Near Infrared Spectroscopy: the Practical Chemical Imaging Solution

NIRSPECTROSCOPY Near infrared spectroscopy: the practical chemical imaging solution

Frederick W.Koehler IV,Eunah Lee, Linda H. Kidder and E. Neil Lewis* Spectral Dimensions, Inc., 3416 Olandwood Ct, Suite 210 Olney,MD 20832, USA. E-mail: [email protected]

microns to kilometers using micro- Introduction scopes or satellite-based remote sens- Foundations of Chemical imaging spectroscopy is an ing systems, makes it truly unique. exciting new analytical advance that The large format infrared cameras cur- chemical imaging answers commonly asked questions rently available, particularly in the The foundation of chemical imaging such as what chemical species are in a NIR, provide high image fidelity and is the concept of a chemical map. As sample, how much of each is present, coupling these arrays to precise shown in Figure 1, a chemical imaging and most importantly, where are they infrared wavelength selection devices data set is represented by a three- located? Through the fusion of tradi- with narrow bandpass capabilities dimensional cube where two axes tional infrared spectroscopy with pow- including Fourier transform spectrom- describe vertical and horizontal spatial erful microscopic and macroscopic eters1,2 and liquid crystal tunable dimensions and the third dimension imaging capabilities, chemical imaging filters3 make the technique fully represents the spectral wavelength spectroscopy answers all these questions hyperspectral. In fact, most commer- dimension. Like a deck of cards, there simultaneously, in a single rapid mea- cially available spectral imaging systems is one image per wavelength interval surement.1 Chemical imaging enables have spectral resolutions comparable to stacked sequentially. The intensity of a the researcher to obtain spatial and those of conventional infrared and single pixel plotted as a function of the spectral information characterising sam- NIR spectrometer systems. This type wavelength dimension represents a ples with unprecedented ease, speed of instrumentation has already started standard NIR spectrum. Intensities for and spatial and spectral resolution. The to supplant time-consuming mapping all pixels at a single wavelength repre- methodology is aimed at providing a techniques, particularly in the infrared sent an image of absorption at a partic- comprehensive analysis of complex spectral region, in which data is ular spectral band across the sample. It heterogeneous samples. recorded one spectrum at a time and is the wavelength dimension that visu- Chemical imaging has advanced sig- the image is constructed by moving alises chemical specificity by segregat- nificantly with the commercialisation the sample in an x,y pattern under the ing different chemical species to differ- of infrared focal plane arrays (FPAs), spectrometer optics. It is likely that ent spectral regions of the chemical which are cameras composed of many infrared array-based imaging approach- image cube. thousands of individual infrared detec- es will completely replace single-point A variety of analytical techniques can tor elements. When coupled with mapping techniques in the future. be used to visualise and process chemi- infrared optics and a means of wave- Although we have applied this gen- cal image data sets. Simple gray scale length selection, these instruments pro- eral hyperspectral imaging approach to chemical maps can be constructed from vide an image where varying contrast is tackle a number of biological and a single image plane. For more detailed derived from the unique infrared industrial problems in the past, and contrast, mapping individual image chemical signature characteristic of have developed a variety of instruments planes onto red, green and blue (RGB) each component within the sample. and methods for mid-infrared, NIR channels can create composite colour Powerful statistical and chemometric and Raman imaging we will focus in chemical images. Individual chemical tools can enhance these chemical maps this paper on industrial chemical imag- images can be ratioed against each by performing valuable data reduction ing problems using NIR diffuse other to normalise for pathlength steps to identify and extract the most reflectance only. We believe that this effects or to perform spectral normali- analytically useful information. This spectral interval strikes a perfect balance sation and further isolate components can be invaluable to winnow the large between sensitivity, flexibility, simplic- of interest. These types of univariate volumes of data collected by this tech- ity and ruggedness that makes it an approaches work well for species with nique. ideal industrial imaging tool. In addi- unique spectral signatures. For cases Infrared and near infrared (NIR) tion, we will illustrate through several with significant spectral overlap, com- imaging, in particular, has the experi- examples the unique features that mon in NIR, or when the maximum mental flexibility to characterise a chemical sensing with array detectors signal to noise ratio is required, statisti- wide variety of samples chemically affords. With the application of quanti- cal and chemometric approaches such using transmission and reflectance tative data analysis in these examples, as principal component analysis (PCA) measurements, while its ability to we will show it’s not just about pretty or partial least squares (PLS) regression tackle samples ranging in size from pictures. can be utilised. Additionally, after iso-

12 © Spectroscopy Europe 2002 Spectroscopy Europe 14/3 (2002) NIRSPECTROSCOPY

the basis from which relative concen- trations can be determined for each spatial location. The entire process of pure component determination and concentration estimation can be derived from a single imaging experiment. While NIR-CI may appear at first sight not to have the analytical specificity of mid-infrared (MIR) imaging or the spatial resolution of Raman imaging, in many real-world situations, it outperforms them both. It does this by taking advantage of the same unique qualities that traditional NIR spectroscopy offers over these same techniques. For example, whilst providing a degree of spatial resolution, which is better than that of mid-infrared imaging systems and competitive with Raman systems, when coupled to a microscope, it can also be readily adapted to a wide variety of fields-of- Figure 1. Infrared chemical imaging data cube of a finished pharma- view (FOV). Global Raman imaging ceutical product. The single pixel spectrum shown corresponds to quickly runs out of steam in the case intensity as a function of wavelength at a fixed spatial location and where larger FOVs are desired because provides the spectral signature of chemical components present in of the inability to flood these areas with that part of the sample. The single channel image shown corresponds sufficient laser power. Mid-infrared to intensity for all pixels at a fixed wavelength. This carries informa- imaging is hampered by the fact that all tion about the spatial distribution of a chemical component with a dis- commercially available systems collect tinguishable spectral signature at that wavelength. data by spectrally modulating infrared energy from the source and do not fil- ter the image. This rules out the possi- lating spectral information unique to it has become an extremely powerful bility of using multiple sources to the sample of interest, standard image adjunct to NIR spectroscopy in a address the ever-increasing power den- processing approaches can be applied number of different ways. sities required to illuminate larger and statistical analysis performed to For example, in contrast to tradition- FOVs. The single globar in a typical characterise numbers of particles, their al spectroscopic techniques which FT-IR system was not designed for size and relative concentration. analyse the sample in bulk and deter- spectral imaging and is unlikely to pro- mine an average composition across the vide the necessary energy. While both entire sample, individual detector ele- mid-infrared (MIR) imaging and ments in a chemical imaging array can Raman imaging perform well as micro Why NIR detect spatially isolated minor chemical techniques, they quickly run into diffi- species. Because the data is collected in culty imaging samples much larger than imaging? parallel the result is not hampered by a a few millimeters. Because NIR-CI Conventional NIR spectroscopy is a “dilution” effect in the same way a sin- uses relatively simple quartz-tungsten- widely adopted method for performing gle spectral bulk measurement would, halogen (QTH) sources, and an image analytical measurements. It has become and trace sample contamination mea- filtering, not a source filtering the obvious choice in many industrial surements can be made extremely approach, wide-field illumination for a applications because it provides both rapidly. While this problem is also variety of magnifications and imaging qualitative and, in particular, quantita- tractable using conventional micro- modes is relatively trivial. Both infrared tive results non-destructively on solids mapping spectrometers, due to lengthy and Raman mapping approaches can and liquids in reflectance or transmit- data collection times, these approaches deal with this FOV issue, but the data tance mode. In many cases sample quickly become impractical for routine collection time is generally not accept- preparation is minimal or non-existent, work. able. making the method extremely user- Recent advances in chemometric NIR-CI is also free of the need for friendly and rapid. In addition, the factor analysis approaches have identi- the very flat samples necessary for glob- optical systems (sources, detectors, lens- fied another interesting property of al Raman imaging. This flatness es, fibre optics etc.) are rugged, exhibit array-based chemical sensing which requirement is imposed by the limited high-performance and are readily avail- enables quantitative information to be depth-of-field of a global Raman imag- able, resulting in a practical technique obtained without running separate cali- ing experiment as a direct result of the for widespread use inside and outside bration samples. The techniques take necessity to use very high numerical of highly regulated laboratory environ- advantage of the large numbers of indi- aperture objectives to collect the limit- ments. Just like its older, conventional vidual spectra recorded in a typical ed amount of signal available. A similar NIR cousin, NIR chemical imaging image cube and utilise the inherent sampling limitation is also encountered (NIR-CI) retains all of these advan- spectral variance across the spatial in MIR imaging for different reasons. tages and gains several more through dimension of heterogeneous samples to In most cases the reflectance mode has the addition of spatial dimensions and obtain pure component spectral esti- proven to be energy limited, suffering parallel data collection. In recent years mates. These spectral estimates provide from a poor signal-to-noise ratio and

14 Spectroscopy Europe 14/3 (2002) NIRSPECTROSCOPY

requiring extended periods of data collection time. Therefore the over- whelming body of MIR imaging data published to date has been recorded in transmittance mode. This requires the samples to be sectioned and, like the global Raman imaging measurement, optically flat. These issues have limited the general applicability and scope of both of these imaging approaches. On the other hand, NIR imaging systems perform particularly well in reflectance mode. Because of the excellent noise characteristics of the NIR arrays and the abundant optical signal available, these systems can be designed without the need for high collection efficiency imaging optics and consequently exhibit large depths-of-field. This design philosophy provides extreme tolerance to variations in sample geom- etry allowing most NIR chemical imaging experiments to be performed on very irregular samples without sample preparation. Analytical targets NIR-CI analysis not only makes the qualitative determination that particular Figure 2. (a) Visible image of an over-the-counter plain medication species are present in the sample, but tablet. (b) PCA score plot for factor 3. (c) ALS “pure” component load- also visualises the spatial distribution of ing spectra for the API and Excipient mixture. (d) Concentration esti- these chemical species throughout the mate image. White pixels represent the API with an estimated bulk sample. For many materials, both man- concentration of 19.38%. made and naturally occurring, the degree of chemical heterogeneity within any given sample is itself a critical this category and the chemical mor- image cube recorded with approxi- functional parameter not easily mea- phology of these materials can dictate mately one minute of data acquisition sured via standard analytical approach- properties such as dissolution rate and time. A standard personal computer es. Analogous to parallel computer pro- therefore potency. This will be the running the Windows 2000TM operat- cessing, the array detector collects spec- subject of a second example applica- ing system was employed for data col- tral data in parallel, allowing NIR-CI tion. lection and analysis (IsysTM, Spectral to investigate multiple samples simulta- Biological tissues, both plant and Dimensions, Inc., Olney, MD, USA). neously. Using this approach any dif- animal, are by far the most complex A second data set consisting of a ferences between individual samples are examples of materials which are both reflectance standard was collected in readily identified using unsupervised chemically and spatially complex. each case and used as a background. statistical approaches. Biology is mediated across spatial Data is presented as log10(1/R) units Samples that benefit from NIR-CI chemical gradients and cannot function computed through an inverse log trans- fall into several categories with the ana- unless this complex structural frame- form after dividing the sample data set lytical goal being slightly different for work is intact. These materials also with the background image cube. each. For mixtures in which the distri- make excellent examples of the utility bution of individual species in the sam- of chemical imaging. ple is largely random, and doesn’t affect Quantitative the functionality, a typical analytical goal is often the rapid determination of Experimental assessment of the relative abundance of each species in the mixture. In other cases the sam- protocol pharmaceutical ple may be assumed to be largely The imaging data used in the follow- homogeneous with a need to screen ing examples were recorded in diffuse formulations and identify contaminants. An example reflectance mode using a commercially In our first example, NIR-CI is used of this latter application is screening available NIR-CI spectrometer to visualise and quantify the spatial dis- cattle feed for animal protein contami- (MatrixNIR, Spectral Dimensions, tribution of the active pharmaceutical nants and will be covered in the article. Inc., Olney, MD, USA) tuning over ingredient (API) in an over the counter For samples in which the spatial dis- the wavelength range 1050–1700 nm (OTC) child’s pain medication tablet. tribution of the components in a mix- with a spectral point spacing of 10 nm. Figure 2(a) consists of a visible image ture is critical for functionality, the This particular imaging system utilises taken of this tablet and demonstrates analytical goal may be to correlate an indium gallium arsenide (InGaAs) the relative uniformity of the sample physical properties with this chemical FPA detector consisting of 240 × 320 surface under visible illumination. heterogeneity. Pharmaceuticals fall into pixels for a total of 76,800 spectra per Using an unsupervised principal com-

16 Spectroscopy Europe 14/3 (2002) NIRSPECTROSCOPY

ponent analysis (PCA) technique to examine the NIR imaging data, spectral variations between the API and excipient are automatically identified and exploited to generate image contrast. The principal component image shown in Figure 2(b) highlights spatial locations with similar spectral signatures by assigning them similar scores and therefore similar contrast. It follows that regions with similar spectral signatures will also have similar chemical composition and in this particular chemical map, red pixels indicate drug rich regions and blue pixels represent the surrounding matrix of excipient components. While the image highlights the high degree of chemical heterogeneity of this particular formula- tion, it is purely qualitative and does not provide any quantitative information, or any pure component spectra. Further analysis was carried out by spatially isolating only those pixels corresponding to the tablet (31,683 total), calculating their second derivative spectra and performing an alternating least squares analysis (ALS).4 Using this approach an estimation of the spectra Figure 3. (a) Bright field image of cattle feed particles with 1% (wt / wt) of the pure chemical components and poultry by-product contamination. (b) Pure component training spec- their relative abundance in the formu- tra of cattle feed (blue) and poultry by-product (red). (c) PLS concen- lation can be determined without the tration image calculated with respect to the poultry by-product class. need to collect spectra of standard ref- The red particle is assigned to the poultry by-product while the yellow erence materials. Figure 2(c) shows the particles show regions of intermiediate concentration. Blue particles calculated pure component NIR spec- are classified as pure cattle feed. tra after performing ALS for the API (blue line) and the bulk excipient mixture (red line). Initial estimates of the pure component spectra were derived farm animals, and ultimately the human corresponding to “pure” cattle feed and from a principal component analysis on population, have mandated that the “pure” poultry by-product. The term these derivative spectra. composition of animal feed be more “pure” here does not refer to a pure These pure component spectra can strictly controlled. In particular, chemical species since each component be used to generate concentration esti- because the disease is postulated to be is a complex aggregate from many dif- mates for each analyte at every spatial spread by ingesting proteins from other ferent sources. The calibration set for location. Figure 2(d) shows the corre- slaughtered, and potentially diseased, each component was extracted from a sponding API concentration distribu- animals, the amount of animal protein single image data set that spanned the tion image calculated for all 31,683 present in animal feedstocks must not chemical variation in each of the pure spatial locations. The white regions exceed 0.1%. Therefore, a critical need samples. Figure 3(b) shows the mean represent areas rich in the API and the exists to rapidly screen and identify low spectrum for each of the components. black pixels correspond to domains of levels of animal protein contamination Figure 3(c) shows the result of per- the excipient mixture. The total num- in bulk feed. forming a partial least squares (PLS) ber of pixels assigned to API is 6,140 Figure 3(a) shows a bright field prediction on the mixture image data corresponding to a bulk concentration image of an 8.5 mg sample of cattle set using a model computed with this estimate of 19.38% [(6140 / 31,683) × feed contaminated with poultry by- calibration set where each component 100]. This result compares favorably product at approximately 1% concen- concentration was set to unity. The with the 20% concentration by weight tration by weight. This image lacks resulting concentration image map reported by the manufacturer. intrinsic contrast between the individ- highlights regions of the mixture that ual particles making it extremely diffi- are rich in the poultry by-product. In cult to identify visually the presence or this example, the red particle is High-throughput absence of the contaminant. Figure 3(c) unequivocally identified as belonging on the other hand shows a NIR chem- to the poultry class, while the yellow screening of ical image of the same group of parti- particles are identified as “suspicious” cles where the contaminant is clearly and may contain some lower fraction biological identified. This data set was collected of poultry contamination. Blue parti- in a manner identical to the pharma- cles are identified as belonging to the material ceutical application but a supervised cattle feed class and would be consid- New regulations imposed attempting data processing approach was ered harmless. From a simple calcula- to contain the spread of Bovine employed. Spectra to be used as a cali- tion of the area occupied by the conta- Spongiform Encephalopathy, BSE, in bration set were collected of samples minant particle relative to the total area

18 Spectroscopy Europe 14/3 (2002) NIRSPECTROSCOPY

covered by all the particles we arrive at minute more recent studies have indi- these parameters affect their intended a statistical estimate of 1.5% for poultry cated that this type of discrimination functionality. contamination. This is in relatively might be possible on many more parti- good agreement with the actual value cles at significantly higher speeds. Acknowledgements of 1%. The authors would like to thank Dr This application represents an exam- Ishiguro of the Japanese Ministry of ple where the focus is not imaging or Conclusions Agriculture for providing the samples characterising spatial heterogeneity, but Infrared arrays and chemical imaging of cattle feed used in the study. rather high-throughput screening. The represent a major enhancement to the large number of individual NIR detec- capabilities of conventional infrared tors (76,800) on a typical array and the and NIR spectroscopy through the References variable FOV of this particular NIR introduction of spatial dimensions and 1. E.N. Lewis, P.J. Treado, R.C. imaging system enables hundreds or parallel data collection. Advanced Reeder, G.M. Story, A.E. even thousands of particles to be chemometric, statistical and image Dowrey, C. Marcott and I.W. screened simultaneously. In essence, analysis routines are available which Levin, Anal. Chem. 67, 337 the instrument performs like hundreds can enhance the chemical image data (1995). of microspectrometers all operating in collected to obtain concise results as 2. L.H. Kidder, I.W. Levin, E.N. parallel, each one responsible for estab- intuitive to interpret as viewing a pic- Lewis, V.D. Kleiman and E.J. lishing the origin of a discrete particle ture. Further, this information can be Heilweil, Opt. Lett. 22, 742 in the mixture. This process provides used to characterise the abundance, dis- (1997). an enormous throughput advantage tribution and mean size of the chemical 3. E.N. Lewis, J.E. Carroll, F. and a corresponding increase in sensi- domains present in complex heteroge- Clarke, NIR news 12(3), 16 tivity over traditional bulk spectroscop- neous samples. This information is (2001). ic or microspectroscopic approaches. often the key to understanding the 4. R. Tauler, A. Smilde, B. While the data presented in this exam- bulk physical properties of many man- Kowalski, J. Chemometr. 9, 31 ple was recorded in approximately one ufactured materials and hence how (1995).

Spectroscopy Europe 14/3 (2002) 19