Journal of Cellular Biochemistry Supplement 39:125–137 (2002)

Model-Based Biological Raman Spectral Imaging

Karen E. Shafer-Peltier,1 Abigail S. Haka,2 Jason T. Motz,2 Maryann Fitzmaurice,3

Ramachandra R. Dasari,2 and Michael S. Feld2*1Department of Biomedical Engineering, Northwestern University, Evanston, Illinois 602082G.R. Harrison Spectroscopy Laboratory, Massachusetts Institute of Technology, Cambridge,Massachusetts 021393Department of Pathology, University Hospitals of Cleveland and Case Western Reserve University,Cleveland, Ohio 44106

Abstract Raman spectral imaging is a powerful tool for determining chemical information in a biological specimen.The challenge is to condense the large amount of spectral information into an easily visualized form with high informationcontent. Researchers have applied a range of techniques, from peak-height ratios to sophisticated models, to produceinterpretable Raman images. The purpose of this article is to review some of the more common imaging approaches, inparticular principal components analysis, multivariate curve resolution, and Euclidean distance, as well as to present anew technique, morphological modeling. How to best extract meaningful chemical information using each imagingapproach will be discussed and examples of images produced with each will be shown. J. Cell. Biochem. Suppl. 39:125–137, 2002. � 2002 Wiley-Liss, Inc.

Key words: Raman spectroscopy; confocal microscopy; imaging; cells; breast; artery; tissue; principal componentanalysis; Euclidean distance; multivariate curve resolution; least-squares modeling

Raman spectroscopy can provide detailedqualitative and quantitative information abouta sample being studied. It is an inelastic scat-tering process in which photons incident on asample transfer energy to or from the sample’svibrational or rotational modes. The differencein energy between the incident and exiting pho-tons corresponds to the transition of a moleculefrom one state to another. Since the energylevels are unique for each molecule, Ramanspectra are chemical specific [McCreery, 2000].The wealth of information obtainable from aRaman spectral image has led to its use in manyfields, including environmental [Nelson et al.,2001], industrial [Andrew et al., 1998], poly-mers [Appel et al., 2000], semiconductors[Schaeberle et al., 2001], food science [Archibald

et al., 1998], and pharmaceutical [Clarke et al.,2001] applications. Raman spectral imaginghas only begun to be applied to the chemicalanalysis of complex biological samples. In thepast few years, however, a number of papershave been published using Raman spectralimaging to monitor chemical contributions atthe cellular and sub-cellular level.

Several approaches have been employed toacquire Raman imaging data sets. The threestandard approaches are point scanning, linescanning, and direct imaging [Delhaye andDhamelincourt, 1975; Turrell and Corset, 1996].Direct imaging involves the collection of a fullimage with a single spectral component. Wave-length selectivity is achieved by using either anacousto-optic or a liquid crystal tunable filterthat sweeps through specified wavelength in-tervals capturing a frame at each. Line scan-ning and point scanning collect a full spectrum(usually covering Raman shifts between 400and 1,800 cm�1 for biological media), eitherwhile imaging a line or a single point. The resul-tant data set from each of these approaches canbe thought of as a hypercube of Raman intensityas a function of Raman shift and two spatialaxes.

� 2002 Wiley-Liss, Inc.

Grant sponsor: NIH; Grant number: P41-RR 02594; Grantsponsor: Pathology Associates of University Hospitals.

*Correspondence to: Michael S. Feld, MassachusettsInstitute of Technology, 77 Massachusetts Ave., Cambridge,MA 02139, USA. E-mail: msfeld@mit.edu

Received 16 October 2002; Accepted 16 October 2002

DOI 10.1002/jcb.10418Published online in Wiley InterScience(www.interscience.wiley.com).

One of the first tissues to be explored usingRaman imaging was human breast tissue. In1996, Treado et al. published Raman images offoreign polymer inclusions (from silicone breastimplants) anchored in the fibrous breast tissuesurrounding the implant [Schaeberle et al.,1996]. Although the Raman spectrum obtainedfrom the normal biopsy tissue had very weak,uninterpretable features, the 1,615 cm�1 bandof Dacron polyester was easily identified, enabl-ing imaging of the polymer material within thebreast tissue matrix. A subsequent study ofbreast tissue demonstrates the use of Ramanimaging to develop a model to understand thechemical and morphological origins of Ramanspectral features observed in normal and dis-eased tissue [Shafer-Peltier et al., 2002]. Thismorphological model, consisting of spectraacquired from the cell cytoplasm, cell nucleus,fat, b-carotene, collagen, calcium hydroxyapa-tite, calcium oxalate dihydrate, cholesterol-likelipid deposits of normal and diseased samplescan be used in a linear combination to fitmacroscopic breast tissue spectra. The modelcan in turn be used to generate Raman mapswith highly specific information content.

In addition to mapping tissue architecture,Raman imaging can be used for in situ chemicalinvestigation of disease processes. One suchexample is atherosclerosis where the end pro-duct of the disease, ceroid, is defined as anautofluorescent lipid product whose chemicalcomposition is unknown. Recently, van de Pollet al. [2002] have explored the chemistry ofceroid using Raman imaging. This study hasprovided new insight into the chemical nature ofceroid and the disease processes that produceit. In another experiment, designed to studyfatigue-related microdamage in bone, spectro-scopically distinct microstructures were corre-lated with tissue damage [Timlin et al., 2000].Raman images have even been acquired tostudy sub-cellular chemistry. Arikan et al.[2002] have used Raman imaging as a tool tomonitor beta-carotene in live corpus luteumcells, while Freeman et al. [1998] have inves-tigated the sub-cellular localization of zincphthalocyanines, photosensitizing agents usedin photodynamic therapy. Surface-enhancedRaman spectroscopy in conjunction with imag-ing has been used to study the chemical com-position of live cells [Kneipp et al., 2002]. Inparticular, the DNA and phenylalanine con-tents of the cells were monitored.

A time-honored technique for creating spec-tral images is by examination of a specific peakheight. In this approach, the intensity of aparticular Raman band at each spatial locationis plotted to produce an image [McCreery, 2000].This method has been widely used and providesinformation about the spatial location of everymolecule in the sample that contributes inten-sity to the vibrational frequency chosen. How-ever, this approach only takes advantage of asmall portion of the data available. In complexbiological samples, where several distinct moi-eties may contribute intensity to a particularRaman band, it is necessary to incorporate all ofthe spectral information in order to differenti-ate them. This is achieved by the application of amodel that utilizes the full spectrum, as is donewith point and line scanning, when creating animage. The key is to compress the informationinto a manageable, yet still informative form.Some common data compression techniques,which will be presented here, are principalcomponent analysis (PCA), multivariate curveresolution (MCR), and Euclidean distance.Morphological modeling is a new approach thatwill also be presented.

Each one of these techniques relies on thebasic assumption that the Raman spectrum of amixture of chemicals can be represented as alinear combination of the mixture’s componentspectra [Buschman et al., 2001a,b; Shafer-Peltier et al., 2002]. Raman images are gener-ated by fitting basis spectra contained withinthe model to the Raman spectrum obtained ateach position in the image. Generally, the morea basis spectrum contributes to a data spec-trum, the larger the fit coefficient and thebrighter that spot appears in the image of thecomponent being examined. In the cases of PCAand MCR, basis spectra are mathematicallyderived, whereas for Euclidean distance andmorphological modeling, basis spectra are ex-perimentally determined.

In PCA, singular-value decomposition is usedto calculate basis spectra [Wold et al., 1987]. Thefirst basis spectrum, or principal component,accounts for the maximum variance in the dataif the data is mean-centered prior to analysis.The second basis spectrum accounts for the nextmost variance, and so on, until the basis spectraaccount only for the noise in the data. Thesebasis spectra are created such that they areorthogonal to each other, and therefore containno overlapping spectral information. The fit

126 Shafer-Peltier et al.

coefficients obtained when these principal com-ponents are fit to the imaging data set can beused to create a two-dimensional image. Thisimage will provide a map of how the spectralfeatures represented by the principal com-ponents are distributed in the sample. In turn,this map can be correlated with morphologicalfeatures observed through another optical tech-nique, such as phase contrast microscopy orlight microscopy with histological staining. Thelineshapes of the principal components mightalso be correlated with the Raman spectra ofknown chemicals, however, this is difficult asthe principal components contain both negativeand positive spectral features.

MCR is designed to extract basis spectra thatare similar to the real Raman spectra of thechemicals present in the sample [Tauler et al.,1994; Andrew and Hancewicz, 1998]. An initialestimate of the concentrations or basis spectrapresent in the sample is used in a constrained,alternating least-squares optimization. Newestimates for the concentrations and basisspectra are generated by iterating betweenleast-squares solutions for basis spectra andconcentrations. These equations can be solvedsubject to non-negativity constraints to ensurethat both the basis spectra and concentrationsare all positive and thus physically relevant.Optimization continues until the changes inthe concentrations and basis spectra from oneiteration to the next are minimal. The morecomplex the system, the better the initial esti-mates need to be to obtain meaningful solutionsto these equations. Due to the high-degree ofoverlap in the spectral features of differentcomponents and the noise inherent in the data,MCR cannot always converge on the correctsolution. However, when a solution is found, thebasis spectra produced resemble the Ramanspectra of the individual chemicals present inthe sample. Once again, the fit coefficients of thebasis spectra can be used to produce an image.

Both PCA and MCR are useful techniqueswhen little is known about the sample a priori.They enable one to extract spectral informationwithout knowing its chemical origin. Euclideandistance measurements and morphologicalmodeling both use information about the knownchemistry of a sample to create an image.Euclidean distance only requires the knowledgeof a few chemicals present whereas morphol-ogical modeling requires knowledge of all ofthe major contributors to the sample’s Raman

spectrum. Despite requiring the most priorknowledge of the sample, morphological model-ing produces the most easily interpretableresults.

Euclidean distance classifies spectral var-iance in the image data from a basis spectrum,usually a pure chemical spectrum, according tothe data’s geometric distance [Potter et al.,2001]. The distance is calculated using the

equation:ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiPlðSðlÞ � PðlÞÞ2

rwhere d is the

Euclidean distance,S is the sample data,P is thepure chemical spectrum, and l represents thewavelengths over which the spectra are acquir-ed. The more a spectrum in the image dif-fers from the basis spectrum, the larger thedistance.

Morphological modeling is a new techniquefor analyzing Raman images, which uses ordin-ary least-squares to fit a set of basis spectra tothe data [Buschman et al., 2001a; Shafer-Peltier et al., 2002]. The origin of the basisspectra is what makes this approach so useful.The basis spectra are acquired from the majormorphological features found in a set of repre-sentative samples using a Raman confocalmicroscope. By using a spectrum of a morpho-logical feature acquired in situ, one obtains aspectrum that represents that morphologicalcomponent in its chemical microenvironment.The basis spectra should account for all of themajor chemicals present in the sample, butboth the signal to noise of the data as well asthe degree of overlap of the basis spectra mustbe considered to determine whether they can beaccurately be resolved. Although basis spectracan be acquired from pure chemical compounds[Brennan et al., 1997], morphologically derivedcomponents [Buschman et al., 2001b; Shafer-Peltier et al., 2002] are preferable as they arederived from actual samples, and are thus closerthan pure chemical spectra to what is observ-ed in situ. Sometimes, a combination of purechemical components and morphologicallyderived components will produce the best resultif the chemicals of interest do not occur in-dependently within a sample. If a model is wellchosen, the images produced can reveal detailedmorphological and chemical structure in thesample.

In this article, the application of morphologi-cal modeling to Raman images of human coloniccarcinoma cells as well as human breast andartery samples will be demonstrated. This new

Raman Spectral Imaging 127

technique of morphological modeling will thenbe compared with other commonly used techni-ques, primarily: peak height analysis, PCA,MCR, and Euclidean distance. The advantagesand disadvantages of each technique, as well aswhen to use them, will be discussed.

MATERIALS AND METHODS

Tissue Handling

Breast tissue samples were obtained fromexcisional biopsy specimens while artery sam-ples were obtained from explanted hearts at thetime of transplant. Once removed, the tissuewas snap frozen in liquid nitrogen and stored at�808C. The tissue samples were then mountedon a cryostat chuck using Histoprep (FisherDiagnostics, Orangeburg, NY) and cut into 6 mmthick sections using a cryomicrotome (Interna-tional Equipment Company, Needham Heights,MA). Several consecutive sections were cut, onemounted on a MgF2 slide (Moose Hill Enter-prises, Inc., Sperryville, VA) for Raman dataacquisition and at least two others on glassslides for histological staining. The stainedslides were used for pathological confirmationof features observed in the Raman maps. Dur-ing measurements, the tissue was kept moistwith phosphate buffered saline (PBS), pH¼ 7.4.In addition to the Raman micro-images, phasecontrast images of the unstained tissue wererecorded via a CCD camera.

Cell studies were performed using the humancolonic carcinoma cell line HT29 (gift of theHarvard Digestive Diseases Center, Boston,MA). They were grown using high-glucoseDulbecco’s modified Eagle medium (DMEM)supplemented with 10% fetal calf serum, 100 U/ml penicillin, and 100 mg/ml streptomycin (allGibco BRL products, Life Technologies, GrandIsland, NY). Cells were grown to confluency at378C in a humidified atmosphere of 5% CO2 inair and dispersed into suspension using trypsin.Cell suspensions were placed on MgF2 flats,rinsed with PBS, buffered at pH¼ 7.4, andallowed to air dry. Drying of the sample wasnecessary in order to immobilize the cells for theentire mapping experiment. The dried sampleswere then rewet with PBS and Raman mapswere subsequently acquired. Raman imagingmicroscope data collected from the dried cellswere compared to data collected from viablecells still in suspension using a bulk Ramansystem. The spectra acquired from the dried

cells were used to model the spectra obtainedfrom the viable cells. No residual from the modelfit was observed.

Instrumentation

The Raman micro-imaging set-up used tocollect the data for the images presented herewas a point scan system. Raman excitation wasprovided by an argon ion laser-pumped Ti:sap-phire laser (Coherent Innova 90/Spectra Phy-sics 3900S, Coherent, Inc., Santa Clara, CA).Typically 50–150 mW of 830 nm excitation lightwas focused through a microscope objective(63� Zeiss Achroplan, infinity corrected, waterimmersion, numerical aperture 0.9) to a spoton the sample with a diameter of < 2 mm. Theexperimental system has been described pre-viously [Shafer-Peltier et al., 2002]. The spec-tral resolution was �8 cm�1. Spectral maps ofthe tissue were created by raster scanning thetranslation stage (Prior Scientific InstrumentsLtd., Cambridge, MA) under the microscopeobjective. Maps were normally acquired with astep size of 2 mm, consistent with the spatialresolution of the confocal microscope. Althoughdata collection time depended on several userdefined parameters, such as the image step size,number of steps, and spectral acquisition time,an entire Raman image was typically generatedin 2–5 h. A CCD camera atop the microscopeallowed for registration of the focused laser spotwith a white light trans-illuminated or phasecontrast image.

Data Processing

All spectral data processing was preformedusing MATLAB (MathWorks, Inc., Natick, MA).The data were corrected for the spectral re-sponse of the system using a tungsten lightsource and then frequency calibrated using theknown Raman lines of toluene. Cosmic rayswere removed with a derivative filter and thesmall background from the MgF2 flat wassubtracted. Data were then fit with a fourthor fifth order polynomial, which was subtractedfrom the spectrum in order to remove anyfluorescence background. All data was peak-height normalized to one. Finally, MATLABwas used to implement the various data com-pression techniques: PCA, MCR, Euclidean dis-tance, and morphological modeling. Algorithmsfor PCA and ordinary least-squares (used as thefitting algorithm for morphological modeling)already exist in MATLAB, while the algorithm

128 Shafer-Peltier et al.

for MCR was a part of PLS_Toolbox (Eigen-vector Research, Inc., Manson, WA). The purechemicals used for spectroscopic modeling of theHT29 cells: triolein, phosphatidyl choline, cho-lesterol, and DNA (calf thymus), were pur-chased from Sigma (St. Louis, MO) [Buschmanet al., 2001a]. The morphologically derivedspectra used to analyze the artery, breast, andHT29 data have been presented previously[Buschman et al., 2001b; Shafer-Peltier et al.,2002].

In order to obtain improved image contrast asmoothing algorithm based on spatial filteringwas applied to all data presented here [Baxes,1984; Bock and Krischer, 1998]. Spatial filteringrelies on the assumption that adjacent pixels ina digital image contain related information. Agroup of pixels surrounding and including thecentral pixel is called a kernel. We base oursmoothing algorithm on a kernel size of 3� 3.Our algorithm uses a mask that weights thecontributing pixels according to the reciprocal oftheir geometric distance from the center of thekernel. The resultant mask is

where each fraction represents the weight of apixel in the kernel.

RESULTS

Morphological modeling is a powerful tool forcollecting architectural and chemical informa-tion on a small scale. In Figure 1, features suchas the cell membrane, nucleus, and cytoplasmare easily identified when spectra of humancolonic carcinoma cells (HT29) are fit with thepure chemical spectra of phosphatidyl choline(A), DNA (B), cholesterol linoleate (C), triolein(D), and ‘‘cell cytoplasm’’ (E), a morphologicallyderived spectrum developed for thebreast tissuemodel,mostlyactin) [Shafer-Peltieretal., 2002].The spectrum (�) corresponding to the voxelindicated in Figure 1E can be seen in G, alongwith the corresponding fit (�) and residual(below). The fit contributions of the individualmodel elements are also shown. The spectralimages agree with the phase contrast image,demonstrating that using a simple model of fivebasis spectra, it is possible to obtain structuraland chemical information about a sample at the

sub-cellular level. As the cell shown in the imageis evenly bisected by the plane of focus of theconfocal microscope, the cell membrane (mostlyphosphatidyl choline and cholesterol) is observ-ed as a ring structure with the cell cytoplasmand DNA contributions observed clearly asdistinct features within. The average nuclearsize for HT29 cells is 10 mm, consistent with thedimensions provided by the Raman image of thecell DNA content [Wax et al., 2002].

Morphological modeling can be applied tohuman tissue samples as well. Figure 2 showsphase contrast images (2A and G) of a mildlyatherosclerotic artery along with Raman im-ages depicting the distribution of some of themorphological structures (2B–F). The imagesclearly show that the cholesterol (2B), foamcells, and necrotic core (2C) are solely confinedto the intima while the smooth muscle cells (2E)are more prominently found in the media. Thisfinding is consistent with the known architec-ture of atherosclerotic vessels. There is only aslight demarcation between one smooth musclecell and the next because they are so closelyspaced and even overlapping in the media. Theimages demonstrate the high spatial resolutionof this technique and show evidence of fenestra-tion of the elastic lamina, a process known tooccur with the development of atherosclerosis[Braunwald et al., 2001]. The fenestration canbe observed in the Raman image of the internalelastic lamina (IEL), Figure 2D. The smoothmuscle cells, shown in Figure 2E, can be seenmigrating through the break in the IEL into theintima. Smooth muscle cell migration is a char-acteristic of atherosclerotic disease progression.In addition, one can identify a prominentcollagen fiber (2F) in the media atop a diffuseconnective tissue background, a feature that isdifficult to fully appreciate from the phasecontrast image.

Figure 3 shows Raman images of a normalhuman breast duct obtained using a morpholo-gical model [Shafer-Peltier et al., 2002] createdspecifically to analyze breast tissue (Fig. 3A–D).These images can be compared with thosecreated by plotting the intensities of two Ramanbands (Fig. 3E and F) characteristic of the DNAphosphate stretch (1,094 cm�1) and the amide Iband (1,664 cm�1). The morphologically basedRaman images represent the regions wherea particular component (cell cytoplasm (3A),cell nucleus (3B), fat (3C), or collagen (3D))contribute strongly to the spectrum (bright

2/28 3/28 2/283/28 8/28 3/282/28 3/28 2/28

Raman Spectral Imaging 129

regions). Histological analysis of the tissuesample showed a normal breast duct with adiameter of approximately 25 mm. A typicalbreast duct of this size consists of a ring ofepithelial cells surrounded by a basementmembrane (primarily collagen). Within andsurrounding the duct is some fat. The morpho-logical model images clearly show the architec-ture of the duct, whereas the peak heightimages produced using the Raman bands foundat 1,094 and 1,664 cm�1 are much less informa-tive. Although the DNA phosphate stretch

(1,094 cm�1, Fig. 3E) should be found primarilyin cellular regions, while the amide I band(1,664 cm�1, Fig. 3F), indicative of protein,should be found mainly in collagenous regions,the images produced show neither the cellularcomponent nor the collagen as clearly as themorphological model images do. This is becausethe amide I band can be found in many proteins,including those that form the cell cytoskeleton,whereas the phosphate stretch overlaps withbands present in the collagen spectrum. Theinability of peak height analysis to accurately

Fig. 1. Raman images (A–E) of HT29 cells with correspondingphase contrast image (F). Raman spectra are fit with phosphatidylcholine (A), DNA (B), cholesterol linoleate (C), triolein (D), andmorphologically derived cell cytoplasm (E) spectra to produce

chemical maps of the cells. G: shows the spectrum (�) acquiredfrom within the box indicated in image E along with thecorresponding fit (�) and residual (below, with zero line drawn).The fit contributions of each model element are listed to the side.

130 Shafer-Peltier et al.

distinguish morphological features due to spec-tral overlap results in a much less informativeimage.

The Raman spectrum in Figure 3G representsa single point in the Raman image. The spect-rum is a mixture of many chemical components,all of which contribute to the Raman spectrum.

By fitting the spectrum with a morphologicalmodel it is possible to account for the majorspectral features in the data. The residual of thefit, also shown in Figure 3G, is predominatelynoise, indicating that all of the information inthe Raman imaging data hypercube can be re-presented by model-based images.

Fig. 2. Phase contrast images (A and G) of a mildly atherosclerotic artery, with the internal elastic lamina(IEL) and collagen fibers highlighted in G. Also shown are the Raman images of cholesterol (B), foam cells andnecrotic core (C), IEL (D), smooth muscle cells (E), and collagen (F). Key morphological features, such as thefenestration of the IEL can be observed.

Raman Spectral Imaging 131

Although morphological modeling is an effec-tive means of representing Raman images,it requires much advanced knowledge of thesample being studied. As discussed earlier,PCA, MCR, and Euclidean distance can also be

used to compress the data into a manageableform and are much more effective when little isknown about the system. Figure 4 shows a side-by-side comparison of PCA, MCR, Euclideandistance, and morphological modeling. The im-ages, generated from the same data set, are of asample of normal breast tissue containing threeductal units (mostly cells) surrounded by a col-lagen matrix. As can be seen, the images createdby all four techniques are similar. The Eucli-dean distance images are shown as inverses (asthey represent differences from input spectrarather than similarities as the other methodsdo) for easy comparison with the other techni-ques. On the left, the contributions attributableto collagen are shown, while on the right, themore subtle contributions of the cell nucleus(mostly DNA) are displayed. Both PCA (Fig. 4A)and MCR (Fig. 4B) were able to find sevenindependently varying basis spectra. Our com-plete morphological model for breast tissuehas nine basis spectra, however, this includesseveral elements, such as microcalcifications,that are pathologically very important but thatare observed only rarely in human breast tissueand not at all in this specimen [Haka et al.,2002].

The first two principal components, two of thespectra derived using MCR, and the collagenand cell nucleus basis spectra are shown inFigure 4E. The first principal component andthe MCR spectra are similar to the collagenspectrum, the largest contributor to the image.The second principal component and the spec-trum produced by MCR both contain somefeatures of the cell nucleus spectrum (asnegative peaks), but as can be seen from theimage produced (Fig. 4A,B, right), they aremuch less effective at extracting the nuclearcontent within the ductal units than the mor-phological model (Fig. 4D, right). The filled-inrounded shape of the ductal units observed in

Fig. 3. Raman images of normal breast duct based on ordinaryleast-squares fitting of morphologically derived components: cellcytoplasm (A), cell nucleus (B), fat (C), and collagen (D). Images Eand F plot the intensity of single bands: the DNA phosphate(1,094 cm�1) and the protein-based amide I (1,664 cm�1) peaksrespectively. Demonstration of the fitting of a morphologicallybased model (�) to the spectrumof an individualpixel (located in aregion with cellular content) in a Raman image (�) is shown inG.The residual of the fit is plotted below the spectrum (with the zeroline drawn).

132 Shafer-Peltier et al.

Fig. 4. Comparison of four different methods for analyzingRaman images of a region with multiple ductal units, separatedby collagen. The images produced by the fit coefficients of thefirst two principal components are shown in A. B: This shows thetwo corresponding images produced by multivariate curveresolution (MCR). C: This shows images based on Euclideandistance, using the collagen (left) and cell nucleus (right) spectrafrom the morphological model. The images in D are producedusing the fit coefficients produced by ordinary least-squares

fitting with the morphological model, only collagen (left) and cellnucleus (right) are shown, but the complete model was used. E:shows the basis vectors used to create the images, from top tobottom: the first two principal components, the correspondingspectra produced by MCR, the morphologically derived spec-trum of collagen and the morphologically derived spectrum ofthe cell nucleus. The last two spectra were used in both theEuclidean distance measurements and morphological modeling.

Raman Spectral Imaging 133

Figure 4D (right) is consistent with the path-ology of this tissue slice.

Figure 5 shows the normalized fit coefficientsof a particular row of the Raman images used in

Figure 4A–D, left (row indicated in Fig. 5A).Although PCA (~), MCR (&), Euclidean dis-tance (*), and morphological model (X) alldisplay some form of transition from the

Fig. 5. A: Raman image (same as Fig. 4D, left) with third row indicated by white line and (B) heights forcorresponding fit coefficients for the indicated row obtained using the four different models: PCA (~), MCR(&), Euclidean distance (*), and morphological model (X).

134 Shafer-Peltier et al.

collagenous to cellular regions of the tissue,indicated by a change in the intensity of thefit coefficient, the transition is sharpest whenusing the morphological model. Therefore, notonly does the morphological model provideinformation about more of the constituentsof the sample (e.g., cell nucleus), but it alsoproduces images with a higher resolution.

DISCUSSION

The simplest method for displaying a Ramanimage is to plot the intensity of a particularRaman band, or alternatively the ratio of twoRaman bands. This method of analysis onlytakes advantage of a small portion of the dataand because most biological samples containmany compounds with similar spectral fea-tures, is not applicable to biological systems.Spectral overlap makes it difficult to obtainstructural or chemical information about asample from a Raman image based solely onpeak height.

In this article, four techniques which utilizethe full spectrum for creating Raman imagesare presented: PCA, MCR, Euclidean distance,and morphological modeling. These imagingtechniques are applicable not only to Raman,but also to many other spectroscopic imagingtechniques, such as fluorescence. Each techni-que has its advantages and disadvantages.Some require no (PCA) or little (MCR) priorknowledge of the sample being studied, whileothers require some (Euclidean distance) orcomplete (morphological modeling) knowledge.The quality of the images produced is usuallyrelated to how much information is known.

PCA requires the least input from the userand consequently is the best tool for studyingnew types of samples. PCA is used to map outregions based on their spectral variance. Due tothe mathematical process by which they arecreated, the principal components will alwaysexplain all of the spectral features present in thedata. However, as the principal componentsthemselves are mathematical constructs, theycan be difficult or impossible to correlate withknown chemicals. Despite this drawback, infor-mation gained from PCA can be used to buildmore sophisticated models, such as the morpho-logical models developed for breast and arterytissues.

While MCR is also mathematically driven,non-negativity constraints can be applied to

ensure that the basis spectra developed havemore identifiable features than those producedby PCA. In fact, spectra determined using MCRcan be very similar to the true chemical spectra.The disadvantage of MCR is that the morecomplex the system being studied, especially ifthere is much overlap in spectral features, themore difficult it is to perform the analysis. Askilled user can recognize when MCR has failedand adjust the parameters accordingly if thesystem is simple enough, but this too becomesmore challenging as more component spectraare added to the sample mixture. In addition, asmore curves are resolved in a complex system,noise plays a larger and larger role. None-theless, MCR is extremely useful for obtainingspectral lineshapes that can be used to directfurther analysis of a sample.

When some, but not all, of the components of asample are known, Euclidean distance is veryeffective. For example, it is not uncommon tohave a sample in which the spectrum of thespecific chemical being studied is known, butwhere the background chemicals are unknown.In this case, Euclidean distance can map thedistribution of that particular chemical withinthe sample, unencumbered by the lack ofknowledge of the background.

For detailed analysis of a system, especiallyfor producing images with similar informationcontent to pathology slides, morphological mod-eling is the best technique. However, develop-ment of a good morphological model can taketime and requires much data acquisition in itsown right. If the model is incomplete, the imageswill give less accurate information. Therefore,morphological modeling is best used when ex-tensive studies are being performed and modeldevelopment is a part of the experiment.

CONCLUSIONS

Raman spectral imaging is a powerful tool fordetermining chemical information in a biologi-cal specimen. The challenge is to capitalize onall of thespectral information, condensing it intoan image with maximal information content. Inthis article, we introduced a new technique,morphological modeling, and reviewed threeof the more common imaging approaches: PCA,MCR, and Euclidean distance. Each techniquehas its time and place. PCA and MCR are ex-cellent for studying samples about which little isknown a priori, whereas Euclidean distance can

Raman Spectral Imaging 135

produce improved images when some infor-mation about the sample is known. Using amorphological model, it is possible to obtainstructural and chemical information about sub-cellular features, although it is ineffective if thesystem being studied is not well understood.

The ability to combine Raman confocal micro-scopy with imaging modalities to produce im-ages of tissue or cells will be important for futurebiological studies. Although, Raman has onlyrecently begun to be used as a tool for studyingbiological processes, because Raman is non-damaging at low laser powers, it will be usedmore and more to study biological samples,including live cells. Soon researchers will bemonitoring sub-cellular processes using Ramanimaging.

ACKNOWLEDGMENTS

The work was carried out at the MIT LaserBiomedical Research Center supported by NIHP41-RR 02594 grant. We thank PathologyAssociates of University Hospitals for fundingpart of this study. Tissue was provided by theCleveland Clinic Foundation and the Coopera-tive Human Tissue Network of the NationalCancer Institute. HT29 cells were providedby Dr. Kamran Badizadegan (Department ofPathology, Children’s Hospital and HarvardMedical School).

REFERENCES

Andrew JJ, Hancewicz TM. 1998. Rapid analysis of Ramanimage data using two-way multivariate curve resolution.Appl Spec 52:797–807.

Andrew JJ, Browne MA, Clark IE, Hancewicz TM,Millichope AJ. 1998. Raman imaging of emulsion sys-tems. Appl Spec 52:790–796.

Appel R, Zerda TW, Waddell WH. 2000. Raman microima-ging of polymer blends. Appl Spec 54:1559–1566.

Archibald DD, Kays SE, Himmelsbach DS, Barton FE.1998. Raman and NIR spectroscopic methods for deter-mination of total dietary fiber in cereal foods: A com-parative study. Appl Spec 52:22–31.

Arikan S, Sands HS, Rodway RG, Batchelder DN. 2002.Raman spectroscopy and imaging of beta-carotene in livecorpus luteum cells. Anim Reprod Sci 71:249–266.

Baxes GA. 1984. Digital image processing: A practicalprimer. Englewood Cliffs, NJ: Prentice-Hall.

Bock RK, Krischer W. 1998. The data analysis breifbook.Berlin: Springer-Verlag.

Braunwald E, Zipes DP, Libby P. 2001. Heart disease: Atextbook of cardiovascular medicine. New York: W.B.Saunders Company.

Brennan JF, Romer TJ, Lees RS, Tercyak AM, Kramer JR,Feld MS. 1997. Determination of human coronary artery

composition by Raman spectroscopy. Circulation 96:99–105.

Buschman HP, Deinum G, Motz JT, Fitzmaurice M,Kramer JR, van der Laarse A, Bruschke AV, Feld MS.2001a. Raman microspectroscopy of human coronaryatherosclerosis: Biochemical assessment of cellular andextracellular morphologic structures in situ. CardioPathol 10:69–82.

Buschman HP, Motz JT, Deinum G, Romer TJ, FitzmauriceM, Kramer JR, van der Laarse A, Bruschke AV, Feld MS.2001b. Diagnosis of human coronary atherosclerosis bymorphology-based Raman spectroscopy. Cardio Path10:59–68.

Clarke FC, Jamieson MJ, Clark DA, Hammond SV, Jee RD,Moffat AC. 2001. Chemical imaging fusion. The synergyof FT-NIR and Raman mapping microsocpy to enable amore complete visualization of pharamaceutical formula-tions. Anal Chem 73:2213–2220.

Delhaye M, Dhamelincourt P. 1975. Raman microprobe andmicroscope with laser excitation. J Raman Spectrosc 3:33–43.

Freeman TL, Cope SE, Stringer MR, Cruse-Sawyer JE,Brown SB, Batchelder DN, Birbeck K. 1998. Inves-tigation of the subcellular localization of zinc phtha-locyanines by Raman mapping. Appl Spec 52:1257–1263.

Haka AS, Shafer-Peltier KE, Fitzmaurice M, Crowe J,Dasari RR, Feld MS. 2002. Identifying microcalcifica-tions in benign and malignant breast lesions by probingdifferences in their chemical composition using Ramanspectroscopy. Cancer Res 62:5375–5380.

Kneipp K, Haka AS, Kneipp H, Badizadegan K, YoshizawaN, Boone C, Shafer-Peltier KE, Motz JT, Dasari RR, FeldMS. 2002. Surface-enhanced Raman spectroscopy insingle living cells using gold nanoparticles. Appl Spec56:150–154.

McCreery RL. 2000. Raman spectroscopy for chemicalanalysis. New York: John Wiley and Sons.

Nelson MP, Zugates CT, Treado PJ, Casuccio GS, ExlineDL, Schlaegle SF. 2001. Combining Raman chemicalimaging and scanning electron microscopy to character-ize ambient fine particulate matter. Aerosol Sci Tech 34:108–117.

Potter K, Kidder LH, Levin IW, Lewis EN, Spencer RGS.2001. Imaging of collagen and proteoglycan in cartilagesections using Fourier transform infrared spectralimaging. Arthritis Rheum 44:846–855.

Schaeberle MD, Kalasinsky JL, Luke JL, Lewis EN, LevinIW, Treado PJ. 1996. Raman chemical imaging: Histo-pathology of inclusions in human breast tissue. AnalChem 68:1829–1833.

Schaeberle MD, Tuschel DD, Treado PJ. 2001. Ramanchemical imaging of microcrystallinity in silicon semi-conductor devices. Appl Spec 55:257–266.

Shafer-Peltier KE, Haka AS, Fitzmaurice M, Crowe J,Myles J, Dasari RR, Feld MS. 2002. Raman microspec-troscopic model of human breast tissue: Implications forbreast cancer diagnosis in vivo. J Raman Spec 33:552–563.

Tauler R, Smilde AK, Henshaw JM, VBurgess LW,Kowalski BR. 1994. Multicomponent determination ofchlorinated hydrocarbons using a reaction-based chemi-cal sensor. 2. Chemical speciation using multivariatecurve resolution. Anal Chem 66:3337–3344.

136 Shafer-Peltier et al.

Timlin JA, Garden A, Morris MD, Rajachar RM, Kohn DH.2000. Raman spectroscopic imaging markers for fatigue-related microdamage in bovine bone. Anal Chem 72:2229–2236.

Turrell G, Corset J. 1996. Raman spectroscopy develop-ments and applications. London: Academic Press.

van de Poll SWE, Bakker Schut TC, van der Laarse A,Puppels GJ. 2002. In situ investigation of the chemical

composition of ceroid in human athersclerosis by Ramanspectroscopy. J Raman Spec 33:544–551.

Wax A, Yang C, Backman V, Badizadegan K, Boone CW,Dasari RR, Feld MS. 2002. Cellular organization andsubstructure measured using angle-resolved low-coher-ence interferometry. Biophys J 82:2256–2264.

Wold S, Esbensen K, Geladi P. 1987. Principal componentanalysis. Chemometrics Intell Lab Sys 2:37–52.

Raman Spectral Imaging 137