributions but are very useful in terms of quantification. An immediate outcome of this comparison is that semi-quantitative imaging is possible via the spectra with substantial noise, provided the hyper-spectral data cube is collected and particularly if the pure component spectra are available. Bead images The Raman image of the bead displayed in Fig. 4 reveals that there are at least four potential layers to analyse. For the purpose of this work, we concentrate only on the major components. Our first task is to relate noise employed in the model with noise in the bead spectra. Fig. 5 shows the series of spectra extracted from different points in the three visible layers with highlighted noise on the peaks. Clearly, for the spectra in all three layers the rsd of noise is inside the boundaries investigated in the model so that the results from imaging the model are expected to reasonably relate with imaging the bead via both univariate and multivariate ways. The univariate points for the layers of the bead are shown in Fig. 6. The outer layer cannot be reliably imaged as the baseline at its univariate point at 360 cm21 is poorly defined. The peak at 1520 cm21 that is uniquely assignable to this layer cannot be employed as the univariate point because the baseline variation in that part of the spectra significantly interferes with the band itself. On the other hand, the inner layer and the core have much better prospects for univariate imaging due to a much better definition of the univariate points (Fig. 6). According to the model, these two layers may be successfully imaged because the rsd at their univariate points is about 10%. If this assumption holds, the difference between the univariate and multivariate images (de-noising by PCA and imaging via the same univariate points) should be small. After binarizing the images of the three layers, the discrepancy of 4.4% in the number and position of the white pixels in the multivariate and univariate images is found for the inner layer, 11% for the core and 25% for the outer layer. The difference in imaging the core is visualized in Fig. 7. This comparison reveals satisfactory agreement between the univariate model and bead images for the given level of noise. For images obtained by the multivariate approach, the de-noised Raman spectra produce chemical images of high quality because PCA effectively removes noise from the spectra. This is also in Analyst, 2004, 129, 1001–1007 1003 Fig. 4 Raman image of the bead at 857 cm21. The size of the image is approximately 1 mm 6 0.5 mm. The symbols mark the analysed layers. Fig. 8 shows the spectra of the layers (compare with Fig. 5). The decrease in the signal-to-noise ratio is substantial and surprising as reducing the measurement time was not expected to increase the rsd’s from Fig. 5 for more than 50%.25 In reality the rsd was found to be between three and four times larger than the values in Fig. 5. Fig. 8 suggests that an image of the outer layer cannot be obtained by any approach, the inner layer still has recognizable features, while many bands of the core are buried in noise. The univariate and multivariate chemical images of the three major layers of bead are displayed in Fig. 9. While only the univariate chemical image of the inner layer may be considered acceptable, the multivariate images of all three layers are acceptable. The binarized univariate image of the inner layer reasonably compares with the multivariate one as the calculated spatial difference is 11%. The difference for the core is 33% while it is pointless to compare images of the outer layer because its univariate image is of intolerable quality. Discussion Fig. 3 The true (A), univariate (B) and multivariate (C) grey images of avicel. agreement with the model that predicts superior performance of the multivariate approach for an rsd of 10%. The obtained results indicate that a reduction of the measurement time, i.e. increase of noise, may still provide satisfactory multivariate images and acceptable univariate images of the inner layer. To test this hypothesis, the measurement time was reduced from 60 to 20 s and the experiment repeated, albeit on another bead sample. 1004 Analyst, 2004, 129, 1001–1007 Owing to the amount of data collected and computations involved, the multivariate approach produces reliable chemical images and allows for full data exploration in both spectral and spatial dimensions. The univariate approach hinges on the premise of a precise definition of the investigated systems. If something unexpected is encountered during the experiment (e.g. the sample changes), or if the instrument premises are not ideal (e.g. many cosmic events, temperature/humidity changes), one may be unable to recognize and resolve the problem. Thus, it is undisputable that the multivariate analysis is the favoured data processing method, not only in chemical imaging but in analytical spectroscopy in general. However, there are technical issues accompanying collection of such a vast number of spectra (usually measured in thousands when mapping or imaging samples). Such Raman experiments are normally performed on line or point mapping systems and may take several days to collect the data. The current state of chemical imaging hardware partially promotes global imaging systems, as they rapidly image employing univariate imaging methodology. While the measurement time is not necessarily a key issue for the pharmaceutical samples analysed here, it may be crucially important whenever biological samples are imaged as these may change with time or environmental conditions. Therefore, there is a need to estimate the relation of univariate versus multivariate data processing in terms of quality of produced chemical images. The results presented in this work quantitatively compare the performance of multivariate and univariate data analysis for samples that are well defined, and are equally applicable to any sample irrespective of its nature. The criteria for estimating the noise within spectra and quality of images employed in this work may appear simple but are recognized in the literature,25 and are universally Fig. 6 Univariate points (ƒ) for the outer (red) and inner layer (–), and for the core (– –). Fig. 7 Cropped and binarized univariate (A) and multivariate (B) images of the core in Fig. 4. Fig. 5 The series of spectra obtained from different positions on the bead: the outer layer (A), the core (B) and the inner layer (C). The numbers denote rsd (in %) at the peak positions. understandable and applicable. One has to be aware, however, that the definition of the threshold for binarizing images is a very sensitive issue specifically if intimately mixed samples are imaged. The criterion selected here seems functional when spatially well resolved solid samples are investigated but it should not be generalized as it may ignore minor components of sub-pixel size mixed samples, or in the images with poor spatial resolution. In such cases there is a significant risk of obtaining overlapped images of the single components. In order to avoid that, a more thorough criterion for pixel thresholding must be proposed. Table 1 and Figs. 5, 7, 8 and 9 illustrate the empirical target values for signal-to-noise ratio in the Raman spectra with respect to producing chemical images. The simulation suggests that even 10% of noise, as defined here, produces relatively small differences between the univariate and multivariate images. When tested on the real sample this prediction proves rather accurate. Furthermore, as one intuitively expects, the only criterion for successfully imaging a component is the strength of a component’s signal provided that there is no interference from neighbouring pixels. The simulation shows that whilst MgSt is present at a low concentration, it is in fact easy to image because of its strong spectral response in the model data. The 60 times more abundant avicel represents a demanding target due to its weakly effective univariate point. In terms of the multivariate approach, both model and bead chemical images are encouraging as they are produced using visually unacceptable spectra (Figs. 1 and 8). The ability of PCA to eliminate noise is impressive though we assume that this may be a consequence of the high compactness of the data explored here. As mentioned above, we briefly touched on the performance of OLS regression. It is clear that, provided there is a good match between the pure component and experimental spectra, OLS images can only be of higher quality than those obtained via PCA because they have higher information content. OLS regression is not often used in analytical spectroscopy as it is rarely possible to confidently recognize all the components that contribute to the experimental spectra. However, there are at least two groups that intensively practise OLS regression in analysing vibrational spectra of biological samples.26–29 The results of this work suggest that if a biological sample under investigation is spectroscopically precisely defined, meaning OLS is applicable as a data processing tool, then there is a good prospect of fast data acquisition from such a sample. If single vibrational spectra are measured, precise concentrations or Analyst, 2004, 129, 1001–1007 1005 classification lines from the spectra of relatively poor quality may be determined via OLS. For the sample analysed here, it is relatively easy to provide the spectra of the components that form the bead, so that the prospects of fast data acquisition and imaging the beads (or samples of comparable structure) are much better. In addition, if anything unexpected takes place during the experiment, it may be trivial to ascertain the chemical nature of the change (e.g. by analysing residuals between the original and via OLS reconstructed data). Our attempt to significantly reduce the collection time without a comprehensive loss of image quality was successful for the multivariate analysis whilst the performance of the univariate approach was somewhat below the expectation. The 60 s data measurement provides acceptable images by univariate analysis for two investigated components while for the third one only a much longer collection time may lead to better images (Figs. 5 and 7). After reducing the collection to 20 s, only one layer is successfully imaged by the univariate approach (Fig. 9, middle). The estimated increase in rsd with reduction of measurement time was based on the dt dependence of signal-to-noise ratio in Raman spectroscopy.25 However, as the spectra used for calculating rsd were selected from various points on the bead, other factors may adversely contribute to the rsd making dt dependence too optimistic. Still, the spectra in Fig. 8 are far too noisy. The unexpectedly higher noise level may be attributed to a higher level of fluorescence (the fluorescence background is eliminated from all the spectra shown here). In addition, the baseline interference is notable. One may argue that a weakness of the model is that it ignores baseline noise as most of the peaks in the model are clearly above the baseline. What this study has not addressed is the situation when the Raman signal is comparable to the background noise. Another study, that carefully monitors and explores baseline variations, is required for this case. Nonetheless, the data from the model are clearly helpful for comparing the quality of univariate and multivariate data processing approaches, and for empirically relating the quality of obtained chemical images with the criterion for signal-tonoise ratio in the hyper-spectral data cubes. Conclusion Fig. 8 The same as in Fig. 5 with the data collection time being only one third of that used for the spectra in Fig. 5. Fig. 9 The univariate (top) and multivariate (below) images of the layers of the bead mapped for 20 s. 1006 Analyst, 2004, 129, 1001–1007 This study reports on the influence of 1, 10, 25 and 35% of noise (ratio of rsd at the peak maximum to the mean maximum in a set of spectra) on the quality of images obtained from the artificial hyper-spectral data cubes. The images are obtained in univariate and multivariate ways, the latter referring to PCA used only for noise elimination, or to OLS regression. The comparison of the binarized true and model images shows that the univariate approach is reasonably effective between 1 and 10% of rsd at the peaks but fails at 25% of rsd. On the other hand, the multivariate approach in its simplest form efficiently reduces noise and produces high quality chemical images despite the presence of substantial, even exaggerated noise in the original data. The results of the OLS regression seem even more optimistic. The model spectra are related to the Raman spectra obtained by mapping pharmaceutical samples. It is found that the spectra collected during routine mapping of a bead are inside the boundaries of noise used in the model so that the conclusions from the model can be verified on the bead images. The differences between the chemical images of the layers of the bead created by the univariate and multivariate approaches are more pronounced than expected. Using the results from the model, a mapping experiment is conducted with dramatically reduced measurement time. As the noise has increased above the prediction, the univariate approach becomes practically inapplicable. However, the multivariate approach still produces high quality images despite the Raman spectra being visually of a very low quality. Acknowledgements We thank Drs R. Brody and A. de Paepe of Pfizer Global Research and Development for useful discussions. References 1 M. Delhaye and P. Dhamelincourt, J. Raman Spectrosc., 1975, 3, 33. 2 F. C. Clarke, M. J. Jamieson, D. A. Clark, S. V. Hammond, R. D. Jee and A. C. Moffat, Anal. Chem., 2001, 73, 2213. 3 S. L. Zhang, J. A. Pezzuti, M. D. Morris, A. Appadwedula, C.-M. Hsiung, A. Leugers and D. Bank, Appl. Spectrosc., 1998, 52, 1264. 4 M. Malecha, C. Bessant and S. Saini, Analyst, 2002, 127, 1261. 5 H. R. Morris, J. F. Turner, II, B. Munro, R. A. Ryntz and P. J. Treado, Langmuir, 1999, 15, 2961. 6 R. Appel, T. W. Zerda and W. H. Waddell, Appl. Spectrosc., 2000, 54, 1559. 7 M. D. Schaeberle, V. F. Kalasinsky, J. L. Luke, I. W. Levin and P. J. Treado, Anal. Chem., 1996, 68, 1829. 8 J. Ling, S. D. Weitman, M. A. Miller, R. V. Moore and A. C. Bovik, Appl. Opt., 2002, 41, 6006. 9 A. Carden, R. M. Rajachar, M. D. Morris and D. H. Kohn, Calcif. Tissue Int., 2003, 72, 166. 10 N. Uzunbajakava, A. Lenferink, Y. Kraan, B. Willekens, G. Vrensen, J. Greve and C. Otto, Biopolymers, 2003, 72, 1. 11 J. Kneipp, T. B. Schut, M. Kliffen, M. Menke-Pluijmers and G. Puppels, Vib. Spectrosc., 2003, 32, 67. 12 C. Krafft, T. Knetschke, A. Siegner, R. H. W. Funk and R. Salzer, Vib. Spectrosc., 2003, 32, 75. 13 J. W. Schopf, A. B. Kudryavtsev, D. G. Agresti, T. J. Wdowiak and A. D. Czaja, Nature, 2002, 416, 73. 14 C. J. L. Constantino, T. Lemma, P. A. Antunes, P. Goulet and R. Aroca, Appl. Spectrosc., 2003, 57, 649. 15 M. Ishikawa, Y. Maruyama, J. Y. Ye and M. Futamata, J. Biol. Phys., 2002, 28, 573. 16 P. J. Treado, M. Nelson, in Handbook of Vibrational Spectroscopy, ed. J. M. Chalmers and P. R. Griffiths, Wiley, New York, 2001, vol. 2. 17 J. A. Timlin, A. Carden, M. D. Morris, R. M. Rajachar and D. H. Kohn, Anal. Chem., 2000, 72, 229. 18 J. M. Shaver, K. C. Christensen and M. D. Morris, Appl. Spectrosc., 1998, 52, 259. 19 W. H. A. M. van den Broek, D. Wienke, W. J. Melssen, C. W. A. de Crom and L. Buydens, Anal. Chem., 1995, 67, 3753. 20 J. -H. Wang, P. K. Hopke, T. M. Hancewicz and S. L. Zhang, Anal. Chim. Acta, 2003, 476, 93. 21 J. J. Andrew and T. M. Hancewicz, Appl. Spectrosc., 1998, 52, 797. 22 S. Schlücker, M. D. Schaeberle, S. W. Huffman and I. W. Lewin, Anal. Chem., 2003, 75, 4312. 23 P. Geladi and H. Grahn, Multivariate Image Analysis, Wiley, New York, 1996. 24 B. G. M. Vandenginste, D. L. Massart, L. M. C. Buydens, S. de Jong, P. J. Lewi and J. Smeyers-Verbeke, Handbook of Chemometrics and Qualimetrics B, Elsevier, Amsterdam, 1998. 25 R. L. McCreery, Raman Spectroscopy for Chemical Analysis, Wiley, New York, 2000. 26 K. E. Shafer-Peltier, A. S. Haka, M. Fitzmaurice, J. Crowe, J. Myles, R. R. Dasari and M. S. Feld, J. Raman Spectrosc., 2002, 33, 552. 27 H. P. Buschman, G. Deinum, J. T. Motz, M. Fitzmaurice, J. R. Kramer, A. van der Laarse, A. V. Bruschke and M. S. Feld, Cardiovasc. Pathol., 2001, 10, 69. 28 S. W. E. van de Poll, T. J. Romer, O. L. Volger, D. J. M. Delsing, T. C. B. Schut, H. M. G. Princen, L. M. Havekes, J. W. Jukema, A. van der Laarse and G. J. Puppels, Arterioscler., Thromb., Vasc. Biol., 2001, 21, 1630. 29 P. J. Caspers, G. W. Lucassen, E. A. Carter, H. A. Bruining and G. J. Puppels, J. Invest. Dermatol., 2001, 116, 434. Analyst, 2004, 129, 1001–1007 1007
© Copyright 2026 Paperzz