analyst - doyoubuzz · 2014. 8. 26. · transmission spectra recorded h erearelesspronetodistorsion...

Analyst

PAPER

aLaboratory for the Structure and Functio

Structural Biology and Bioinformatics, Uni

Triomphe 2, CP206/2, B1050 Brussels, Belg

2-650-53-82; Tel: +32-2-650-53-86bBreast Cancer Translational Research La

Universite Libre de Bruxelles, Departmen

Brussels, BelgiumcDepartment of Pathology, Institut Jules Bor

† Electronic supplementary informa10.1039/c3an01454a

Cite this: Analyst, 2014, 139, 1044

Received 31st July 2013Accepted 5th December 2013

DOI: 10.1039/c3an01454a

www.rsc.org/analyst

1044 | Analyst, 2014, 139, 1044–1056

Infrared imaging in breast cancer: automated tissuecomponent recognition and spectralcharacterization of breast cancer cells as well as thetumor microenvironment†

Audrey Benard,a Christine Desmedt,b Margarita Smolina,a Philippe Szternfeld,a

Magali Verdonck,a Ghizlane Rouas,b Naima Kheddoumi,b Françoise Rothe,b

Denis Larsimont,c Christos Sotirioub and Erik Goormaghtigh*a

Current evaluation of histological sections of breast cancer samples remains unsatisfactory. The search for

new predictive and prognostic factors is ongoing. Infrared spectroscopy and its potential to probe tissues

and cells at the molecular level without requirement for contrast agents could be an attractive tool for

clinical and diagnostic analysis of breast cancer. In this study, we report the successful application of

FTIR (Fourier transform infrared) imaging for breast tissue component characterization. We show that

specific FTIR spectral signatures can be assigned to the major tissue components of breast tumor

samples. We demonstrate that a tissue component classifier can be built based on a spectral database of

well-annotated tissues and successfully validated on independent breast samples. We also demonstrate

that spectral features can reveal subtle differences within a tissue component, capturing for instance

lymphocytic and stromal activation. By investigating in parallel lymph nodes, tonsils and wound healing

tissues, we prove the uniqueness of the signature of both lymphocytic infiltrate and tumor

microenvironment in the breast disease context. Finally, we demonstrate that the biochemical

information reflected in the epithelial spectra might be clinically relevant for the grading purpose,

suggesting potential to improve breast cancer management in the future.

Introduction

Breast cancer is the second most common cancer aer lungcancer with 10.9% of all cancer diagnosed and 6.1% of all cancerdeaths.1,2 The identication of clinico-pathological parametersand their subsequent integration in cancer management arecrucial for the assessment of patient diagnosis/prognosis as wellas the prediction of therapy response/resistance.3,4 The currentlyused parameters include the histological type and grade, tumorsize, lymph node involvement, vascular invasion as well as theage of the patient. The expression status of hormonal receptorsand the expression/gene amplication status of the HER-2oncogene are both prognostic and predictive factors linked with

n of Biological Membranes, Center for

versite Libre de Bruxelles (ULB), Bld du

ium. E-mail: [email protected]; Fax: +32-

boratory (BCTL), Institut Jules Bordet,

t of Pathology, Institut Jules Bordet,

det, Brussels, Belgium

tion (ESI) available. See DOI:

the efficacy of anti-hormonal and anti-HER2 treatments,respectively.4 Additionally, it is becoming increasingly apparentthat immune regulation and stromal–epithelial cross-talk greatlyinuence the natural course of the tumor and its response tochemotherapy as well as to targeted treatments.5–16 Therefore,several parameters clearly rely on the visualization of the struc-ture and distribution of cellular components in stained tissuesections by a pathologist. Unfortunately, most of these param-eters, such as the histological grade, are associated with inter-and intra-observer discrepancies, in addition to suffer from aquantication problem.3,17

The need for objectivation and quantication of these path-ological parameters has led to a renewal of interest for spectro-scopic approaches and notably for Fourier transform infrared(FTIR) spectroscopy. Based on the absorption of infrared light byvibrational transitions in covalent bonds, FTIR spectroscopyleads to characteristic spectral features related by complexrelationships to the chemical content and conformation of themolecules present in the sample. Since the energy needed for avibrational or rotational transition is highly dependent on thechemical bond environment, the FTIR spectrum is sensitive tothe molecular structure. Therefore, characteristic spectralfeatures can be correlated with biological properties of the

This journal is © The Royal Society of Chemistry 2014

Paper Analyst

sample.18–31 Since it is the intrinsic molecular vibrations thatprobe the chemical composition of the sample, this uniquesignature of the sample is obtained in the absence of anystaining or further processing. By preserving the histo-morpho-logical aspects of the tissue sample but also allowing thedetection and quantication of molecular biochemistry, FTIRimaging also has the advantage of enabling the investigation ofthe geographical heterogeneity of the tissue. This technologycould potentially help automate part of the pathologicalassessment of the tumors, limiting the inter- and intra-observercontributions in the diagnosis and prognosis decisions but alsooff-load pathologists from their enormous daily work. Further-more, given the fact that biochemical changes in cells andtissues most of the time precede morphological alterationswhich are the basis of the histopathology assessment, FTIRimaging could also, through collection of spatially resolvedhigh-quality data, potentially rene the evaluation of variouspathological parameters.

Becoming a new innovative opportunity to analyze tissuespecimens for disease diagnosis/prognosis and progressionmonitoring,32–37 FTIR imaging has already demonstrated itspotential for histopathological issues from which some appli-cations are reported in the literature: characterization of xeno-graed carcinomas,38 as well as bone and cartilage;39 analysis ofbreast lesions,40–42 axillary lymph nodes,43 stem cell identica-tion in the inter-follicular epidermis,44 cell cervical carcinomas45

and colorectal adenocarcinomas.46 These initial data40–42 clearlydemonstrated the potential value of FTIR spectroscopy fordiscriminating benign from malignant breast lesions.

In this study, we developed a tissue component classierbased on a spectral database of histologically well-annotatedbreast tissues. This tissue component classier enables recog-nizing and potentially quantifying the various tissue compo-nents present in a tumor. Since the tumor microenvironmenthas been reported to be an active player in the carcinogenesis,as well as in the efficacy of certain anti-cancer therapies, we alsoinvestigated whether FTIR imaging could identify subtledifferences associated with activation phenotype of lympho-cytes and changes in the extracellular matrix (ECM). Altogether,this study demonstrates the potential added value of automatedFTIR imaging for breast cancer characterization.

Materials and methodsTissue specimens

Breast cancer formalin-xed paraffin-embedded (FFPE) samplesfrom 50 patients were provided by the tumor tissue bank of theJules Bordet Institute (Brussels, Belgium). This project wasapproved by the ethics committee of the Institute. FFPE breastcancer samples from ve patients were considered for thedevelopment of the tissue component classier, including thefurther investigation of the tumor microenvironment (immuneinltration and tumor-associated stroma). Eleven additionalFFPE breast cancer samples were considered for the indepen-dent validation of the tissue component classier. Table 1reports on the clinical and pathological characteristics of thebreast cancer patients considered in this study.


To rene our knowledge on the immune and stromalresponses, seven lymph nodes and three tonsils as well as sevenscars from mastectomy biopsied tissue samples, one cutaneousscar and four biopsies chosen for their extracellular matrixcharacteristic of scar remodelling were analyzed.

A tissue microarray (TMA) including four grade I and eightgrade III invasive breast carcinomas was built. Six breast cancercell lines (BCC) (LY2; SKBR3; T47D; MCF-7; MDA-MB-231;MDA-MB-361) obtained from the American Type CultureCollection (ATCC, Manassas, VA) were used for the comparisonof epithelial spectral features between in vivo and in vitroconditions. Cell culture and processing are reported in the ESI.†FFPE processing of cells in culture has been shown not todegrade the information contained in the FTIR spectra.47

Histological assessment was evaluated by a trained breastcancer pathologist on 3 mm haematoxylin and eosin (H&E) stainedtissue section and grading was assessed based on Nottinghamcriteria. Carcinoma samples were also part of a previous studywhich dened the genomic grade as assessed by microarray tech-nology,48,49 conrming the pathologist histological grade. For eachsample, an adjacent 3 mm section was mounted on a 40 � 26 �2 mm3 barium uoride (BaF2) window (Korth Kristalle GmbH,Germany) for FTIR imaging analysis. These adjacent sections weresubsequently deparaffinized, rehydrated and dried but not stained.

FTIR imaging and data processing

The FTIR data were collected using aHyperion 3000 FTIR imagingsystem (Bruker Optics, Ettlingen, Germany), equipped with a 64�64 Mercury Cadmium Telluride (MCT) Focal Plane Array (FPA)detector (2560 � 2560 mm2) and a 15� objective (NA ¼ 0.4). Thedata were collected in transmission mode from sample regions of170 � 170 mm2. Every element of the FPA acts as an independentand discrete detector from which a full spectrum is obtained. Thecorresponding pixel covers an area of 2.7 � 2.7 mm2. One FTIRimage (unit image) results in 4096 spectra, each one being theaverage of 256 scans recorded in a spectral range from 3900 to800 cm�1 (ca. 5 minutes). To cover larger sample areas, severalFTIR images were juxtaposed in order to obtain one FTIR map. Inthose cases, pixels have been binned so that every detector coversan area of 5.8� 5.8 mm2. The spectral resolution was set to 8 cm�1

and data points encoded every 1 cm�1. All the spectra werepreprocessed as follows.50 The water vapor contribution wassubtracted as described previously51–53 with 1956–1935 cm�1 asthe reference peak. In order to eliminate any intensity variationcaused by changes in the thickness of the tissue section orquantity of the cellular material, the spectra were normalized forequal area between 1725 and 1481 cm�1. An 11-point baselinecorrection was subtracted. Preprocessed spectra were retained forfurther analyses when, on the amide I and II regions (from 1750 to1480 cm�1), the absorbance was superior to�5 and inferior to 120and when the signal-to-noise ratio (S/N) was better than 300 : 1.This ratio is calculated considering the signal as the maximumabsorbance within 1750–1480 cm�1 spectral range and noise asthe standard deviation within 2200–2100 cm�1 range.

Dispersion artefacts leading to dispersive band shapes super-imposed on absorbance features are sometimes observed.

Analyst, 2014, 139, 1044–1056 | 1045

Table 1 Clinical and pathological characteristics of breast cancer patients

Analyst Paper

Transmission spectra recorded here are less prone to distorsion(the so-called Mie scattering) than transection spectra.54 Severaltechniques have been reported to reduce or remove thesedispersion artefacts from infrared micro-spectral data.55–58 Yet,these checks eliminated only a small number of spectra from ourspectral database with marked Mie scattering effects. Correctionfor Mie-scattering or resonant Mie scattering was thus not furtherapplied.

Statistical analysis

We used Principal Component Analysis (PCA) and K-meansclustering to investigate the natural grouping of spectra. PCA isan unsupervised multivariate method enabling variable reduc-tion by building linear combinations of wavenumbers varyingtogether, called the Principal Component (PC).59 The rst prin-cipal component explains most of the data variance. The secondprincipal component, uncorrelated with the rst one, accountsfor most of the residual variance and so on. Usually 2 to 6 PCs aresufficient to explain the major proportion of the original varianceof the dataset, reducing the description of each spectrum to 2 to 6numbers representing the projection (scores) of the spectrum onthe PCs. Based on these scores, the spectra can be presented as apoint on a 2D and 3D PC space. K-means clustering is based on anon-hierarchical process and is particularly efficient for dealingwith large datasets as it is less demanding of computationalresources.60 The process minimizes the intra-cluster variance andmaximizes the inter-cluster variance. The hypotheses on thenumber of clusters have to be made before computation.

MANOVA, the generalized form of the univariate analysis ofvariance (ANOVA) was considered to investigate whether or notthe means were equal or not among BCC spectra.

To develop our tissue component classier, we used thesupervised statistical method Partial Least Squares Discriminant

1046 | Analyst, 2014, 139, 1044–1056

Analysis (PLS-DA) that requires a priori knowledge about theclasses of spectra. It allows both data reduction and discrimi-native investigation. This approach consists of the application ofPLS regression resulting in fewer uncorrelated variables. Linearcombination of variables explaining the membership assign-ment will then serve as discriminant rules, minimizing intra-group and maximizing inter-group separations.

ResultsGeographical distribution of the main tissue components inbreast cancer

Breast tumors are composed of tumor epithelial cells, bro-blasts, immune cells (mainly lymphocytes), extracellular matrix(ECM) and blood vessels, which together play key roles in breasttumor development. Currently, the detection of the varioustissue components is limited in unstained tissues and oenrequires specic immuno-histological staining to identify thespecic cell type. Here we propose an infrared imagingapproach for identication as well as geographical distributionof the main tissue components present in histology sections.

Fig. 1 shows a schematic diagram of the workow presentedin this study. A two-step approach was employed: unsupervisedmethods were rst applied on spectral datasets giving evidenceof the potential of FTIR imaging for tissue component recog-nition in histopathological evaluation and secondly a super-vised approach was used to optimize automated tissuecomponent identication.

Unsupervised classication of breast cancer tissuecomponents by FTIR imaging

Using ca. 13 000 breast cancer tissue spectra (420 FTIR images;from 10 to 50 manually selected spectra per FTIR image)


Fig. 1 The current histopathological gold standard procedure for breast cancer involves morphological assessment of stained tissue sections bya trained pathologist (left). FTIR-based pattern recognition could serve as a complementary tool to standard pathological assessments of thetumor. A two-step approachwas employed in this investigation: unsupervisedmethods of analysis (K-means clustering, PCA and chemical groupmapping) were applied on spectral data recorded on breast cancer tissues for exploring disease characteristics. It resulted in the development ofan extensive database of breast tissue components. However, transferring bench knowledge to practical tool for diagnostic decision requires theuse of a supervised approach. A PLS-DA was then conducted on the database built up on five breast disease samples for discrimination purpose,resulting in an FTIR-based tissue component classifier, which was extensively validated on independent datasets. The FTIR imaging potential forbreast tissue structure discrimination is highlighted by comparing spectral pseudo color maps with pathologist morphological assessment.Further insight was realized on the epithelial grading problem as well as the tumor microenvironment. (A) The 4096 spectra acquired on a breastcancer tissue sample are represented on the entire spectral region (from 3800 to 1000 cm�1) and absorbance values are reported on the Z axis.(B) Photomicrograph of a 550 � 700 mm2 section of the H&E stained breast lobular structure. (C) K-means cluster map acquired from infrareddata of the adjacent unstained tissue section. Each color is associated with a particular cluster. Distances for the clustering were computedacross the 1800–1000 cm�1 region. (D) PCA pseudo color map based on the PC2 accounting for 18% of the data variance. (E) IR image rep-resenting the ratio between protein and lipid contents on the breast sample. The image contrast is provided by the absorbance ratio1656/1468 cm�1. (F) PLS pseudo colormap obtained after applying the PLS-DAmodel on the spectral dataset. The color of the pixel describes theresult of the tissue component classifier assignment: epithelial cells (blue), erythrocytes (cyan), lymphocytes (yellow) and ECM (red). For all ofthose spectral maps, spectra rejected based on the signal to noise ratio, minimum and maximum absorbance criteria are black, light and darkgrey color coded, respectively.

Paper Analyst

acquired from ve patients (see Table 1), we rst investigatedany natural separation that may occur among the variouscomponents of a tumor using an unsupervised PCA method.The results of the PCA were then confronted with the annota-tions [normal/tumoral epithelial cells, ECM, proteic matter(present into the lactiferous ducts), erythrocytes and lympho-cytes – see typical areas in the ESI, Fig. S1†] made by thepathologist on the adjacent H&E slide.


Fig. 2A displays a score plot obtained aer PCA decomposi-tion of the breast spectral database (ca. 13 000 spectra) carriedout on the spectral region (1400–1000 cm�1; dominated by DNA/RNA contributions). The rst two PCs explain more than 90% ofthe total variance of the entire spectral database. Every point inFig. 2A is the projection of a spectrum of the database inPC1–PC2 space. At rst glance, it appears from Fig. 2A that thereis a signicant separation according to the various cell types

Analyst, 2014, 139, 1044–1056 | 1047

Fig. 2 PCA score plots obtained from histologically well-defined spectra. (A) PCA analysis was computed on the breast database considering the1400–1000 cm�1 spectral region for decomposition. Each spectrum is represented as a point in the PC1–PC2 space that account for 90% of thetotal variance of the breast spectral database. Only 12 179 spectra of the database were retained for this PCA decomposition. The spectra ofthe entire database (13 569 spectra) were submitted per class to a 3 PC clean up process in which spectra with projection on PC1 and PC2

components beyond 2 times the standard deviation were rejected (see ESI).† Spectra plotted in (B), (C) and (D) are the same as for (A), except thatfor panel (B), the 1741 lymphocyte spectra from breast tissues were replaced by 1250 spectra recorded in non-active lymph nodes (625 spectra,light purple) and active lymph nodes and tonsils (625 spectra, dark purple); for panel (C), the 6498 ECM spectra acquired from breast tissues werereplaced by 1500 spectra acquired from newly healed scars (750 spectra, light pink) and old scar tissues (750, dark pink); for panel (D), the2660 spectra from healthy and malignant epithelial cells from breast tissues were replaced by 1450 spectra from six breast cancer cell lines inculture (from light to dark blue). The same principal components from panel (A) were used to plot spectra in (B), (C) and (D).

Analyst Paper

and ECM; PC1 explains that the greatest part of the data vari-ance is responsible for the separation between ECM and theother components of the tumor while PC2 highlights separationbetween the erythrocytes, lymphocytes and epithelial cells.These data suggest that the FTIR spectra might contain suffi-cient information to identify the different components presentin breast tissue.

The robustness of this PCA separation clearly appears whenindependent spectra obtained from other tissue types andorgans are analyzed and projected on the PC1–PC2 axes deter-mined in Fig. 2A: lymphocytes from lymph nodes and tonsils

1048 | Analyst, 2014, 139, 1044–1056

(Fig. 2B), ECM from scars (Fig. 2C) and BCC in culture (Fig. 2D).It can be observed that these new spectra fall on the exact spatiallocation on the PCA score plots as their breast tissue counter-parts, demonstrating that the breast tissue component databaseaccounts for specic features independently from the tissuetype, organ or sample. Interestingly, at this level, there seems tobe no clear sub-separation within a given tissue component. Forinstance, no difference can be observed between healthy andmalignant epithelial cells or among carcinoma cell lines: theyare all located in the same PCA score plot area. Similarly, novisible distinction appears among lymphocyte spectra. The


Paper Analyst

origin of the lymphocytes (lymph nodes, tonsils or the breasttumor microenvironment) and the activation state do notemerge on this PCA analysis. Again, spectra acquired fromhealing wounds (still invaded by myobroblasts or not) andtumor reactive stroma cannot be differentiated in Fig. 2. Thisapparent inability to distinguish sub-types within the maintissue components can be either due to the fact that (i) infraredspectra do not contain sufficient signicant features to achieveany further separation or (ii) the scale as well as the spectralrange of investigation are not fully appropriate to achieve such adistinction.

Sub-classication of the major tissue components byunsupervised multivariate analysis through investigation ofother spectral ranges

Lymphocytes. In order to investigate whether the spectralsignature contains sufficient information to discriminate thestate of activation of lymphocytes in immune organs and inbreast tumors, a PCA analysis was computed on spectra recor-ded from seven lymph nodes, three tonsils and breast tumorlymphocytic inltrates from three patients (Fig. 3A). Lympho-cyte spectra were similarly obtained from ve lymph nodes inwhich no lymphocyte activation has been noticed, two lymphnodes and three tonsils showing signs of activation or stimu-lation (germinal centers found around). Five FTIR images perimmune organ and een in total for breast cancer lymphocyticinltrates were recorded; one hundred lymphocyte spectra werethen randomly extracted for each FTIR image, leading to a totalof 2500 spectra from non-activated lymphocytes, 2500 spectrafrom activated lymphocytes (active lymph nodes and tonsils)and 1500 spectra from invasive carcinoma-associated lympho-cytic inltrates. While all lymphocyte spectra are located on thesame region of the score plots in Fig. 2A and B, Fig. 3A revealsdistinct spectral characteristics of lymphocytes based on theiractivation status. While the spectral signature of lymphocyteactivation status is independent of the immune organ of origin,this analysis highlights the distinct spectral signature of tumor-inltrating lymphocytes when compared to those present inlymph nodes or tonsils. Interestingly, the immune response tothe invasive disease has a signicantly distinct spectral signa-ture from the one observed during lymphocyte activation intonsils and lymph nodes.

ECM. Fig. 3B presents a PCA analysis computed on spectrarecorded in twelve scar tissues and tumor reactive stroma fromtwo invasive carcinomas. Seven mastectomy biopsied tissuesamples, one cutaneous scar and four biopsies chosen for theirECM characteristic of scar remodelling were analyzed, resultingin 60 FTIR images (5 per patient). For each measure recorded innewly healed scars (30 FTIR images) and old scar tissues(30 FTIR images), twenty-ve spectra were randomly extracted,leading to a total of 1500 spectra. 527 spectra were manuallyselected in the tumor-associated stromal margins of two gradeIII invasive carcinomas chosen for their large invasive core(about 3 cm) and their tumor-associated ECM in intimatecontact with the carcinoma invasive front. Wound healingreactive stroma (light pink) cannot be distinguished in Fig. 3B


from old scar tissues (dark pink). However, a clear-cut separa-tion is observed in Fig. 3B between tumor-associated stromaand wound healing microenvironment, highlighting thedistinct spectral signature of the tumor reactive stroma.

Epithelial cells. While no visible separation could beobserved between healthy and malignant breast epithelial cellsand between BCC lineages (see Fig. 2), we decided to investigatefurther whether the biochemical information reected in FTIRspectra could be clinically relevant for grading breast carci-nomas as well as differentiating cell lineage. Spectral data wereobtained from healthy epithelial cells (two patients), grade I(four patients) and grade III (eight patients) carcinomas.615 spectra of normal epithelial cells were extracted by manualpixel selection from entire breast samples. For low and highgrades, 22 and 61 FTIR images respectively were recorded froma TMA and twenty-ve epithelial spectra per FTIR image werethen randomly extracted, leading to a total of 550 and1575 spectra. The result of a PCA computed on healthy, grade Iand III epithelial cell spectra and carried out on the 1400–1000 cm�1 spectral range is displayed in Fig. 3C. This analysishighlights, in an unsupervised way, the molecular changesassociated with the carcinogenesis process since an evolutionfrom normal to highly malignant epithelial cells is observed.Additionally, six breast cancer cell lines were considered,leading to 59 FTIR images (from 9 to 10 measures per cell line).Quite interestingly, spectra from these epithelial cells inculture, once formalin-xed and paraffin-embedded and thendeparaffinized according to the protocol used for tissues werefound in the same area of the score plot than the ones found inbreast samples (Fig. 2D). However, once analyzed in the absenceof the other tissue components, the subtle differences amongcell lines appear in a PCA analysis (Fig. 3D). A MANOVA analysisconcluded that every breast cell line present particular spectralfeatures (data not shown).

It can be concluded that evidence for distinct tissuecomponent FTIR signatures appear in unsupervised analysissuch as PCA in the 1400–1000 cm�1 range (Fig. 2). Re-submit-ting every single group alone to a new PCA analysis allows subtledifference to appear. Using other spectral ranges of investiga-tion could also help evidencing sub-classications; for instance,the identication of the specic spectral signature of theimmune (Fig. 3A; 1735–1481 cm�1) and stromal responses tothe invasive disease (Fig. 3B; 1800–1400 cm�1). The spectralrange largely dominated by nucleic acid contributions high-lights carcinogenesis evolution from healthy to malignant state(Fig. 3C; 1400–1000 cm�1). Such a hierarchical use of differentspectral ranges is common in the eld of bacteria.18,26,29

Classication of the major tissue components by supervisedmultivariate analysis

The classications described so far were completely unsuper-vised. They have the advantage of highlighting the naturalseparation between spectra of tissue components but they arenot optimized to achieve a specic classication. However,precise diagnostics require as little false positive and falsenegative as possible. Here, a Partial Least Squares Discriminant

Analyst, 2014, 139, 1044–1056 | 1049

Fig. 3 Sub-classification of the major breast tissue components by an unsupervised statistical method (PCA). Each point of these score plots isthe projection of a spectrum in a 3D-PC space. (A) Lymphocytes: non-active lymph node lymphocytes (2500 spectra, light purple), active lymphnode and tonsil lymphocytes (2500 spectra, dark purple) and invasive carcinoma-associated lymphocytic infiltrates (1500 spectra, blue). PCAdecomposition was computed on the amide I & II spectral regions (from 1735 to 1481 cm�1) containing the most discriminant wavenumbers forthe distinction between activated and non activated lymphocytes as revealed by a Student's statistical test (data not shown). (B) ECM: newlyhealed scars (750 spectra, light pink), old scar tissues (750, dark pink) and tumor microenvironment (527 spectra, orange). The protein spectralregion (from 1800 to 1400 cm�1) was considered for the PCA decomposition. (C) Breast carcinoma staging: healthy epithelial cells (615 spectra,blue), grade I (550 spectra, green) and grade III (1575 spectra, red). PCA decomposition was carried out in the DNA/RNA spectral region, from1400 to 1000 cm�1. (D) BCC phenotype: MCF-7, its antiestrogen variant LY2, T47D, SKBR3, MDA-MB-361 and -231. For each cell line, nine to tenFTIRmeasures were considered and in total, 59mean spectra are plotted on the PCA score plot. PCA decomposition was carried out in the entirefingerprint region, from 1800 to 1000 cm�1.

Analyst Paper

Analysis (PLS-DA) was conducted on the database, describedabove. Two-third of the database was used as the training set tobuild the recognition model and the remaining third as the testset from which sensitivity, accuracy and specicity of thediscriminative tool are assessed. A bootstrapping procedureallowed repeating the process by selecting randomly a trainingset and applying the resulting prediction model to theremaining spectra that did not participate in the building ofthe model. Table 2 summarizes the results of this analysis. The

1050 | Analyst, 2014, 139, 1044–1056

diagonal reports the percentages of correctly assigned spectra;other values reported in a same line or column are the falsenegative and false positive percentages and represent respec-tively the sensitivity and specicity of the PLS classication.Almost 98% of epithelial spectra tested were well recognized bythe prediction model with only 4.8% of false positive and 2.1%of false negative. The poorest sensitivity of the predictionmodel is observed for lymphocytes that exhibit the highestfalse negative rates (5.8%) with most of the misclassied


Table 2 Result of the PLS-DA classification computed on the breast spectral database. Predicted values are expressed as % of the true values.Two-third of the spectral database was used for the prediction model training. The remaining third was used as a validation set. This procedurewas repeated 100 times, yielding a standard deviation for each value. The spectral range of investigation is the concatenation of these twospectral ranges 1800–1600 cm�1 & 1500–1000 cm�1. 2660manually selected spectra of epithelial cells, 1741 of lymphocytes, 1023 arising fromerythrocytes and 6498 spectra from ECM were considered for the PLS-DA training

PLS-DA (in %) Predicted Epithelial cells Erythrocytes Lymphocytes ECM False negatives False positives

Epithelial cells 97.9 � 0.7 0.2 � 0.2 1.8 � 0.7 0.1 � 0.2 2.1 4.8Erythrocytes 3.3 � 0.9 96.4 � 1.0 0.3 � 0.3 0.0 � 0.0 3.6 0.6Lymphocytes 5.5 � 1.1 0.0 � 0.1 94.1 � 1.2 0.3 � 0.3 5.8 2.9ECM 0.0 � 0.0 0.0 � 0.0 0.0 � 0.0 100.0 � 0.0 0 0.1

Paper Analyst

lymphocyte spectra recognized by the model as epithelial cells.An interval PLS-DA analysis was run in order to investigate theimplication of the different spectral ranges in the classicationobtained. During the analysis, spectral regions of consecutive100 cm�1 (from 2000 to 901 cm�1) were challenged as con-taining information for tissue component discrimination byPLS-DA. Success rates of the different prediction models arereported according to the different components of the breasttumor in Fig. 4. Several conclusions can be pointed out fromthis analysis. First, all spectral ranges tested between 1800 and900 cm�1 were found to contain the information pertinent forthe classication of the four different tissue components. Thisindicates that the discriminant information is present in alltypes of cell molecules and not only in proteins (which absorbessentially between 1800 and 1400 cm�1) or nucleic acids(having their major contribution in the 1400–1000 cm�1).

Fig. 4 Success rate (in %) of the prediction models as a function of the spwas computed on the breast spectral database. The percentages of correpredicted values are expressed as % of the true values. As for the PLS-training and the remaining third as the test set. The procedure was repeatspectral range between 2000 and 901 cm�1 was investigated by steps ospectra of epithelial cells, 1741 of lymphocytes, 1023 arising from erythtraining.


Second, since spectra were normalized for equal area on the1725–1481 cm�1 range, discriminative information for classi-cation contained in this spectral range can only be due tosubtle spectral shape variations; for instance, an overall changein the secondary structure of proteins. Third, considering allthe spectral ranges together from 1800 to 900 cm�1 yields abetter prediction than any isolated sub range. For instance,none of the 100 cm�1 sub ranges reported on Fig. 4 allows a90% correct prediction for lymphocytes, while taken together, asuccess rate of 94% was reached (Table 2). It reveals that thedifferent sub ranges contain non-redundant information thatcan be exploited together for building a more efficient predic-tion model. However, a systematic analysis showed that theconcatenation of 1800–1600 & 1500–1000 cm�1 were the mostsuited spectral ranges for tissue component discrimination(data not shown).

ectral range selected (see color legend). Interval PLS-DA classificationct assignment by the prediction models are reported on the Y axis; theDA classification, two-third of the spectral database was used for theed 100 times, yielding a standard deviation for each % value. The entiref 100 cm�1 as indicated in the color legend. 2660 manually selected

rocytes and 6498 spectra from ECM were considered for the PLS-DA

Analyst, 2014, 139, 1044–1056 | 1051

Analyst Paper

Independent validation of the tissue component classier

A completely independent data subset of eleven breast cancerswas used for validation (Table 1). In total, more than 270 000spectra from 44 FTIR spectral maps (four per patient) wererecorded from these new cases and submitted to the tissuecomponent classier. A spectral map usually comprisedbetween 1024 to 9216 spectra taken from tissue regions between184 and 552 mm length. Results of the classication are bestreported as color-coded images as the classication processapplied to every pixel of the FTIR image determines themembership probability of the unknown spectral prole andassigns it to the histological class for which the higher proba-bility is observed. It leads to pseudo color maps that, correlatedwith morphological assessment of the pathologist, provide aqualitative measure of the classication accuracy. Three exam-ples are discussed below.

The le panel of Fig. 5 compares the photomicrograph of anH&E stained grade II invasive adenocarcinoma (A) and thecorresponding color-coded FTIR image based on the PLSclassication model (B). Four major tissue components arevisible on the stained histological structure: (A1) healthyepithelium bilayer, (A2) connective tissue, (A3) lymphocytessituated all around the healthy epithelium structure and (A4)some malignant cells inltrating the surrounding breastconnective tissue. It can be observed that the assignmentprovided by the PLS-DAmodel on the adjacent unstained tissuesection displays a high degree of correspondence with H&Emorphological assessment, demonstrating the robustness ofthe method. Importantly, the inltrative malignant cells arewell recognized by the classier, presenting the FTIR-basedclassier as a tool of choice for the study of the inltratingphase of the carcinoma evolution.

The right panel of Fig. 5 compares the histomorphologicalfeatures observed on a stained low grade invasive adenocarci-noma (C) with the FTIR-based histology obtained aer applyingthe PLS classier on the spectral map (D). The region encom-passes a lactiferous duct presenting both epithelial hyperplasia(C1) and micropapillary clusters of carcinoma cells (C2). Thepresence of erythrocytes (C3) in the lactiferous duct has to berelated to some bleeding associated with the cancer disease. Acapillary blood vessel (C4) is also observed surrounded by theECM (C5). Again, a very good correlation is observed since thehistomorphological features highlighted in the H&E stainedsection are well preserved by the FTIR-based assignments.

The bottom panel of Fig. 5 describes the potential of FTIRimaging as an efficient tool for highlighting immune inltra-tion in tumor malignancy. The photomicrograph of a highgrade invasive adenocarcinoma undergoing immune response(E) is compared to the FTIR-based pseudo color map (F). A largelymphocyte inltration (E1) is separated from the invasivecarcinoma mass (E2) by collagen bers (E3). Some invasivemalignant cells (E4) are observed among tumor-inltratedlymphocytes. Despite the fact that few collagen bers werereported to delimitate the junction between invasive breastcarcinoma and lymphocyte response, red colored pixels areobserved in the pseudo color map, demonstrating the

1052 | Analyst, 2014, 139, 1044–1056

robustness of the FTIR-based histology for highlighting thevarious histological features of the tumor and itsmicroenvironment.

Discussion

Breast cancer is a public health problem. A better under-standing and characterization of the heterogeneous nature ofthe tumor and its microenvironment is essential to individu-alize breast cancer management.

The ability to obtain spatially resolved data has opened thedoor to applications of FTIR spectroscopy in pathology byenabling to take into account the heterogeneity of tissuesamples. The inherent chemical differences between cells of thetissue give rise to molecule-specic vibrational signaturespresenting the pathologists a complementary tool for histo-pathological examinations. Automated histopathology recogni-tion may benet the medical community by providing moreaccurate and reproducible readout of the cellular andmolecularcomposition of the tumor, potentially rening the readingsfrom the pathologists. Furthermore, FTIR pattern recognitioncan be fully automated, so that non trained users can interpretthese pseudo color images.

In this study we showed the potential of FTIR imaging forautomation of the histopathology procedure by tissue compo-nent classication and sub-classication as well as demon-strating specic stromal and immune responses to cancerdisease.

Principal component analyses show that the ve main tissuecomponents found on breast carcinomas histological sliceshave infrared spectra that display a clear trend to separate alongthe rst two PCs, i.e. accounting for the largest variance presentin the dataset (Fig. 2). Importantly, this separation is totallyunsupervised, occurring without any a priori knowledge aboutthe tissue structure and composition. Furthermore, the resultsindicate that infrared spectra of each component grouptogether according to their cellular nature and not according totheir tissue of origin. For instance, epithelial cells are all foundin the same area of the PCA score plot (Fig. 2), independently onthe patient and independently on whether they are from healthystructures or various carcinoma types. More unexpectedly, thesix breast cancer cell lines in culture display the same spectralsignature as epithelial structures from tissue biopsies bysharing the exact same location of the PCA score plot (Fig. 2D).So, spectroscopically, breast cancer cell lines in culture are well-suited tumor models. Again, lymphocyte spectra gathertogether during a PCA whether they are activated or not andinterestingly whether they are acquired from breast tumors,lymph nodes or tonsils (Fig. 2A & B). Finally, whether acquiredfrom the tumormicroenvironment or scars, ECM spectra form asingle group distinct from the others. It therefore appears thateven unsupervised classication provides robust clusteringaccording to the cell type or ECM.

The results presented here also show that spreading thescale of the PCA score plot by submitting each of these groups toa new PCA in the absence of the others, and potentially inother spectral ranges, resulted in clear separation trends,


Fig. 5 FTIR-based histopathological characterization of three breast tissue regions based on the PLS-DA classification model during the inde-pendent validation. Comparison of the histological features observed on the 3 mm thick H&E stained tissue sections and the infrared-basedhistological assignment by the PLS-DA model applied on the adjacent unstained tissue section. (A) Breast invasive adenocarcinoma tissue section(1114� 371 mm2). (A1) Healthy epithelium bilayer, (A2) connective tissue, (A3) lymphocytes and (A4) malignant cells infiltrating the connective tissue.(B) FTIR pseudo color map obtained after applying the PLS-DA model on the 12 288 spectra mapping recorded on the next adjacent unstainedslice. (C) Lactiferous duct with epithelial hyperplasia andmicropapillary clusters. The region outlined defines the area selected for FTIR investigation(371 � 557 mm2). (C1) Epithelial hyperplasia, (C2) micropapillary clusters of carcinoma cells, (C3) erythrocytes, (C4) capillary blood vessels and (C5)ECM. (D) The corresponding 6144 spectral map assigned by the PLS-DA model. (E) Grade III invasive adenocarcinoma undergoing immuneinfiltration (557 � 371 mm2). (E1) Lymphocyte infiltration, (E2) invasive carcinoma mass, (E3) collagen fibers and (E4) invasive malignant cells sur-rounded by lymphocytes. (F) FTIR-based assignment on the corresponding 6144 spectral map. All photomicrographs were taken with a 40�magnification. The color coding provided represents the class assignment of the PLS-DA model. Spectra with S/N ratio below threshold (S/N <500) were turned to black while spectra not passing the minimum and maximum quality test are respectively light and dark grey colour coded.

Paper Analyst

highlighting the potential of FTIR spectroscopy for the investi-gation of subtle biochemical changes associated for examplewith lymphocyte activation and reactive stroma. Also, healthyepithelial cells could clearly be distinguished from the malig-nant ones (Fig. 3C).61 However, only an imperfect separationcould be observable between histological grade I and III tumors.Yet, no rm conclusion can be drawn at this stage concerningthe potential of FTIR imaging for identifying and gradingvarious sub-types of carcinoma cells. A much larger set of breastcarcinoma patients should be investigated in future studies.


With regard to tumor-inltrating lymphocytes, they showeda distinct spectral signature from those present in lymph nodesand tonsils. Also, lymph node and tonsil lymphocytes weredemonstrated to be correlated with their state of activation,independently from their origin. Interestingly while in thepresent work, lymphocyte activation was triggered by some kindof infection, it has been demonstrated elsewhere that infraredspectra of B-lymphocytes from lymph nodes partially invaded bymetastases cluster according to the primary cancer type and notaccording to the patient.62 This is important as lymphocytes

Analyst, 2014, 139, 1044–1056 | 1053

Analyst Paper

present in breast tumors have a spectrum distinct from bothactivated and non-activated lymphocytes from lymph nodes andtonsils, possibly suggesting a specic effect of the breast tumor.It is also interesting to note that B- and T-lymphocytes can bedifferentiated by FTIR spectroscopy.63 Finally, although newlyhealed scar tissues are characterized inter alia by an increasednumber of myobroblasts, the spectral signature obtained fromthese wound healing reactive stroma could not be distinguishedfrom the one obtained in old scar tissues in which myobro-blast apoptosis and normal phenotype restoration greatlydecrease the broblast presence (Student's t-test, data notshown). However, though cancers are sometimes described aswounds that do not heal,12,13,64,65 the stromal response todeveloping tumor presents a spectral signature distinct fromthe one observed in wound healing stroma (Fig. 3B), suggestingthat specicities of the ECM close to the tumor margin could beinvestigated with this approach. Correlating those signatures topatient outcome and therapeutic response could help betterunderstand the immune and stromal implication into breastdisease.

The supervised approach dedicated to tissue componentdiscrimination achieves a much more better separation amongthe different components of breast cancer tissue. ECM isrecognized with 100% accuracy and epithelial cells with close to98%, followed by erythrocytes at 96%. Lymphocytes have adecent score of 94%, and the misclassied spectra are recog-nized as epithelial cells (5%). Importantly, the validation con-ducted on distinct eleven clinical cases demonstrated thatpatient-to-patient variations are smaller than those due to thedifferent tissue components and that the associated spectralngerprints could be accurately recognized by the tissuecomponent classier.

Conclusion

The present work demonstrates the potential of FTIR imagingfor automated tissue component assignment using the chem-ical composition specicity of the different breast tissuecomponents reected in their infrared spectrum. FTIR imaginghas signicant assets: (i) neither staining of the tissue slices orchemical reagent addition are necessary for sample preparationas intrinsic molecular vibrations probe the chemical composi-tion and structural properties (e.g. protein structure) of thesample;51,66,67 (ii) FTIR data can be collected and interpretedwithin minutes, promoting FTIR imaging as a rapid technique;(iii) as this technology is non-invasive and non-destructive,additional analyses like immunohistochemistry staining arepossible in a second step on the same tissue slice; (iv) the lowcost of this approach may facilitate its access and incorporationas routine medical tools and nally (v) the FTIR-basedpathology approach can objectively detect and quantify eachtissue component class by counting the assigned pixels and so,the covered area. This could be really useful for the investiga-tion and quantication of lymphocytic inltration, which hasrecently been shown to play an important role in breast cancerprognosis.6,7,68–70

1054 | Analyst, 2014, 139, 1044–1056

In conclusion, we demonstrate here that FTIR imagingthrough collection of spatially resolved high-quality data allowsthe investigation of the subtle alterations in biochemical compo-sition of the tissue associated with the carcinogenesis process.

Abbreviations

3D
Three dimensional; BaF2 Barium uoride; BCC Breast cancer cell line; DNA Deoxyribonucleic acid; ECM Extracellular matrix; FFPE Formalin-xed, paraffin-embedded; FPA Focal plane array; FTIR Fourier transform infrared spectroscopy; H&E Hematoxylin–eosin; LI Lymphocytic inltrate; MANOVA Multivariate analysis of variance; MCT Mercury cadmium telluride; PC Principal component; PCA Principal component analysis; PLS-DA Partial least square discriminant analysis; RNA Ribonucleic acid; S/N Signal/noise ratio; TMA Tissue microarrays; TS Tumor stroma.
Acknowledgements

Grant numbers and sources of support: this research has beensupported by grants from the National Fund for ScienticResearch (FRFC 2.4533.10, 2.4527.10, 2.4526.12 and PDRT.0155.13). A.B. was supported by the Fund for Research andEducation within Industry and Agriculture (FRIA) from theFNRS (Belgium) and David & Alice Van Buuren fund. She is nowScientic Collaborator at the Cancer Centre in the ScienticInstitute for Public Health (WIV-ISP), Brussels, Belgium. E.G. isDirector of Research with the National Fund for ScienticResearch (FNRS) (Belgium), and M.V. is a Research Fellowsupported by the Fund for Research and Education withinIndustry and Agriculture (FRIA) from the FNRS (Belgium).

References

1 J. Ferlay, H. R. Shin, F. Bray, D. Forman, C. Mathers andD. M. Parkin, Int. J. Cancer, 2010, 127, 2893–2917.

2 D. M. Parkin, F. Bray, J. Ferlay and P. Pisani, Int. J. Cancer,2001, 94, 153–156.

3 J. S. Meyer, C. Alvarez, C. Milikowski, N. Olson, I. Russo,J. Russo, A. Glass, B. A. Zehnbauer, K. Lister andR. Parwaresch, Mod. Pathol., 2005, 18, 1067–1078.

4 S. E. Pinder, I. O. Ellis and C. W. Elston, J. Clin. Pathol., 1995,48, 981–983.

5 C. Desmedt, B. Haibe-Kains, P. Wirapati, M. Buyse,D. Larsimont, G. Bontempi, M. Delorenzi, M. Piccart andC. Sotiriou, Clin. Cancer Res., 2008, 14, 5158–5165.


Paper Analyst

6 R. D. Schreiber, L. J. Old and M. J. Smyth, Science, 2011, 331,1565–1570.

7 C. Denkert, S. Loibl, A. Noske, M. Roller, B. M. Muller,M. Komor, J. Budczies, S. Darb-Esfahani, R. Kronenwett,C. Hanusch, C. von Torne, W. Weichert, K. Engels,C. Solbach, I. Schrader, M. Dietel and G. von Minckwitz,J. Clin. Oncol., 2010, 28, 105–113.

8 G. M. Clark, J. Natl. Cancer Inst. Monogr., 2001, 30, 17–21.9 C. W. Elston, I. O. Ellis and S. E. Pinder, Crit. Rev. Oncol.Hematol., 1999, 31, 209–223.

10 M. A. Cichon, A. C. Degnim, D. W. Visscher andD. C. Radisky, J. Mammary Gland Biol. Neoplasia, 2010, 15,389–397.

11 M. J. Bissell, D. C. Radisky, A. Rizki, V. M. Weaver andO. W. Petersen, Differentiation, 2002, 70, 537–546.

12 E. S. Radisky and D. C. Radisky, Rev. Endocr. Metab. Disord.,2007, 8, 279–287.

13 R. Kalluri and M. Zeisberg, Nat. Rev. Cancer, 2006, 6, 392–401.

14 P. Micke and A. Ostman, Lung Cancer, 2004, 45(suppl. 2),S163–S175.

15 S. Loi, N. Sirtaine, F. Piette, R. Salgado, G. Viale, F. VanEenoo, G. Rouas, P. Francis, J. P. Crown, E. Hitre, E. deAzambuja, E. Quinaux, A. Di Leo, S. Michiels,M. J. Piccart and C. Sotiriou, J. Clin. Oncol., 2013, 31, 860–867.

16 M. Ignatiadis, S. K. Singhal, C. Desmedt, B. Haibe-Kains,C. Criscitiello, F. Andre, S. Loi, M. Piccart, S. Michiels andC. Sotiriou, J. Clin. Oncol., 2012, 30, 1996–2004.

17 P. Robbins, S. Pinder, N. de Klerk, H. Dawkins, J. Harvey,G. Sterrett, I. Ellis and C. Elston, Hum. Pathol., 1995, 26,873–879.

18 L. P. Choo-Smith, K. Maquelin, T. van Vreeswijk,H. A. Bruining, G. J. Puppels, N. A. Ngo Thi, C. Kirschner,D. Naumann, D. Ami, A. M. Villa, F. Orsini, S. M. Doglia,H. Lamfarraj, G. D. Sockalingum, M. Manfait, P. Allouchand H. P. Endtz, Appl. Environ. Microbiol., 2001, 67, 1461–1469.

19 A. Derenne, R. Gasper and E. Goormaghtigh, Analyst, 2011,136, 1134–1141.

20 A. Gaigneaux, J. M. Ruysschaert and E. Goormaghtigh, Eur. J.Biochem., 2002, 269, 1968–1973.

21 A. Gaigneaux, C. Decaestecker, I. Camby, T. Mijatovic,R. Kiss, J. M. Ruysschaert and E. Goormaghtigh, Exp. CellRes., 2004, 297, 294–301.

22 R. Gasper, J. Dewelle, R. Kiss, T. Mijatovic andE. Goormaghtigh, Biochim. Biophys. Acta, 2009, 1788, 1263–1270.

23 R. Gasper, T. Mijatovic, A. Benard, A. Derenne, R. Kiss andE. Goormaghtigh, Biochim. Biophys. Acta, 2010, 1802, 1087–1094.

24 R. Gasper, T. Mijatovic, R. Kiss and E. Goormaghtigh, Appl.Spectrosc., 2011, 65, 584–594.

25 D. M. Haaland, D. T. J. Howland and E. V. Thomas, Appl.Spectrosc., 1997, 51, 340–345.

26 C. Kirschner, K. Maquelin, P. Pina, N. A. Ngo Thi, L. P. Choo-Smith, G. D. Sockalingum, C. Sandt, D. Ami, F. Orsini,


S. M. Doglia, P. Allouch, M. Mainfait, G. J. Puppels andD. Naumann, J. Clin. Microbiol., 2001, 39, 1763–1770.

27 D. C. Malins, N. L. Polissar and S. J. Gunselman, Proc. Natl.Acad. Sci. U. S. A., 1997, 94, 259–264.

28 L. M. McIntosh, M. Jackson, H. H. Mantsch, M. F. Stranc,D. Pilavdzic and A. N. Crowson, J. Invest. Dermatol., 1999,112, 951–956.

29 D. Naumann, D. Helm and H. Labischinski, Nature, 1991,351, 81–82.

30 B. Rigas, S. Morgello, I. S. Goldman and P. T. Wong, Proc.Natl. Acad. Sci. U. S. A., 1990, 87, 8140–8144.

31 B. Rigas and P. T. Wong, Cancer Res., 1992, 52, 84–88.32 R. Bhargava, Anal. Bioanal. Chem., 2007, 389, 1155–1169.33 M. Diem, M. Romeo, S. Boydston-White, M. Miljkovic and

C. Matthaus, Analyst, 2004, 129, 880–885.34 D. C. Fernandez, R. Bhargava, S. M. Hewitt and I. W. Levin,

Nat. Biotechnol., 2005, 23, 469–474.35 P. Garidel and M. Boese, Microsc. Res. Tech., 2007, 70, 336–

349.36 P. Lasch, M. Diem, W. Hansch and D. Naumann, J. Chemom.,

2007, 20, 209–220.37 I. W. Levin and R. Bhargava, Annu. Rev. Phys. Chem., 2005, 56,

429–474.38 R. Wolthuis, A. Travo, C. Nicolet, A. Neuville, M. P. Gaub,

D. Guenot, E. Ly, M. Manfait, P. Jeannesson and O. Piot,Anal. Chem., 2008, 80, 8461–8469.

39 A. Boskey and C. N. Pleshko, Biomaterials, 2007, 28, 2465–2478.

40 H. Fabian, P. Lasch, M. Boese and W. Haensch, Biopolymers,2002, 67, 354–357.

41 H. Fabian, P. Lasch, M. Boese and W. Haensch, J. Mol.Struct., 2003, 661–662, 411–417.

42 H. Fabian, N. A. Thi, M. Eiden, P. Lasch, J. Schmitt andD. Naumann, Biochim. Biophys. Acta, 2006, 1758, 874–882.

43 B. Bird, M. Miljkovic, M. J. Romeo, J. Smith, N. Stone,M. W. George and M. Diem, BMC Clin. Pathol., 2008, 8, 8.

44 I. I. Patel, W. J. Harrison, J. G. Kerns, J. Filik, K. Wehbe,P. L. Carmichael, A. D. Scott, M. P. Philpott, M. D. Frogley,G. Cinque and F. L. Martin, Anal. Bioanal. Chem., 2012,404, 1745–1758.

45 W. Steller, J. Einenkel, L. C. Horn, U. D. Braumann,H. Binder, R. Salzer and C. Kra, Anal. Bioanal. Chem.,2006, 384, 145–154.

46 P. Lasch, W. Haensch, D. Naumann and M. Diem, Biochim.Biophys. Acta, 2004, 1688, 176–186.

47 M. Verdonck, N. Wald, J. Janssis, P. Yan, C. Meyer, A. Legat,D. E. Speiser, C. Desmedt, D. Larsimont, C. Sotiriou andE. Goormaghtigh, Analyst, 2013, 138, 4083–4091.

48 C. Sotiriou, P. Wirapati, S. Loi, A. Harris, S. Fox, J. Smeds,H. Nordgren, P. Farmer, V. Praz, B. Haibe-Kains,C. Desmedt, D. Larsimont, F. Cardoso, H. Peterse,D. Nuyten, M. Buyse, M. J. Van de Vijver, J. Bergh, M. Piccartand M. Delorenzi, J. Natl. Cancer Inst., 2006, 98, 262–272.

49 C. Sotiriou and C. Desmedt, Ann. Oncol., 2006, 17(suppl. 10),259–262.

50 A. Benard and E. Goormaghtigh, Advances in BiomedicalSpectroscopy, IOS Press, 2012, 6, 195–211.

Analyst, 2014, 139, 1044–1056 | 1055

Analyst Paper

51 E. Goormaghtigh, V. Cabiaux and J. M. Ruysschaert, Subcell.Biochem., 1994, 23, 363–403.

52 E. Goormaghtigh and J. M. Ruysschaert, Spectrochim. Acta,Part A, 1994, 50, 2137–2144.

53 E. Goormaghtigh, Adv. Biomed. Spectrosc., 2009, 2, 104–128.

54 K. Wehbe, J. Filik, M. D. Frogley and G. Cinque, Anal.Bioanal. Chem., 2013, 405, 1311–1324.

55 B. Bird, M. Miljkovic and M. Diem, J. Biophotonics, 2010, 3,597–608.

56 K. R. Bambery, B. R. Wood and D. McNaughton, Analyst,2012, 137, 126–132.

57 P. Bassan, A. Sachdeva, A. Kohler, C. Hughes, A. Henderson,J. Boyle, J. H. Shanks, M. Brown, N. W. Clarke andP. Gardner, Analyst, 2012, 137, 1370–1377.

58 A. Dazzi, A. Deniset-Besseau and P. Lasch, Analyst, 2013, 138,4191–4201.

59 R. A. Johnson and D. W. Wichern, Applied multivariatestatistical analysis, Prentice Hall, New Jersey, 1998.

60 L. Wang and B. Mizaikoff, Anal. Bioanal. Chem., 2008, 391,1641–1654.

1056 | Analyst, 2014, 139, 1044–1056

61 A. Benard, C. Desmedt, V. Durbecq, G. Rouas, D. Larsimont,C. Sotiriou and E. Goormaghtigh, Spectroscopy, 2010, 24, 67–72.

62 A. I. Mazur, J. L. Monahan, M. Miljkovic, N. Laver, M. Diemand B. Bird, J. Biophotonics, 2013, 6, 101–109.

63 C. Kra, R. Salzer, G. Soff and M. Meyer-Hermann,Cytometry, Part A, 2005, 64, 53–61.

64 O. De Wever, P. Demetter, M. Mareel and M. Bracke, Int. J.Cancer, 2008, 123, 2229–2238.

65 O. E. Franco, A. K. Shaw, D. W. Strand and S. W. Hayward,Semin. Cell Dev. Biol., 2010, 21, 33–39.

66 E. Goormaghtigh, V. Raussens and J. M. Ruysschaert,Biochim. Biophys. Acta, 1999, 1422, 105–185.

67 E. Goormaghtigh, J. M. Ruysschaert and V. Raussens,Biophys. J., 2006, 90, 2946–2957.

68 M. L. Disis, J. Clin. Oncol., 2010, 28, 4531–4538.69 S. M. Mahmoud, E. C. Paish, D. G. Powe, R. D. Macmillan,

M. J. Grainge, A. H. Lee, I. O. Ellis and A. R. Green, J. Clin.Oncol., 2011, 29, 1949–1955.

70 R. Mouawad, J. P. Spano and D. Khayat, J. Clin. Oncol., 2011,29, 1935–1936.


analyst - doyoubuzz · 2014. 8. 26. · transmission spectra recorded h erearelesspronetodistorsion...

Documents