journal of biophotonics - ubc physics & astronomyhzeng/s/zeng lung clinical paper.pdf ·...

www. biophotonics-journal.orgJournal of

BIOPHOTONICS

REPRINT

FULL ARTICLE

Real-time endoscopic Raman spectroscopyfor in vivo early lung cancer detection

Hanna C. McGregor1, 2, Michael A. Short1, Annette McWilliams1, Tawimas Shaipanich1, 4,Diana N. Ionescu3, Jianhua Zhao1, Wenbo Wang1, Guannan Chen1, Stephen Lam1, 2, 4,and Haishan Zeng*, 1, 2, 3

1 ImagingUnit – IntegrativeOncologyDepartment, BCCancerAgency ResearchCentre, Vancouver, British Columbia, Canada2 Interdisciplinary Oncology Program, University of British Columbia, Vancouver, British Columbia, Canada3 Department of Pathology and Laboratory Medicine, University of British Columbia, Vancouver, British Columbia, Canada4 Department of Medicine, University of British Columbia, Vancouver, British Columbia, Canada

Received 30 July 2015, revised 22 October 2015, accepted 24 November 2015Published online 9 January 2016

Key words: in vivo tissue Raman spectroscopy, endoscopy, lung cancer, early cancer detection, spectral diagnosis

1. Introduction

Lung cancer is the leading cause of cancer-relateddeaths worldwide with an overall 5 year survival rate

of just 17% after diagnoses [1]. The reasons for thepoor prognosis are that patients tend to be diag-nosed at an advanced stage, coupled with a lack ofeffective treatments for them. This is in contrast to

* Corresponding author: e-mail: [email protected]

Currently the most sensitive method for localizing lungcancers in central airways is autofluorescence broncho-scopy (AFB) in combination with white light broncho-scopy (WLB). The diagnostic accuracy of WLB + AFBfor high grade dysplasia (HGD) and carcinoma in situ isvariable depending on physician’s experience. WhenWLB + AFB are operated at high diagnostic sensitivity,the associated diagnostic specificity is low. Raman spec-troscopy probes molecular vibrations and gives highlyspecific, fingerprint-like spectral features and has high ac-curacy for tissue pathology classification. In this study wepresent the use of a real-time endoscopy Raman spectro-scopy system to improve the specificity. A spectrum is ac-quired within 1 second and clinical data are obtained from280 tissue sites (72 HGDs/malignant lesions, 208 benignlesions/normal sites) in 80 patients. Using multivariateanalyses and waveband selection methods on the Ramanspectra, we have demonstrated that HGD and malignantlung lesions can be detected with high sensitivity (90%)and good specificity (65%).

J. Biophotonics 10, No. 1, 98–110 (2017) / DOI 10.1002/jbio.201500204

© 2016 by WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim

patients found with early stage (0) carcinoma in situ(CIS) or stage 1A (tumor less than 2 cm without me-tastatic spread) where the 5 year survival is >70%[2]. Bronchoscopy has been used for decades for lo-calizing early stage cancers of the central airways,during which time there have been many technologi-cal improvements to the bronchoscope itself, and thedevelopment of helpful adjunct devices [3–9]. In the1990s autofluorescence bronchoscopy (AFB) wasdeveloped and increasingly used in adjunct to thestandard white light bronchoscopy (WLB) to im-prove the localization of early stage cancers and highgrade dysplasia (HGD: moderate dysplasia and se-vere dysplasia) of the main airways [3, 4]. A recentmeta-analyses of 21 studies performed by Sun et al.[10] showed that combined WLB + AFB examgreatly improved the sensitivity for localizing HGDor CIS, but with decreased specificity as comparedto WLB exam alone. The pool relative sensitivity ofWLB + AFB versus WLB was 2.04, while the poolrelative specificity of WLB + AFB versus WLB was0.65. This highlights the need of developing new ad-junct technologies for improving the specificity.Although the evidence for progression of a lesionwith HGD to invasive cancer (IC) is weak, there isstrong evidence for a field effect where HGD is arisk factor for the development of IC somewhere inthe lung [11]. Thus identifying lesions in the centralairways with HGD provides valuable informationabout patients requiring close follow-up. There washowever considerable variation in the sensitivities(41–100%) obtained across all 21 studies that used acombined WLB + AFB for indentifying HGD. Thelargest variability in sensitivities occurred at highspecificities which probably reflect differences inbiopsy protocols, and operator experience. Thosedata suggested that in order to increase the probabil-ity of achieving a high sensitivity a liberal biopsyprotocol is required otherwise the risk of missing aHGD escalates significantly. The drawback withsuch an approach is that it will result in many biop-sies which are negative for HGD or worse, whichmay lead to longer procedural times, greater health-care costs and a higher incidence of procedure com-plications. These factors hinder the practical adop-tion of this technology for widespread clinical uses.The same meta-analysis also showed that the specifi-city obtained across all 21 studies varied from 18–86%. The general trends are that for studiesachieved high sensitivities with WLB + AFB, theirspecificities were relatively low.

It is desirable for a clinical diagnostic tool to havehigh diagnostic sensitivity and good specificity. Wecarefully reviewed all the studies cited in the meta-ana-lysis and focused on studies differentiating HGD + CISfrom benign lesions with large number of cohortsand good prevalence. Three such studies [12–14]suggested that it was possible to achieve better than

90% sensitivity for HGD + CIS localization usingWLB + AFB although the associated specificity be-came low (between 18% and 32%). Our goal is todevelop a new adjunct technology, such as pointRaman spectroscopy that is capable of providinghigh specificity for HGD/CIS identification, to beused together with a WLB + AFB system operatingat high sensitivity level (>90%) for improved HGD/CIS localization. This will ultimately lead to practicalearly detection of lung cancers, thus greatly im-proved survival rate.

A number of technologies have been explored toimprove the specificity of an AFB + WLB examina-tion. The ratio of red to green (R/G) fluorescence ofa lesion is the easiest to implement requiring mini-mal additional equipment, but so far the results havebeen mixed. Another technology that has been ex-plored by our group and others to improve the spe-cificity in identifying HGD during a bronchoscopy isreflectance spectroscopy. However, initial in vivotests were conducted on a relatively small number oflesions which had either a benign/normal or ICpathology; very few, if any HGD were included [15,16]. Nevertheless the concept should be exploredfurther especially with our latest technology devel-opment that does not require a separate optical fibrecatheter, but can obtain the reflectance spectra di-rectly from the WLB imaging camera [16].

In contrast to the broad spectral features of fluor-escence and reflectance technology, Raman spectro-scopy, based on inelastic light scattering, probes mo-lecular vibrations and gives very specific, fingerprint-like spectral features and has high accuracy for dif-ferentiation between malignant and benign tissues[17]. It is potentially useful for improving the specifi-city of early lung cancer localization. However, theRaman signal is exceedingly low, and as a conse-quence, long integration times are required to ac-quire sufficiently strong Raman signals for a singlespectrum. A traditional Fourier-transform Ramansystem requires up to 30 minutes of integration timeto acquire one Raman spectrum with reasonablygood signal-to-noise ratio (SNR). This has hinderedthe clinical applications of Raman technology. Ourgroup has successfully developed a rapid, real-timeRaman spectrometer system and a dedicated endo-scopy Raman catheter for lung measurements thatsubstantially reduces the spectral acquisition time toless than 1 second [18]. This system employed pro-prietary technologies for improving the spectro-meter’s SNR [19], Raman catheter fiber backgroundfluorescence elimination, and size miniaturization[18]. We realized real-time in vivo Raman measure-ments of the lung for the first time [18]. In a pilotstudy, we conducted point Raman spectroscopymeasurements of suspicious areas identified byWLB + AFB imaging for improving early lung can-cer detection [20]. The results very well demon-

J. Biophotonics 10, No. 1 (2017) 99

© 2016 by WILEY-VCH Verlag GmbH & Co. KGaA, Weinheimwww.biophotonics-journal.org

strated the technical feasibility and clinical compat-ibility. In this manuscript, we report a study usingthe developed Raman spectroscopy technology as anadjunct device to WLB + AFB exam to improve thespecificity of localizing HGD/CIS of the central lungairways, while maintaining high detection sensitivity.

2. Patients and methods

This study was approved by the University of BritishColumbia –BCCancer Agency Research Ethics Board(certificate number: H06–00010). Patients who were at-tending the BC Cancer Agency Vancouver Center fora previously scheduled bronchoscopy were invited tovolunteer for this study. Patients must have alreadyconsented to a bronchoscopy as part of a standard diag-nostic procedure or as part of an approved lung cancerprevention study, before being approached to volun-teer. Patients were excluded if they had a cardiac pace-maker or implanted defibrillator device, had a knownallergic reaction toXylocaine, were taking a blood thin-ner such as warfarin or heparin, or had any medicalcondition such as acute or chronic respiratory failure,which could jeopardize the safety of the patient duringparticipation in the study. Women who were premeno-pausal were excluded unless they were surgically sterileor on the birth control pill.

The Raman system used to take measurementswas similar to the ones described previously by ourgroup [18, 20]. The main differences were the inclu-sion of a new thermal electrically cooled CCD detec-tor for faster start up times and reduced opticalnoises (etaloning effect), as well as a new spectro-graph with a holographic reflection type grating, al-lowing tunable wavelength range [21]. Thesechanges were implemented to improve the SNR,thus allowing for a more reliable extraction of theRaman signal from the fluorescence background in-stead of the more obtuse and glossy 2nd order deri-vative processing of the data used by us previously[20]. Figure 1 shows the schematic diagram of theendoscopic laser Raman spectroscopy system. Theinserts show the arrangement of the excitation (red)and collection fibers (green). The Raman excitationlight was produced by a wavelength stabilized785 nm diode laser, and delivered to the tissue sur-face by a detachable 1.8 mm size fiber optic probe(Raman catheter) passed down the instrument chan-nel of the bronchoscope. The maximum excitationpower at the tissue surface was 150 mW. The samecatheter collected emission from the tissue and deliv-ered it to the spectrometer for analyses. The collec-tion fibers were connected to the spectrographthrough a special round to parabolic fiber bundle tocorrect the spectral imaging distortion to achievebetter SNR. The catheter itself contained ultra low

OH impurity fibers, and a gold coated excitation fi-ber to avoid cross-talk between the excitation andcollection fibers. Optical filters were coated at thedistal end of the probe to filter out laser noise, fiberemission, and to attenuate all collected light with wa-velengths ≤820 nm (≤540 cm−1 relative to the785 nm excitation). At its proximal end, the probewas attached to a second set of optical filters withsimilar transmission characteristics, but higher OD(optical density) in the rejection bands for further,and better, attenuation of the aforementionedunwanted emissions. The spectrograph grating wastuned to cover the high wavenumber range of2050–3100 cm−1 which is known to have much lesstissue autofluorescence than lower wavenumbers(<1800 cm–1), and yet still have Raman bands sensi-tive to biomolecular changes [18, 20]. The spectralresolution in this wavenumber range was estimatedto be 8 cm–1. Although the SNR increases with exci-tation energy (power × time), an acquisition time of1 second was used as this was the time that the ca-theter and excitation spot could be reliably main-tained in the same position on the tissue surface.After data acquisition, custom designed softwarewas used to subtract the fluorescence backgroundusing a modified polynomial fit (9th order) and todisplay the calibrated Raman spectrum all within afraction of a second [22]. An example of the fluor-escence background removal was shown in Supple-mentary Information Figure S1.

The procedure was for patients to first undergo astandard WLB + AFB exam which took place in theEndoscopy Suite and the location of any autofluor-

Figure 1 Schematic diagram of the endoscopic laser RamanSpectroscopy system. The inserts show the arrangement ofthe excitation (red) and collection fibers (green). The col-lection fibers were connected to the spectrograph through aspecial round to parabolic fiber bundle to correct the spec-tral imaging distortion to achieve better SNR.

H. C. McGregor et al.: Real-time endoscopic Raman spectroscopy for in vivo early lung cancer detection100

© 2016 by WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim www.biophotonics-journal.org

escence positive lesions identified. A three stageWLB + AFB visual grading of each site was deter-mined by the physician following the criteria listedin Table 1 [4]. After the AFB, the Raman probe wasinserted into the instrument channel of the broncho-scope by the physician and directed toward an auto-fluorescence positive lesion or a normal tissue con-trol site. A probe tip to tissue distance of between5–10 mm was used which generated an excitation il-lumination spot diameter on the tissue surface of be-tween 2–4 mm. A lesion was measured between 1–6times, depending on size, and different lesions wereconsidered to be separate sites with unique distinctpathology for analyses. Biopsies were taken as direc-ted by the physician, and analyzed by an experi-enced pathologist using standard histopathology as-sessment for lung lesions [4]. The pathological cod-ing system and the corresponding tissue diagnosisare listed in Table 2. If different biopsy fragments ofthe same tissue site had different histological assess-ments, the worst assessment was taken for analysis.Multiple Raman spectra from a tissue site were aver-

aged to generate a single spectrum to represent thesite. The final data file matched the histopathologyresults to the WLB + AFB visual grades and the Ra-man spectrum for each site if all possible.

Between May 2011 and November 2012, Ramanspectra were obtained from 80 patients. Ramanmeasurements were taken from a total of 280 sites.Of the 280 sites, 214 sites were biopsied after theRaman measurements and histologically assessed.Pathology reports showed that 72 sites were gradedas either a HGD or worse, 90 sites were graded asbenign (inflammation, metaplasia, hyperplasia ormild dysplasia), and 52 were graded normal.Although biopsying normal tissue was not standardclinical procedure, some normal control sites werebiopsied as these were part of another study with aprotocol which required the subject to consent tothe taking of additional tissue. The remaining 66control sites were not biopsied; instead they weredetermined to be normal by the physician throughvisual grading during the WLB and AFB exam. De-tailed patient demographics are shown in Table 3.

Table 1 Visual grading system and the corresponding tissue description.

Visual Grade Description

1 (Normal) No visual abnormality2 (Abnormal) Visual changes suggestive of inflammation, trauma, hyperplasia, metaplasia, mild dysplasia3 (Suspicous) Visual changes suggestive of moderate dysplasia, severe dysplasia, carcinonma in situ or invasive cancer

Table 2 Pathological coding system and the corresponding tissue diagnosis.

Pathological Code Description

Normal & Benign 1 Normal2 Inflammation3 Hyperplasia or Metaplasia4 Mild Dysplasia

HGD* & Malignant 5 Moderate or Severe Dysplasia6 Carcinoma in situ (CIS)7 Microinvasive Cancer8 Invasive Cancer (IC)

* HGD refers to High Grade Dysplasias, specifically Moderate Dysplasia and Severe Dysplasia

Table 3 Patient demographics and location of the Raman reading. A total of 280 sites were measured.

Diagnosis Patient Demographics Raman Reading Location

Mean age(range)

Male Female Right Lung Left Lung Central(trachea)

Invasive SCC and CIS 66 (41–86) 21 10 17 12 2Severe + Moderate Dysplasia 62 (40–78) 32 9 24 13 4Mild Dysplasia + Metaplasia 62 (40–80) 52 16 40 18 10Hyperplasia 67 (40–78) 14 2 9 7 0Inflammation 65 (61–72) 6 0 2 3 1Normal 67 (40–84) 86 32 56 33 29Total 211 69 148 86 46



3. Statistical analysis

Multivariate statistical methods have been widelyused for classification of Raman spectra in a numberof studies related to cancer diagnoses including skinand were proven to be very effective and reliable[17]. In this study, principal components with gener-alized discriminant analysis (PC-GDA) and partialleast-squares (PLS or called PLS-DA) were used forspectral classification with leave-one-out cross-vali-dation (LOO-CV) where successive single spectrawere left out for test with the remaining spectra usedfor training.

For PC-GDA analysis, we followed the metho-dology in [23]. Briefly, the training process consistedof the following procedures: (1) the mean and stan-dard deviation of the training spectra data set werecalculated, (2) each spectrum in the training data setwas standardized by subtracting the mean and thendividing by the standard deviation, (3) the standar-dized training spectra were analyzed with principalcomponent analysis. PC factors of the training casesand PC loadings were obtained, and (4) a general-ized (linear) discrimination model was developedfrom the PC factors which could be used directly topredict the new cases. The testing process consistedof the following procedures: (1) the test spectrumwas standardized by removing the mean and divid-ing by the standard deviation obtained from thetraining spectra, (2) the PC factors of the test spec-trum was calculated based on the PC loadings fromthe training spectra, and (3) a posterior probabilityof the testing spectrum was obtained from the discri-mination model developed in the training procedure.The above procedures were repeated until all thespectra were left out once (and only once) for test-ing.

For PLS analysis, we followed the methodologyin [24]. The training process consisted of the follow-ing procedures: (1) the mean and standard deviationof the training spectra data set were calculated, (2)each spectrum in the training data set was standar-dized by subtracting the mean and then dividing bythe standard deviation, (3) the standardized trainingspectra were analyzed with NIPALS algorithm (non-linear iterative partial least squares) with the classifi-cation of the training spectra setting to the knownvalue. The weight factors, the loadings, the regres-sion coefficient and the factor scores of the trainingspectra were obtained, (4) a general discriminationmodel was developed from the training spectrawhich could be used directly to predict new cases.The testing process consisted of the following proce-dures: (1) the test spectrum was standardized by re-moving the mean and dividing by the standard de-viation obtained from the training spectra, (2) thefactor scores of the test spectrum were calculatedbased on the weight factors from the training spec-

tra, and (3) A posterior probability of the testingspectrum was obtained from the discrimination mod-el developed in the training procedure. The aboveprocedures were repeated until all the spectra wereleft out once (and only once) for testing.

Before the statistical analysis, all the Ramanspectra were normalized to their respective inte-grated spectral areas under the curve (AUC). PCA-DA and PLS analyses were performed on the fullspectra based data set and also on data set with se-lected discrete wavebands generated using a numberof feature selection strategies including: stepwisemultiple regression (STEP) [25], least absoluteshrinkage selection operator (LASSO) [26], and ge-netic algorithm (GA) [27]. These feature selectionalgorithms have been previously utilized to improvemodel prediction/classification in a number of stu-dies, for example, genomics and proteomics [28],lung cancer [29], breast cancer [30], and oral cancers[31]. All the multivariate classification analyses inthis study were implemented using MATLAB (ver-sion 2013b, Math-Works).

Waveband selection results depend on the sam-ple spectra and sample size. In order to find reliableoptimal wavebands, a LOO-CV protocol was used.In the LOO-CV waveband selection procedure, asingle spectrum was left out with the remaining spec-tra used for waveband selection operation. A set ofwavebands were selected which gave the best diag-nostic performance of the training spectra. By re-peating this procedure, every spectrum was left outonce for wavelength selection purpose. At the end,n sets of wavebands were selected, where n was thetotal number of cases. The n sets of wavebands werethen accumulated. The wavebands with higher oddsfrom the LOO-CV analysis were chosen for subse-quent PCA-DA and PLS analyses. Waveband selec-tion was also tested using three-fold cross-validationwhere the spectra were divided equally and ran-domly into three groups with two groups for trainingand one group for testing. Similar results were ob-tained for three-fold cross-validation with those ofLOO-CV. In order to prevent over-fitting or select-ing spurious wavebands, a window size of 5 pixels,corresponding to an average of 5.3 cm–1 spectralrange, was chosen based on recommendations fromRefs. [32, 33].

The receiver operating characteristic (ROC)curve was calculated from the posterior probabilitiesderived from each of the analysis models describedabove and represents the diagnostic performance ofeach model. The AUC of each ROC was calculatedusing the trapezoidal rule [34]. The significance ofthese AUCs and comparisons between differentAUCs were carried out in a standard fashion [35].All ROC analyses were based on nonparametrictechniques and were conducted separately for thePC-GDA and PLS analyses. To compare the differ-



ent statistical methods used, and to compare the uti-lity of Raman spectroscopy with other non-invasivediagnostic techniques, the specificities were calcu-lated for fixed sensitivity levels of: 90% and 95%.

4. Results

The Raman spectra were dominated by strong CHstretching related bands in the wavenumber rangefrom 2775 cm–1 to 3040 cm–1. For wavenumbers be-low 2775 cm–1 the spectra did not contain any clearRaman signals, but a significant amount of noise at-tributed to etaloning effects and residual optical fi-bre emissions. The mean Raman spectra from2775 cm–1 to 3040 cm–1 for each histopathologygroup are shown in Figure 2. Mild dysplasia and me-taplasia were grouped together because both pathol-ogies are considered to be low grade preneoplasticlesions, carrying the same risk of progression. Simi-larly, moderate and severe dysplasias were groupedtogether because both are considered to be highgrade preneoplastic lesions which carry a similar riskof progression into invasive cancer. Furthermore,due to a small number of CIS lesions (n = 2) as wellas the higher risk of progression into invasive cancer,CIS was grouped with invasive cancers (IC). Andwe did not get any microinvasive cancer cases in thisstudy. For these reasons, Figure 2 only has 6 cate-gories. Major Raman peaks are seen at 2850 cm–1,2885 cm–1, 2940 cm–1, 2965 cm–1, 2990 cm–1, and3020 cm–1. These peaks were assigned to variousfundamental CH, CH2, and CH3 stretching modes[36] and overtones of CH2 and CH3 bending modes[37]. The peak at 2850 was assigned to the CH2 sym-metric stretching modes of fatty acids and lipids,while the peak at 2885 cm–1 was for the CH3 sym-

metric stretching modes. [36]. The main peak at2940 cm–1 was assigned to a mixture of CH vibra-tions in proteins and CH3 asymmetric stretchingmodes of lipids and nucleic acids [20, 36]. And thepeaks at 2965 cm–1 and 2990 cm–1 were assigned toin-plane and out-of-plane anti-symmetric CH3stretching in lipid and fatty acid molecules [36, 38,39]. The peak 3020 cm–1 was assigned to the asym-metric stretching of =C–H group in RCH=CHR mo-lecules [39], in which the R stands for an alkenefunctional group.

Apart from the main Raman peaks there was evi-dence for smaller Raman peaks, or inflection pointsat 2790 cm–1, 2825 cm–1 and 2920 cm–1 that did notappear to be related to noise. The origins of thesebands were uncertain. The 2920 cm–1 band was mostlikely due to Fermi resonance interactions betweenthe main stretching modes and CH bending over-tones. Whereas the 2825 cm–1 band may be due toone of the pair of CH stretching modes of aldehydicfunctional groups, with the other lost in the noise atlower wavenumbers [40]. No explanation can be of-fered currently for the 2790 cm–1 band.

Despite there being clear Raman bands thatwere probably connected to the abundance of dif-ferent biomolecules, there were no unique peaksthat could be assigned to lung cancer alone.Although on average there was a distinctive loss ofthe lipid peak at 2850 cm–1 seen in the spectra frommalignant lesions, the amount lost for individual le-sions showed considerable variation. There werealso differences in the spectra from inflamed tissuecompared to the other pathologies. Specifically, in-tensity of the inflammation group was relativelyhigher than all other categories between 2850 cm–1

and 2900 cm–1. To extract a more reliable correla-tion of spectra with pathology, multivariate statisti-cal techniques were used.

Figure 2 Mean Raman spectra bydiagnosis. All spectra were normal-ized to their respective area undercurve (AUC) before averaging bydiagnosis. CIS: carcinoma in situ.



4.1 Classification based on HGD andmalignant versus benign andnormal lung tissues

When the Raman spectra (full range) were used todistinguish HGD and malignant (CIS & invasivecancers) lung tissues (n = 72) from benign and nor-mal lung tissues (n = 208), the AUCs of the resultingROCs for PLS and PC-GDA analyses were almostidentical at 0.83 (column 2, Table 4) and statisticallysignificant (p < 0.001). Waveband selection techni-ques STEP, LASSO and GA were applied beforePC-GDA and PLS analysis to improve the diagnos-tic performances. Figure S2 and Figure S3 in theSupplementary Information show the wavebands se-lected by STEP and LASSO respectively. Some ofthe selected wavebands, but not all of them, were lo-cated at the Raman peaks. Some regions off thepeak positions of the spectra were also selected andable to help with classification as well. The three wa-veband selection methods increased the AUC oftheir respective ROCs to between 0.85 and 0.88 (col-umns 3–5, Table 4). The diagnostic power of theseanalysis methods are shown in Table 5. At a sensitiv-ity level of 90%, the full spectrum PLS analysis andPC-GDA analysis provided the same specificity of51% (95% CI: 0.44–0.58). Use of the waveband se-lection techniques STEP, LASSO and GA increasedthe diagnostic specificity (columns 4–6, Table 5). For

a 90% (95% CI: 0.81–0.96) sensitivity the PLS ana-lyses with STEP waveband section provided the bestspecificity of 65% (95% CI: 0.58–0.71). As an exam-ple, the posterior probabilities and the ROC curvecorresponding to the STEP PC-GDA analysis areshown in Figure 3.

Figure 3A shows the posterior probability (PP)for each lesion to be classified as a lung cancer (ma-lignant) or precancerous lesion (HGD). From thedistribution of the posterior probabilities, the ROCcurve with 95% CIs is generated and shown in Fig-ure 3B. Figure 3C shows the box plot representationof the posterior probability distributions accordingto lesion subcategories. The PP values are relativelylow for normal tissue and all benign lesion cate-gories, but increase greatly for the HGD and malig-nant groups.

4.2 Classification based on modelingmalignant versus normal lung tissues

A different classification of all the histopathologies(n = 72 vs. n = 208) was tried by using models gener-ated from spectra of the two extremes only: IC(n = 29) and normal sites (n = 118). These modelswere then used to classify all histopathologies. Therationale for this was that all histopathologies wereexpected to lie between IC and normal sites accord-

Table 4 Area under the ROC curve and 95% CI based on full range and waveband selection algorithms for discrimi-nating precancers (HGD) and malignant lung cancers (n = 72) from benign lung diseases and normal lung tissues(n = 208).

Full range STEP LASSO GA

PLS analysis 0.83(0.77–0.88)

0.88(0.84–0.92)

0.86(0.81–0.91)

0.85(0.80–0.90)

PC-GDA analysis 0.83(0.78–0.88)

0.88(0.84–0.92)

0.85(0.80–0.90)

0.85(0.80–0.90)

Table 5 Summary of Raman spectroscopy diagnostic parameters derived from ROCs: specificity values according todifferent levels of sensitivity for full range and waveband selection algorithms. HGD and malignant (n = 72) versusbenign lung lesions and normal lung tissues (n = 208).

Sensitivity level(95% Cl)

Full range STEP LASSO GA

PLSanalysis

0.95(0.88–0.99)

0.42(0.35–0.49)

0.49(0.42–0.56)

0.46(0.39–0.53)

0.44(0.37–0.51)

0.90(0.81–0.96)

0.51(0.44–0.58)

0.65(0.58–0.71)

0.53(0.46–0.60)

0.53(0.46–0.60)

PC-GDAanalysis

0.95(0.88–0.99)

0.40(0.35–0.49)

0.46(0.39–0.53)

0.43(0.39–0.53)

0.44(0.37–0.51)

0.90(0.81–0.96)

0.51(0.44–0.58)

0.64(0.57–0.70)

0.46(0.39–0.53)

0.50(0.43–0.57)



ing to their probability to be a cancer. These testsprovided an ROC AUC of 0.85 (95% CI: 0.81–0.90),based on full spectrum PC-GDA, and an ROCAUC of 0.86 (95% CI: 0.81–0.91) based on full spec-trum PLS. The posterior probabilities and the ROC

curve corresponding to the full spectrum PC-GDAanalysis are shown in Figure 4, and demonstratedthat all the histopathologies, including those cate-gories not trained, can be well classified based onthis approach.

Figure 3 Lesion classification by Raman spectroscopy based on STEP PC-GDA analysis. (A) The posterior probability plotfor distinguishing cancerous lesions (HGD, CIS and invasive cancer; n = 72) from benign lesions (mild dysplsia, metaplasia,hyperplasia, inflammation and normal; n = 208) (B) The ROC curves and 95% CIs derived from the posterior probabilities.(C) The box plot representation of the posterior probability distributions according to lesion subcategories. The lower boundon each box shows the 25th percentile, where the upper bound shows the 75th percentile, meaning 25% and 75% of the datapoints are below these two bounds respectively. The line between these two bounds is the median. The whiskers on the plotshow the 5th and 95th percentile, meaning 5% and 95% of the data are below these whiskers respectively. Data found outsidethe 5–95th percentile whiskers are outliers, as shown by separate data points.

Figure 4 Lesion classification based on modeling using the extremes cases (29 Invasive Cancers and 118 Normal Sites) for train-ing (full spectra PC-GDA). Testing was done for the entire data set (72 HGD/malignant vs. 208 benign/normal). (A) The poster-ior probability plot for distinguishing cancerous lesions (HGD, CIS and invasive cancer; n = 72) from benign lesions (mild dys-plsia, metaplasia, hyperplasia, inflammation and normal; n = 208). (B) The ROC curves and 95%CIs derived from the posteriorprobabilities. (C) The box plot representation of the posterior probability distributions according to lesion subcategories.



4.3 Sensitivity and specificityof visual diagnosis

Before performing Raman spectrum measurement, avisual grade was also determined for the majority ofsites during the WLB + AFB exam by the physician(Table 1). There were three categories of visualgrading: grade 1 for normal, grade 2 for abnormal(suggestive of inflammation, metaplasia, hyperplasia,and mild dysplasia), and grade 3 for suspicious (sug-gestive of moderate dysplasia, severe dysplasia, CISand IC cancers) [4]. The total number of sites thathad a WLB + AFB visual grade, a Raman spectrum,and a matching histopathology assessment from abiopsy was 193, of these 38 were positive for HGDand 24 were positive for IC/CIS based on histo-pathology (see Figure 5). The STEP-PLS diagnosticalgorithms based on Raman spectra of these 193sites has an area under the ROC curve of 0.85, veryclose to the value, 0.88 for the STEP-PLS algorithmgenerated from the full 280 tissue sites. When

comparing the WLB + AFB visual grades to thecorresponding histopathology, WLB + AFB grade 3(suspicious) identified 27 out of 62 lesions that weremoderate dysplasia or worse, representing a 44%sensitivity; the specificity was 68% because it cor-rectly identified 89 of the 131 lesions that were milddysplasia or lower pathology grade according to Fig-ure 5. Alternatively visual grades 2 and 3 combined(suspicious + abnormal) identified 59 out of 62 le-sions that were moderate dysplasia or worse, a sensi-tivity of 95%; the specificity decreased to 13% be-cause only 17 of the 131 benign lesions/normal siteswere correctly identified (Figure 5). The posteriorprobabilities generated from the STEP-PLS Ramanalgorithm identified 24 more HGD and malignant le-sions (17 HGD, 1 CIS, and 6 IC) than the WLB+ AFB grade 3 while keeping the same specificitylevel of 68%. Thus the relative sensitivity improve-ment of WLB + AFB+Raman versus WLB + AFBgrade 3 was (27 + 24)/27 = 1.89. Alternatively if wekeep the same sensitivity of 95% provided by the

Figure 5 The histopathology distribution of bronchoscopy (WLB + AFB) visual grading of lung lesions and normal tissuesites. The number in the parenthesis is the total number of cases within a particular visual grading subcatergory/histopathol-ogy subcategory combination.



WLB + AFB grades 2+3 combined, the same Ramanalgorithm (STEP-PLS) were able to correctly iden-tify 48 more benign lesions/normal sites (18 normal,1 inflamed, 13 metaplasia/hyperplasia, and 16 milddysplasia) than WLB + AFB grades 2 + 3 combined.Thus the relative specificity improvement of WLB+ AFB + Raman versus WLB + AFB grades 2 + 3was (17 + 48)/17 = 3.82.

5. Discussion

The results from multivariate statistical analysis de-monstrated that HGD and malignant lung lesionscan be differentiated from benign lesions and normaltissue using point laser Raman spectroscopy. Wehave previously shown that this was possible using amodel algorithm based on linear discrimination ana-lysis (LDA) of spectra processed using a 2nd orderderivative to remove the background fluorescence[20]. The findings in the current study were obtainedfor a much larger sample size, using a system withalternative components to reduce system noise, andfocus solely on the pure Raman spectra, rather thantheir noisier 2nd order derivatives.

The spectral range from 2775 cm–1 to 3040 cm–1

was chosen as the optimal wavenumber region foranalyses as it contained the only significant Ramanemissions in the range from 2050 cm–1 to 3100 cm−1

which was measured. These emissions were assignedto various CH stretching modes, although a smallpeak at 2790 cm–1, which helped with the discrimina-tion between groups, was unidentified. The optimalupper bound cut off was determined to be 3040 cm–1

as the signal beyond this wavenumber had signifi-cantly more noise. Raman emissions due to watermolecule stretching modes that occur for wavenum-bers from 3200–3500 cm–1 were outside the range ofour system.

There were some specific characteristics in themain Raman peaks that correlated with histopathol-ogy grades. Invasive carcinoma, for example, hadspectral shapes that consistently demonstrated a re-duced intensity of the CH2 symmetric stretch emis-sions at 2850 cm–1 and increased intensity at2940 cm–1 from CH3 asymmetric stretching modes.Movasagi et al. suggested that the 2850 cm–1 was agood indicator for the change in the amount of lipidin the samples [36]. The 2940 cm–1 peak has been as-sociated with both proteins and nucleic acids [20,36]. A previous study that used Raman spectroscopyon ex-vivo lung tissue samples also found a reduc-tion in the lipid content for malignant tumor tissue[41]. This was attributed to the decrease of phospho-lipids found in the cancerous tissue when comparedto normal controls. The same study also found thatthere was an increase in certain amino acids while

others decreased, [41] and many of these are knownto have strong Raman emissions in the 2775 cm–1 to3040 cm–1 range [20]. Thus changes in the abundanceof these amino acids may have contributed to thechanges seen in the Raman spectra measured in thisstudy.

It was also clear that the average spectrum of in-flamed tissue was significantly different when com-pared to those of other histopathologies. Althoughthere were only a small number of spectra from in-flamed tissue, all six cases came from six differentpatients, indicating that the abnormalities in the in-flammation spectra were most likely not due tochance. Inflammation has been suggested to be arisk factor in certain cancers [42], particularlychronic inflammation. All six biopsies from the in-flammation sites contained both chronic and acuteinflammation histopathology, meaning that the Ra-man spectra could not be separated into the twosubgroups. Due to this co-pathology, it remains un-known if the difference in the Raman readings frominflamed tissue was due to the chronic or acute diag-nosis. Nevertheless the unique Raman signatures ofinflammation tissues support the idea that Ramanspectroscopy can help to improve the diagnostic spe-cificity of AFB because inflammation is a significantcause of false positives in AFB lung cancer localiza-tion [13].

The change in intensity of individual Ramanpeaks in spectra from lung tissues with different his-topathology were found to be insufficient for reliablypredicting the histopathology of a random lung site.Multivariate classification techniques fared muchbetter in predicting the histopathology of a randomlung site. It was found that the ROC curves gener-ated by PLS and PC-GDA analyses on the full(2775 cm–1 to 3040 cm–1) spectral range had similarAUC of 0.83 (Table 4), and the AUC increased to0.88 with STEP waveband selection. Two other wa-veband selection methods (LASSO or GA) returnedsimilar AUC values, indicating the discrimination re-sults were reliable. The sensitivities and specificitiesgenerated from these ROC curves were high (Table 5).At 90% sensitivity, 65% specificity is achieved by theSTEP-PLS method in localizing HGD and malignantlesions. In comparison the AFB clinical trial in Ref. [4]that led to its FDA approval had a similar specificityof 66%, but the sensitivity is 67% only.

A trend for the spectra as a whole was apparentas shown in Figure 3C. A relatively large change inthe posterior probability occurs between the milddysplasia and the moderate dysplasia, indicated bythe differences between the groups in the box plot.This high probability trend continues for Ramanspectra from malignant tissue sites. All four multi-variate techniques showed this trend, and the algo-rithm generated by using only the malignant casesand normal tissue spectra also showed the same



trend (Figure 4C). These would appear to indicatethat there was a significant change in the spectrafrom tissue with HGD compared to those from be-nign or normal tissue, and this change becomes morepronounced for malignant tissue. This is consistentwith increasing severity of bio-molecular changesthat accompany HGD and malignancy transforma-tions respectively. Currently it is known that notevery HGD will progress into a malignant state [11,43], and those which progress are not able to be vi-sually determined by AFB or WLB endoscopy fromthose which do not. Understanding the biochemicalsignatures of those lesions which spontaneously re-gress versus those which do not, would provide valu-able clinical information. Early detection, especiallydetection of lesions which would be known to be-come invasive could improve therapy. The trends inFigure 3C and Figure 4C suggest a hypothesis thatthe posterior probability values calculated from Ra-man spectrum may be an indicator of the likelihoodof a HGD lesion progressing to malignancy. Theposterior probability values may also be used to pre-dict the prognosis of invasive cancers. These hypoth-eses are worth to be tested in future studies.

Although the adjunct use of AFB to WLB hassignificantly increased the detection of HGD com-pared to WLB alone, there remains the inherentproblem of poor specificities [10]. The specificity ofthe AFB visual grading used in this study rangedfrom 13% for grades 2 + 3 to 68% for grade 3 only.The corresponding sensitivities swung drasticallyfrom 95% to 44% respectively. Although the latterwas a little lower than the values obtained in someAFB studies for similar specificities, it was not in-consistent with the high variability in sensitivities forhigh specificities as shown in the meta-analyses ofSun et al. [10]. This variability highlights how diffi-cult it is to consistently obtain high sensitivities withgood specificities. The adjunct use of Raman spec-troscopy made a significant improvement to the spe-cificity of detecting lesions in this study with HGDor worse. If high specificity is required, it was shownthat for keeping the 68% specificity achieved byWLB + AFB visual grade 3, Raman identified 89%more true positives (51 sites compared to 27 respec-tively), representing a 1.89 times improvement insensitivity. Alternatively if high sensitivity is re-quired, it was shown that for keeping 95% sensitivityachieved by WLB + AFB visual grades 2 + 3 com-bined, Raman reduced the false positives by 42%(66 sites compared to 114 respectively), representinga 3.82 times improvement in specificity.

This study showed that the adjunct use of Ramanspectroscopy improves the specificity of detectingHGD or malignant lung lesions of the main airwayscompared to AFB visual grading. No loss in sen-sitivity by Raman diagnosis occurred relative toWLB + AFB visual grades 2 and 3 combined; but

the specificity increased significantly. Other methodshave been tried to improve specificity, but withlimited success. The fluorescence R/G ratio methodlooks particular promising but so far only one studyshowed any significant benefit [44], but its per-formance at high sensitivity settings are still inferiorcompared to our Raman results: at 95% sensitivityRaman has a better specificity of 49%, compared tothe 32% for the fluorescence R/G ratio method; at90% sensitivity Raman has a better specificity of65%, compared to the 53% for the fluorescence R/Gratio. These results lend support to the concept thatRaman spectroscopy provides the physician with ob-jective secondary information compared to the moresubjective visual appearance of lesions using WLB/AFB examination. This will reduce the number offalse positive biopsies, procedural time, and health-care costs. Most importantly, maintaining high sensi-tivity (≥90%) is critical for accurate diagnosis andmanagement.

Raman scattering within the main lung airwayscan be measured within 1 second and used to im-proving the localization of lung cancer and precan-cerous lesions. We envision that an algorithm de-rived from a database of Raman spectra would beable to classify a lesion in less than half a second,making this approach feasible for real-time lung can-cer localization. Different from subjective interpreta-tion of WLB/AFB images, the Raman algorithm re-presents an automatic, objective diagnosis based onquantitative Raman spectral analysis. Point Ramanmeasurement on lesions identified by WLB + AFBvisual grading 2 and 3 combined is a promising newclinical method for real-time localization of lungcancer/precancerous lesions.

6. Conclusions

In conclusion, a single centre clinical investigation ofthe adjunct use of real time Raman spectroscopy tothe standard WLB and AFB for in vivo lung cancerlocalization of the central airways was conducted. Invivo real-time point laser Raman spectroscopy wasperformed on 280 lung lesions and normal tissuesites with a measurement time of 1 second per spec-trum. Using multivariate techniques and wavebandselection methods on the Raman spectra, it wasshown that HGD and malignant lung lesions can bedifferentiated from benign lung lesions and normallung tissues with high sensitivity (90%) and goodspecificity (65%). Compared to WLB + AFB visualgrade 3 based diagnosis, Raman + WLB + AFB im-proved the sensitivity of localizing HGD andmalignant lesions by 1.89 times, while compared toWLB + AFB visual grade 2 + 3 combined diagnosis,Raman + WLB + AFB improved the specificity of



localizing HGD and malignant lesions by 3.82times. Different from subjective interpretation ofWLB/AFB images, the Raman algorithm could po-tentially become an automatic and objective diagno-sis method based on quantitative spectral analysis.Further multi-center clinical trials are warranted tofully test the potential of this technology for im-proving lung cancer and precancerous lesion locali-zation.

Supporting Information

Additional supporting information can be found inthe online version of this article at the publisher’swebsite.

Acknowledgements This work was supported by the Ca-nadian Institutes of Health Research (CIHR grant # PPP-79109 and MOP-85011) and the Canadian Cancer SocietyResearch Institute (Grant # 20352). The authors wouldlike to thank Myles McKinnon, Jennifer Campbell, andDr. Keishi Ohtani for endoscopy procedure support, andDr. Martial Guillaud for providing statistical expertise forthe Raman data analysis.

References

[1] M. P. Coleman, D. Forman, H. Bryant, J. Butler, B.Rachet, C. Maringe, U. Nur, E. Tracey, M. Coory,J. Hatcher, C. E. McGahan, D. Turner, L. Marrett,M. L. Gjerstorff, T. B. Johannesen, J. Adolfsson,M. Lambe, G. Lawrence, D. Meechan, E. J. Mor-ris, R. Middleton, J. Steward, and M. A. Richards,Lancet 377, 127–138 (2011).

[2] R. Rami-Porta, D. Ball, J. Crowley, D. J. Giroux,J. Jett, W. D. Travis, M. Tsuboi, E. Vallieres, P. Gold-straw, and I. S. Committee, J. Thorac. Oncol. 2, 593–602 (2007).

[3] S. Lam, C. MacAulay, J. Hung, J. LeRiche, A. Profio,and B. Palcic, J. Thorac Cardiovasc Surg 105, 1035–1040 (1993).

[4] S. Lam, T. Kennedy, M. Unger, Y. E. Miller, D. Gel-mont, V. Rusch, B. Gipe, D. Howard, J. C. LeRiche,A. Coldman, and A. F. Gazdar, Chest 113, 696–702(1998).

[5] P. Lee, H. A. P. Brokx, P. E. Postmus, and T. G.Sutedja, Lung Cancer 58, 44–49 (2007).

[6] P. T. Nguyen, O. Salvado, I. B. Masters, C. S. Farah,and D. Fielding, J. Bronchology Interv. Pulmonol. 17,109–116 (2010).

[7] A. McWilliams, T. Shaipanich, and S. Lam, Thorac.Surg. Clin. 23, 153–161 (2013).

[8] S. Leong, T. Shaipanich, S. Lam, and K. Yasufuku, J.Thorac. Dis. Suppl 5, S498-S510 (2013).

[9] K. Ohtani, A. Lee, and S. Lam, Respirology 17, 261–269 (2012).

[10] J. Sun, D. H. Garfield, B. Lam, J. Yan, A. Gu, J. Shen,and B. Han, J. Thorac. Oncol. 6, 1336–1344 (2011).

[11] T. Ishizumi, A. McWilliams, C. MacAulay, A. Gazdar,and S. Lam, Cancer Metastasis Rev. 29, 5–14 (2010).

[12] P. Chhajed, K. Shibuya, H. Hoshino, M. Chiyo, K. Ya-sufuku, K. Hiroshima, and T. Fujisawa, Eur. Respir. J.25, 951–955 (2005).

[13] K. Shibuya, T. Fujisawa, H. Hoshino, M. Baba, Y. Sai-toh, T. Iizasa, M. Suzuki, M. Otsuji, K. Hiroshima, andH. Ohwada, Lung Cancer 32, 19–25 (2001).

[14] P. Vermylen, P. Pierard, C. Roufosse, T. Bosschaerts,A. Verhest, J.-P. Sculier, and V. Ninane, Lung Cancer25, 161–168 (1999).

[15] M. P. Bard, A. Amelink, M. Skurichina, V. N. Hegt,R. P. Duin, H. J. Sterenborg, H. C. Hoogsteden, andJ. G. Aerts, Chest 129, 995–1001 (2006).

[16] Y. S. Fawzy, M. Petek, M. Tercelj, and H. Zeng, J.Biomed. Opt. 11, 044003 (2006).

[17] Lui, J. Zhao, D. McLean, and H. Zeng, Cancer Re-search 72, 2491–2500 (2012).

[18] M. A. Short, S. Lam, A. McWilliams, J. Zhao, H. Lui,and H. Zeng, Opt. Lett. 33, 711–713 (2008).

[19] Z. Huang, H. Zeng, I. Hamzavi, D. I. McLean, and H.Lui, Opt. Lett. 26, 1782–1784 (2001).

[20] M. A. Short, S. Lam, A. M. McWilliams, D. N. Iones-cu, and H. Zeng, J Thorac Oncol 6, 1206–1214 (2011).

[21] H. C. Pawluk, M. A. Short, S. Lam, A. M. McWil-liams, D. N. Ionescue, and H. Zeng, Proc. SPIE 8207,820737/1–82037/6 (2012).

[22] J. Zhao, H. Lui, D. I. McLean, and H. Zeng, Appl.Spectrosc. 61, 1225–1232 (2007).

[23] T. Hill and P. Lewicki, Statistics: Methods and Appli-cations (Statsoft Inc, Tulsa, OK, USA 2005), pp. 231–244

[24] T. Hill and P. Lewicki, Statistics: Methods and Appli-cations (Statsoft Inc, Tulsa, OK, USA 2005), pp. 393–404

[25] D. L. Streiner, Can J Psychiatry 39, 191–196 (1994).[26] R. Tibshirani, J. R. Statist. Soc. B 58, 267–288 (1996).[27] R. Leardi and A. Lupiáñez González, Chemometrics

Intell. Lab. Sys. 41, 195–207 (1998).[28] Y. Saeys, I. Inza, and P. Larranaga, Bioinformatics 23,

2507–2517 (2007).[29] Z. Hu, X. Chen, Y. Zhao, T. Tian, G. Jin, Y. Shu, Y.

Chen, L. Xu, K. Zen, C. Zhang, and H. Shen, J ClinOncol 28, 1721–1726 (2010).

[30] H. Guo and A. K. Nandi, Pattern Recogn. 39, 980–987(2006).

[31] D. Elashoff, H. Zhou, J. Reiss, J. Wang, H. Xiao, B.Henson, S. Hu, M. Arellano, U. Sinha, A. Le, D. Mes-sadi, M. Wang, V. Nabili, M. Lingen, D. Morris, T.Randolph, Z. Feng, D. Akin, D. A. Kastratovic, D.Chia, E. Abemayor, and D. T. Wong, Cancer Epide-miol Biomarkers Prev 21, 664–672 (2012).

[32] C. M. Andersen and R. Bro, J Chemometrics 24, 728–737 (2010).

[33] L. Norgaard, A. Saudland, J. Wagner, J. P. Nielsen, L.Munck, and S. B. Engelsen, Appl Spectrosc 54, 413–419 (2000).



[34] J. N. Matthews, D. G. Altman, M. J. Campbell, and P.Royston, BMJ 300, 230–235 (1990).

[35] J. A. Hanley and B. J. McNeil, Radiology 148, 839–843 (1983).

[36] Z. Movasaghi, S. Rehman, and I. U. Rehman, ApplSpectrosc Rev 42, 493–541 (2007).

[37] I. Hill and I. Levin, J Chem Phys 70, 842–851 (2008).[38] C. Krafft, L. Neudert, T. Simat, and R. Salzer, Spec-

trochim Acta A Mol Biomol Spectros 61, 1529–1535(2005).

[39] V. Baeten, P. Hourant, M. T. Morales, and R. Apari-cio, J. Agric. Food. Chem. 46, 2638–2646 (1998).

[40] E. L. Saier, L. R. Cousins, and M. R. Basila, Anal.Chem. 34, 824–826 (1962).

[41] Z. Huang, A. McWilliams, H. Lui, D. I. McLean, S.Lam, and H. Zeng, Int. J. Cancer 107, 1047–1052(2003).

[42] A. Dalgleish and K. O’Byrne, in: A. Dalgleish and B.Haefner (eds.), The Link Between Inflammation andCancer, (Springer US, 2006), pp. 1–38.

[43] R. H. Breuer, A. Pasic, E. F. Smit, E. van Vliet,A. Vonk Noordegraaf, E. J. Risse, P. E. Postmus, andT. G. Sutedja, Clinical Cancer Res 11, 537–543 (2005).

[44] P. Lee, R. M. van den Berg, S. Lam, A. F. Gazdar, K.Grunberg, A. McWilliams, J. Leriche, P. E. Postmus,and T. G. Sutedja, Clin. Cancer Res. 15, 4700–4705(2009).



journal of biophotonics - ubc physics & astronomyhzeng/s/zeng lung clinical paper.pdf ·...

Documents