Portal de Periódicos da CAPES

Automated Quantification of Tumor Viability in a Rabbit Liver Tumor Model after Chemoembolization Using Infrared Imaging

2015; Elsevier BV; Volume: 185; Issue: 7 Linguagem: Inglês

10.1016/j.ajpath.2015.03.023

1525-2191

Hadrien D’inca, Julien Namur, S. Ghegediban, Michel Wassef, Florentina Pascale, Alexandre Laurent, Michel Manfait,

Infrared Thermography in Medicine

The rabbit VX2 tumor is a fast-growing carcinoma model commonly used to study new therapeutic devices, such as catheter-based therapies for patients with inoperable hepatocellular carcinoma. The evaluation of tumor viability after such locoregional therapies is essential to directing hepatocellular carcinoma management. We used infrared microspectroscopy for the automatic characterization and quantification of the VX2 liver tumor viability after drug-eluting beads transarterial chemoembolization (DEB-TACE). The protocol consisted of K-means clustering followed by principal component analysis (PCA) and linear discriminant analysis (LDA). The K-means clustering was used to classify the spectra from the infrared images of control or treated tumors and to build a database of many tissue spectra. On the basis of this reference library, the PCA-LDA analysis was used to build a predictive model to identify and quantify automatically tumor viability on unknown tissue sections. For the DEB group, the LDA model determined that the surface of tumor necrosis represented 91.6% ± 8.9% (control group: 33.1% ± 19.6%; Mann-Whitney P = 0.0004) and the viable tumor 2.6% ± 4% (control group: 62.2% ± 15.2%; Mann-Whitney P = 0.0004). Tissue quantification measurements correlated well with tumor necrosis (r = 0.827, P < 0.0001) and viable tumor (r = 0.840, P < 0.0001). Infrared imaging and PCA-LDA analysis could be helpful for easily assessing tumor viability. The rabbit VX2 tumor is a fast-growing carcinoma model commonly used to study new therapeutic devices, such as catheter-based therapies for patients with inoperable hepatocellular carcinoma. The evaluation of tumor viability after such locoregional therapies is essential to directing hepatocellular carcinoma management. We used infrared microspectroscopy for the automatic characterization and quantification of the VX2 liver tumor viability after drug-eluting beads transarterial chemoembolization (DEB-TACE). The protocol consisted of K-means clustering followed by principal component analysis (PCA) and linear discriminant analysis (LDA). The K-means clustering was used to classify the spectra from the infrared images of control or treated tumors and to build a database of many tissue spectra. On the basis of this reference library, the PCA-LDA analysis was used to build a predictive model to identify and quantify automatically tumor viability on unknown tissue sections. For the DEB group, the LDA model determined that the surface of tumor necrosis represented 91.6% ± 8.9% (control group: 33.1% ± 19.6%; Mann-Whitney P = 0.0004) and the viable tumor 2.6% ± 4% (control group: 62.2% ± 15.2%; Mann-Whitney P = 0.0004). Tissue quantification measurements correlated well with tumor necrosis (r = 0.827, P < 0.0001) and viable tumor (r = 0.840, P < 0.0001). Infrared imaging and PCA-LDA analysis could be helpful for easily assessing tumor viability. The VX2 tumor model originates from a squamous cell carcinoma that developed as a result of malignant changes in the cells of a Shope virus–induced skin papilloma in a domestic rabbit.1Kidd J.G. Rous P. A transplantable rabbit carcinoma originating in a virus-induced papilloma and containing the virus in masked or altered form.J Exp Med. 1940; 71: 813-838Crossref PubMed Scopus (162) Google Scholar, 2Rous P. Kidd J.G. Smith W.E. Experiments on the cause of the rabbit carcinomas derived from virus-induced papillomas. II. Loss by the Vx2 carcinoma of the power to immunize hosts against the papilloma virus.J Exp Med. 1952; 96: 159-174Crossref PubMed Scopus (86) Google Scholar This tumor model is serially transplantable in allogenic adult rabbits, easily implantable, and grows quickly in many types of organs, such as lungs,3Tu M. Xu L. Wei X. Miao Y. How to establish a solitary and localized VX2 lung cancer rabbit model? a simple and effective intrapulmonary tumor implantation technique.J Surg Res. 2009; 154: 284-292Abstract Full Text Full Text PDF PubMed Scopus (14) Google Scholar liver,4Ke S. Ding X.M. Kong J. Gao J. Wang S.H. Cheng Y. Sun W.B. Low temperature of radiofrequency ablation at the target sites can facilitate rapid progression of residual hepatic VX2 carcinoma.J Transl Med. 2010; 8: 73Crossref PubMed Scopus (74) Google Scholar or rectum.5Liang X.M. Tang G.Y. Cheng Y.S. Zhou B. Evaluation of a rabbit rectal VX2 carcinoma model using computed tomography and magnetic resonance imaging.World J Gastroenterol. 2009; 15: 2139-2144Crossref PubMed Scopus (17) Google Scholar Therefore, the VX2 tumor blood supply is almost entirely from the hepatic artery, similar to that of humans,6Ramirez L.H. Juliéron M. Bonnay M. Koscielny S. Zhao Z. Gouyette A. Munck J.N. Stimulation of tumor growth in vitro and in vivo by suramin on the VX2 model.Invest New Drugs. 1995; 13: 51-53Crossref PubMed Scopus (21) Google Scholar and rabbit hepatic arteries are large enough to permit hepatic artery catheterization.7Geschwind J.F. Artemov D. Abraham S. Omdal D. Huncharek M.S. McGee C. Arepally A. Lambert D. Venbrux A.C. Lund G.B. Chemoembolization of liver tumor in a rabbit model: assessment of tumor cell death with diffusion-weighted MR imaging and histologic analysis.J Vasc Interv Radiol. 2000; 11: 1245-1255Abstract Full Text Full Text PDF PubMed Scopus (174) Google Scholar This makes the VX2 tumor a common animal model for the preclinical evaluation of new anticancer treatments8Zheng L.F. Li Y.J. Wang H. Zhao J.L. Wang X.F. Hu Y.S. Zhang G.X. Combination of vascular endothelial growth factor antisense oligonucleotide therapy and radiotherapy increases the curative effects against maxillofacial VX2 tumors in rabbits.Eur J Radiol. 2011; 78: 272-276Abstract Full Text Full Text PDF PubMed Scopus (10) Google Scholar, 9Han K. Wang Z. Peng X. Chen B. Wen X. Dong Y. Wu C. Transarterial chemoembolization using docetaxel-loaded phytantriol cubic phase precursor for the treatment of hepatocellular carcinoma.J Pharm Sci. 2010; 100: 2240-2247Crossref PubMed Scopus (9) Google Scholar, 10Kang J. Wu X. Wang Z. Ran H. Xu C. Wu J. Wang Z. Zhang Y. Antitumor effect of docetaxel-loaded lipid microbubbles combined with ultrasound-targeted microbubble activation on VX2 rabbit liver tumors.J Ultrasound Med. 2010; 29: 61-70PubMed Google Scholar, 11Nitta N. Sonoda A. Seko A. Ohta S. Nagatani Y. Tsuchiya K. Otani H. Tanaka T. Kanasaki S. Takahashi M. Murata K. A combination of cisplatin-eluting gelatin microspheres and flavopiridol enhances anti-tumour effects in a rabbit VX2 liver tumour model.Br J Radiol. 2010; 83: 428-432Crossref PubMed Scopus (11) Google Scholar and for new therapeutic devices, such as catheter-based therapies for patients with inoperable hepatocellular carcinoma. The evaluation of tumor response after locoregional therapies is essential in directing management for hepatocellular carcinoma. An understanding of the various therapeutic strategies and their posttherapy imaging appearance is essential for accurately assessing treatment response. The evaluation of tumor response should include not only anatomical evaluation, such as reduction in tumor size, but also the reduction of tumor viability, the degree of induced necrosis,12Bruix J. Sherman M. Llovet J.M. Beaugrand M. Lencioni R. Burroughs A.K. Christensen E. Pagliaro L. Colombo M. Rodes J. Clinical management of hepatocellular carcinoma: conclusions of the Barcelona-2000 EASL conference.J Hepatol. 2001; 35: 421-430Abstract Full Text Full Text PDF PubMed Scopus (3769) Google Scholar, 13Livraghi T. Goldberg S.N. Lazzaroni S. Meloni F. Ierace T. Solbiati L. Gazelle G.S. Hepatocellular carcinoma: radio-frequency ablation of medium and large lesions.Radiology. 2000; 214: 761-768Crossref PubMed Scopus (961) Google Scholar, 14Yan K. Chen M.H. Yang W. Wang Y.B. Gao W. Hao C.Y. Xing B.C. Huang X.F. Radiofrequency ablation of hepatocellular carcinoma: long-term outcome and prognostic factors.Eur J Radiol. 2008; 67: 336-347Abstract Full Text Full Text PDF PubMed Scopus (97) Google Scholar and the biochemical changes that occur inside tumor tissues. During the past decade, there has been considerable interest in spectroscopy-based microscopies, which use the intrinsic optical properties of the tissues to produce histologic images without staining or exogenous markers. It was first revealed in the 1950s that neoplastic and normal tissues could be discriminated based on their infrared absorption spectra15Cook E.S. Jansen C.H. Kreke C.W. Motzel W. A new tool for cancer investigation: qualitative and quantitative infrared spectroscopy of proteins and enzymes.Acta Unio Int Contra Cancrum. 1956; 12: 503-507PubMed Google Scholar, 16Fong C.T. Lippincott S.W. Eriksen N. Infrared spectroscopy of crystalline albumin in human neoplasia.J Natl Cancer Inst. 1957; 18: 271-275PubMed Google Scholar due to the differences in their biochemical composition. Infrared spectra directly provide a multivariate nonperturbing molecular description able to distinguish and quantify histologic constituents from tissues. Chemical concentrations are quantified by spectral absorbance properties at specific frequencies, and subtle molecular structural changes are indicated by spectral peak shifts, band shapes, and relative intensity changes.17Fernandez D.C. Bhargava R. Hewitt S.M. Levin I.W. Infrared spectroscopic imaging for histopathologic recognition.Nat Biotechnol. 2005; 23: 469-474Crossref PubMed Scopus (497) Google Scholar, 18Diem M. Papamarkakis K. Schubert J. Bird B. Romeo M.J. Miljković M. The infrared spectral signatures of disease: extracting the distinguishing spectral features between normal and diseased states.Appl Spectrosc. 2009; 63: 307A-318ACrossref PubMed Scopus (31) Google Scholar The coupling of the infrared spectrometer with an imaging system (microspectroscopy) further permitted to combine the measured spectrum to a spatial position on the sample and to record infrared images directly on thin tissue sections. Comprehensive studies using this technique revealed the potential for fast, stain-free, nondestructive molecular histopathologic analysis with a high spatial resolution and established the capability of infrared microspectroscopy (IRMS) to complement histopathologic tools for cancerous tissue diagnosis in different organs.19Bellisola G. Sorio C. Infrared spectroscopy and microscopy in cancer research and diagnosis.Am J Cancer Res. 2012; 2: 1-21PubMed Google Scholar, 20Wang J.S. Shi J.S. Xu Y.Z. Duan X.Y. Zhang L. Wang J. Yang L.M. Weng S.F. Wu J.G. FT-IR spectroscopic analysis of normal and cancerous tissues of esophagus.World J Gastroenterol. 2003; 9: 1897-1899Crossref PubMed Scopus (62) Google Scholar, 21Beleites C. Steiner G. Sowa M. Baumgartner R. Sobottka S. Schackert G. Salzer R. Classification of human gliomas by infrared imaging spectroscopy and chemometric image processing.Vibrational Spectroscopy. 2005; 38: 143-149Crossref Scopus (61) Google Scholar, 22Kendall C. Isabelle M. Bazant-Hegemark F. Hutchings J. Orr L. Babrah J. Baker R. Stone N. Vibrational spectroscopy: a clinical tool for cancer diagnostics.Analyst. 2009; 134: 1029-1045Crossref PubMed Scopus (248) Google Scholar, 23Meurens M. Wallon J. Tong J. Noel H. Haot J. Breast cancer detection by Fourier transform infrared spectrometry.Vibrational Spectroscopy. 1996; 10: 341-346Crossref Scopus (57) Google Scholar, 24Lasch P. Haensch W. Naumann D. Diem M. Imaging of colorectal adenocarcinoma using FT-IR microspectroscopy and cluster analysis.Biochim Biophys Acta. 2004; 1688: 176-186Crossref PubMed Scopus (363) Google Scholar, 25Dukor R.K. Vibrational spectroscopy in the detection of cancer. Handbook of vibrational spectroscopy. Wiley, Hoboken, NJ2002: 3335-3361Google Scholar The most recent developments in IRMS aimed at automating the procedure of tissue recognition and quantification by using statistical methods and prediction algorithms.26Gaigneaux A. Ruysschaert J.M. Goormaghtigh E. Infrared spectroscopy as a tool for discrimination between sensitive and multiresistant K562 cells.Eur J Biochem. 2002; 269: 1968-1973Crossref PubMed Scopus (56) Google Scholar, 27Gazi E. Dwyer J. Lockyer N. Gardner P. Vickerman J.C. Miyan J. Hart C.A. Brown M. Shanks J.H. Clarke N. The combined application of FTIR microspectroscopy and ToF-SIMS imaging in the study of prostate cancer.Faraday Discuss. 2004; 126 (discussion 77-92): 41-59Crossref PubMed Scopus (80) Google Scholar, 28Krafft C. Sobottka S.B. Geiger K.D. Schackert G. Salzer R. Classification of malignant gliomas by infrared spectroscopic imaging and linear discriminant analysis.Anal Bioanal Chem. 2007; 387: 1669-1677Crossref PubMed Scopus (60) Google Scholar, 29Ly E. Piot O. Durlach A. Bernard P. Manfait M. Differential diagnosis of cutaneous carcinomas by infrared spectral micro-imaging combined with pattern recognition.Analyst. 2009; 134: 1208-1214Crossref PubMed Scopus (58) Google Scholar, 30Nallala J. Diebold M.D. Gobinet C. Bouché O. Sockalingum G.D. Piot O. Manfait M. Infrared spectral histopathology for cancer diagnosis: a novel approach for automated pattern recognition of colon adenocarcinoma.Analyst. 2014; 139: 4005-4015Crossref PubMed Google Scholar, 31Salman A. Argov S. Ramesh J. Goldstein J. Sinelnikov I. Guterman H. Mordechai S. FT-IR microscopic characterization of normal and malignant human colonic tissues.Cell Mol Biol. 2001; 47: OL159-166PubMed Google Scholar These methods are based on the measurement of a large number of tissue specimens. The information gathered from this large data set is used to produce a reference spectral library of each tissue type. On the basis of this spectral library, a predictive model is elaborated and validated. Applying a model to a new tissue section generates a false colored image in which each color corresponds to a type of tissue and where the surface of each tissue is automatically calculated. Once the model is set up and validated, it can be applied to a new infrared image in only 1 minute. The infrared imaging technique is a solution to visualize on the same image the morphologic information and the molecular composition of tissues. It appears to be a helpful technique for studying objectively and quantitatively tumor response. Our aim was to validate the use of IRMS to automate the recognition and the quantification of VX2 liver tumor viability after treatment with drug-eluting beads transarterial chemoembolization (DEB-TACE). We worked on a pool of untreated VX2 tumors and a pool of DEB-TACE–treated tumors. First, the infrared spectra characteristics of each tissue of interest were recorded and a prediction model of tissue types was developed. Second, the model was applied to a set of new test VX2 samples to assess the surface of viable and necrotic tumor. A validation procedure was included at each step of the data processing. Infrared results were correlated by histopathologic analysis as standard of reference. This study was approved by the Animal Care Committee and was performed in accordance with our institutional guidelines. Adult New Zealand white rabbits weighing 6 to 8 lb underwent implantation of rabbit VX2 tumor in the liver. The tumors were induced by an injection of a VX2 cells suspension (0.25 × 106 cells/mL) directly in the liver. We included 27 rabbits with VX2 liver tumors: 16 rabbits were subjected to a DEB-TACE treatment and compared with a control group of 11 rabbits. Animals from DEB-TACE group were treated after 12 days of tumor development and were euthanized 3 days after the embolization procedure (15 days of tumor development). In the same model, one experiment on our laboratory found a significant increase of tumor necrosis 3 days after embolization compared with untreated tumors. Animals in the control group were euthanized after 14 days of tumor development. Then, tumor-bearing livers were resected and samples were formalin fixed and paraffin embedded. Two adjacent sections were cut from each sample using a microtome. The first section (10-μm thick) was mounted on a calcium fluoride window suitable for IRMS. The second section (5-μm thick) was put on a standard glass slide, dewaxed, and rehydrated by means of successive baths of xylene and alcohol and stained with hematoxylin-eosin–saffron (HES) to serve as a control for infrared imaging. Infrared spectral images were collected with an infrared microscope (Spectrum Spotlight 300 Imaging System, PerkinElmer, Courtaboeuf, France) coupled to a Spectrum One Fourier transform infrared (FTIR) spectrometer using the image mode. The device is equipped with a nitrogen-cooled mercury cadmium telluride 16-pixel-line detector for imaging and a computer-controlled stage to collect large spectroscopic images from a sample. The microscope was isolated in a venting Plexiglas housing to enable purging with dry air and to eliminate atmospheric interferences. Before acquisition, a visible image of the sample was recorded and the area of interest was selected by comparison to the corresponding H&E-stained adjacent section. In this study, 38 spectroscopic images were recorded from VX2 liver tumors and could vary in size from 8 × 8 mm (64 mm2) to 10 × 10 mm (or 100 mm2). Each pixel sampled a 25 × 25 μm (625 μm2) area at the sample plane, providing images that contained 120,000 to 160,000 individual infrared spectra (depending on the size of the image). Spectral data were acquired in transmission mode. All spectral measurements were recorded using a spectral resolution of 4 cm−1 and two scans per pixels, each spectrum containing 1601 values of absorbance, spanning the spectral range of 800 cm−1 to 4000 cm−1. A background spectrum was collected (75 accumulations, 4-cm−1 resolution) outside the sample (on the calcium fluoride window area) to ratio against the single-beam spectra. The resulting spectra were then automatically converted into absorbance. All subsequent data treatment protocols were performed with Spectrum image-Spotlight version 300 (PerkinElmer), Opus version 5.5 (Bruker Optik, Ettlingen, Germany), and Matlab version 7.12 (Mathworks, Natick, MA) software using protocols validated in our laboratory. All preprocessing and processing data were obtained directly from spectral images in the infrared absorption range of 900 to 1800 cm−1 (451 values of absorbance) considered as the most informative region. Paraffin exhibits strong absorption bands at 1368 and 1467 cm−1. To correct the contribution of paraffin in FTIR spectra, we used an automated processing method based on extended multiplicative signal correction (EMSC) validated and commonly used in our laboratory.32Ly E. Piot O. Wolthuis R. Durlach A. Bernard P. Manfait M. Combination of FTIR spectral imaging and chemometrics for tumour detection from paraffin-embedded biopsies.Analyst. 2008; 133: 197-205Crossref PubMed Scopus (108) Google Scholar, 33Wolthuis R. Travo A. Nicolet C. Neuville A. Gaub M.P. Guenot D. Ly E. Manfait M. Jeannesson P. Piot O. IR spectral imaging for histopathological characterization of xenografted human colon carcinomas.Anal Chem. 2008; 80: 8461-8469Crossref PubMed Scopus (51) Google Scholar It is a numerical dewaxing without xylene treatment (chemical dewaxing). Briefly the EMSC algorithm constrain the bands of paraffin at the same intensity on all tissue spectra. The variations of the paraffin bands intensity are eliminated from tissue spectra. The EMSC method was also used for eliminating the infrared spectra with low signal-to-nose ratio.34Kohler A. Kirschner C. Oust A. Martens H. Extended multiplicative signal correction as a tool for separation and characterization of physical and chemical information in Fourier transform infrared microscopy images of cryo-sections of beef loin.Appl Spectrosc. 2005; 59: 707-716Crossref PubMed Scopus (128) Google Scholar Consequently, the result of classification process [K-means (KM) and linear discriminant analysis (LDA)] is based only on the spectral differences between each type of tissue. The spectra were paraffin and baseline corrected in the same step, and finally they were vector normalized on the 900- to 1800-cm−1 spectral range. For a more comprehensive description of the EMSC method, the reader should refer to previous articles.33Wolthuis R. Travo A. Nicolet C. Neuville A. Gaub M.P. Guenot D. Ly E. Manfait M. Jeannesson P. Piot O. IR spectral imaging for histopathological characterization of xenografted human colon carcinomas.Anal Chem. 2008; 80: 8461-8469Crossref PubMed Scopus (51) Google Scholar, 35Martens H. Stark E. Extended multiplicative signal correction and spectral interference subtraction: new preprocessing methods for near infrared spectroscopy.J Pharm Biomed Anal. 1991; 9: 625-635Crossref PubMed Scopus (363) Google Scholar On a first series of 14 infrared images (Table 1), we performed a KM classification. This statistical classification method gathered spectra that have similar spectral characteristics in clusters whose number is determined by the operator.24Lasch P. Haensch W. Naumann D. Diem M. Imaging of colorectal adenocarcinoma using FT-IR microspectroscopy and cluster analysis.Biochim Biophys Acta. 2004; 1688: 176-186Crossref PubMed Scopus (363) Google Scholar, 29Ly E. Piot O. Durlach A. Bernard P. Manfait M. Differential diagnosis of cutaneous carcinomas by infrared spectral micro-imaging combined with pattern recognition.Analyst. 2009; 134: 1208-1214Crossref PubMed Scopus (58) Google Scholar, 32Ly E. Piot O. Wolthuis R. Durlach A. Bernard P. Manfait M. Combination of FTIR spectral imaging and chemometrics for tumour detection from paraffin-embedded biopsies.Analyst. 2008; 133: 197-205Crossref PubMed Scopus (108) Google Scholar, 36Travo A. Piot O. Wolthuis R. Gobinet C. Manfait M. Bara J. Forgue-Lafitte M.E. Jeannesson P. IR spectral imaging of secreted mucus: a promising new tool for the histopathological recognition of human colonic adenocarcinomas.Histopathology. 2010; 56: 921-931Crossref PubMed Scopus (49) Google Scholar Then, the pixel on infrared image (coordinates x and y) corresponding to the spectrum is colored with the same color as the cluster. The color is randomly attributed to each cluster by the Matlab software. The result of the KM analysis is a false color image in which each color corresponds to a cluster. All the eliminated spectra by EMSC algorithm were colored as white pixels in the KM clustered images (spectra with low signal-to-nose ratio). Each image was analyzed independently; thus, colors of the different KM images were not comparable. The number of clusters varied from 2 to 10. The accurate number of clusters was chosen with the pathologist in such a way that each tissue type identified on HES-stained sections was represented by at least one cluster. Each cluster had a unique histologic assignment.Table 1Sample Repartition in Each Study GroupSampleKM followed by PCA-LDATest samplesTotalCTRLDEB-TACECTRLDEB-TACETumors4371327Tissue sections and infrared images7771738CTRL, control; DEB-TACE, drug-eluting beads transarterial chemoembolization; KM, K-means; PCA-LDA, principal component analysis and linear discriminant analysis. Open table in a new tab CTRL, control; DEB-TACE, drug-eluting beads transarterial chemoembolization; KM, K-means; PCA-LDA, principal component analysis and linear discriminant analysis. To confirm the KM clustering, a total of 255 areas measuring 200 μm × 200 μm each were selected on HES-stained tissue sections in zones of viable tumor, tumor necrosis, fibrosis, normal liver parenchyma, and liver parenchyma necrosis (65, 75, 45, 40, and 30 areas, respectively). Each area was then located on the 14 KM images to determine whether each cluster color corresponded to a particular tissue. Areas that contained multiple clusters were considered unallocated areas. We calculated the sensitivity of each cluster by dividing the number of areas correctly matched by the sum of true-positive and false-negative results. Then, we calculated the specificity of each cluster by dividing the number of areas correctly matched by the sum of true-negative and false-positive results. The spectra of validated KM clusters were used to create a database containing thousands of spectra assigned to a specific type of tissue: viable tumor, tumor necrosis, fibrosis, liver parenchyma, and liver parenchyma necrosis. On the basis of the tissue spectral database, LDA looks for the variables containing both the greatest interclass variance and the smallest intraclass variance. Because we analyzed a huge quantity of spectra in the same time (approximately 430,000), we choose to apply a principal component analysis (PCA) before the LDA. The PCA is a commonly used spectral data processing method that reduces the size of the data while still retaining the variance. This variance is represented by principal components. The resulting scores were then used as inputs for LDA. The PCA-LDA model obtained is a linear combination of the variables to discriminate between the classes. To evaluate the accuracy of the PCA-LDA model, the database was randomly divided into two uneven sets of spectra: a training set corresponding to two-thirds of spectra and a validation set corresponding to one-third of spectra. The training data set was used to establish and optimize the PCA and LDA parameters that would provide the best possible classification. The validation data set, which is labeled with the correct answer, was used to test the accuracy of LDA model. The confrontation between the training set and the validation set gives the confusion matrix of the prediction model. The sensitivity of each class is calculated by dividing the number of spectra labeled as true-positive results by the sum of the number of true-positive and false-negative results. The specificity of each class is calculated by dividing the number of spectra labeled as true-negative results by the sum of true-negative and false-positive results. We chose the PCA-LDA model that displayed the best sensitivity and specificity for the five tissue types of interest. Once the model is validated, it can be applied to new test samples.26Gaigneaux A. Ruysschaert J.M. Goormaghtigh E. Infrared spectroscopy as a tool for discrimination between sensitive and multiresistant K562 cells.Eur J Biochem. 2002; 269: 1968-1973Crossref PubMed Scopus (56) Google Scholar, 29Ly E. Piot O. Durlach A. Bernard P. Manfait M. Differential diagnosis of cutaneous carcinomas by infrared spectral micro-imaging combined with pattern recognition.Analyst. 2009; 134: 1208-1214Crossref PubMed Scopus (58) Google Scholar, 30Nallala J. Diebold M.D. Gobinet C. Bouché O. Sockalingum G.D. Piot O. Manfait M. Infrared spectral histopathology for cancer diagnosis: a novel approach for automated pattern recognition of colon adenocarcinoma.Analyst. 2014; 139: 4005-4015Crossref PubMed Google Scholar In this study, the predictive model was applied to infrared images of 24 new test tumor sections (Table 1) from the 20 remaining tumors. Unidentified test spectra were analyzed by the PCA-LDA model that identified their tissue classes and colored them in accordance with their class color (dark blue color to represent tumor necrosis, yellow for viable tumor, blue for the fibrosis, green for liver parenchyma, and orange for liver parenchyma necrosis). When the maximum probability for a pixel was <0.75, the tissue assignment was considered ambiguous by the model and the pixel was colored in black. The result of the PCA-LDA model analysis is a false color image for which each color corresponds to a type of tissue. The LDA images obtained were quantitatively compared again to control HES-stained adjacent section as for KM images validation step. To confirm the LDA classification, a total of 300 areas measuring 200 μm × 200 μm each were selected on HES-stained tissue sections in zones of viable tumor, tumor necrosis, fibrosis, normal liver parenchyma, and liver parenchyma necrosis (66, 77, 53, 56, and 48 areas, respectively). The number of pixels corresponding to each tissue on LDA images was automatically recorded by the algorithm and used to calculate the percentage of tumor surface occupied by viable tumor or necrotized areas. The values obtained on LDA images were compared with histologic measurements previously obtained on digitalized stained sections from the same paraffinized samples (NanoZoomer 2.0 HT slide scanner at a ×20 objective, Hamamatsu, Hamamatsu City, Japan). Measurements were performed with ICS FrameWork version 2.6 (Tribvn, Châtillon, France). We calculated correlations between the PCA-LDA model tissues quantification and histopathologic measurements using the Spearman coefficient (r) for nonparametric correlation. P < 0.05 was regarded as significant. Because the spectral absorption bands are correlated to biochemical tissue composition, it is possible to determine which biomarkers varied according to the tissue type. The Matlab randfeatures function (available in the MatLab Statistics Toolbox and MatLab Bioinformatics Toolbox) was used to identify the most discriminant infrared wave numbers among our five populations of spectra (tumor necrosis, viable tumor, fibrosis, liver parenchyma, and liver parenchyma necrosis).37Nguyen T.T. Happillon T. Feru J. Brassartâ Passco S. Angiboust JFo Manfait M. Piot O. Raman comparison of skin dermis of different ages: focus on spectral markers of collagen hydration.J Raman Spectrosc. 2013; 44: 1230-1237Crossref Scopus (36) Google Scholar, 38Poplineau M. Trussardi-Régnier Al Happillon T. Dufer J. Manfait M. Bernard P. Piot O. Antonicelli F. Raman microspectroscopy detects epigenetic modifications in living Jurkat leukemic cells.Epigenomics. 2011; 3: 785-794Crossref PubMed Scopus (19) Google Scholar The randfeatures function randomly selects a subset of 15 wave numbers and reduces all the spectra from 451 wave numbers to the 15 sel