Int J Clin Exp Med 207;0(2):2925-293 www.ijcem.com /ISSN:940-590/IJCEM004693 Original Article Detection of lymp node metastases in colangiocanma by fourier transform infrared spectroscopy Min Wu *, Huan-Rong Hu 2*, Long Cui, Zi Xu Department of General Surgery, Peking University Tird Hospital, Beijing 009, Cina; 2 Department of Hematology, National Center for Clinical Laboratories and Beijing Hospital, Beijing 00730, Cina; * Equal contributors. Received July 5, 206; Accepted December, 206; Epub February 5, 207; Publised February 28, 207 Abstract: Background: Te fast development of Fourier transform infrared (FTIR) spectroscopy as proposed a novel avenue for distinguising cancerous tissue from normal one. Te aim of tis researc is to investigate te possibility of employing FTIR spectroscopy wit support vector macine (SVM) classification as a new spectral discriminant metod for classifying metastatic and non-metastatic lymp nodes. Metods: A total of 63 lymp nodes, from colangiocarcinoma patients wo underwent surgery for colangiocarcinomaectomy, were obtained and ten measured by FTIR spectroscopy. Te average FTIR spectrum of metastatic and non-metastatic lymp nodes was made. Standard normal variate (SNV) metod was exploited to reduce scatter effect, and support vector macine (SVM) classification model was utilized to discriminate spectrum of metastatic from non-metastatic lymp nodes. Leaveone-out cross validation (LOOCV) was used to evaluate SVM model effects. Results: 20 metastatic and 43 nonmetastatic lymp nodes were patologically diagnosed. Two parameters, including I 743 (related to Lipid) and I 080 (related to Nucleic acid), are founded statistically different between metastatic and non-metastatic lymp nodes. Te average FTIR spectra of metastatic group was pronouncedly different from non-metastatic group. Te accuracy, sensitivity, and specificity of te SVM model were 88.9%, 75.0%, 95.3%, respectively. Conclusion: A novel approac to differentiate metastatic and non-metastatic lymp nodes in colangiocarcinoma by employing te FTIR spectroscopy combined wit SVM model, is establised and demonstrated, and could be applied in colangiocarcinoma clinical stage value in future clinical practice. Keywords: Colangiocarcinoma, lymp node metastases, fourier transform infrared spectroscopy, standard normal variate metod, support vector macine classification model Introduction Cancer is a serious ealt problem disease tat brings about lots of deats in worldwide []. Altoug colangiocarcinoma as a relatively low incidence rate, as one of te most aggressive uman carcinomas, it as te clinical features of a ig mortality rate and a low long-term survival rate [2]. Tus, te corrective valuation of colangiocarcinoma clinical stage and timely following treatment is extremely important for improving colangiocarcinoma patients prognosis. Te tumor node metastasis (TNM) system classification and clinical staging system is an excellent predictor for te prognosis of colangiocarcinoma and widely used nowadays [3]. TNM staging is closely associated wit lymp node metastases. Tus, it is extremely important to measure and identify metastatic from non-metastatic lymp nodes. Current widely applied metods for assessing lymp nodes include frozen-section and istopatology, wic ave some disadvantages including being laborious intensive and time-consuming [4]. Fourier transform infrared (FTIR) spectroscopy is an absorption spectra tecnology, wic can reflect te subtle vibrational canges of cellular biomolecules composition or structure, including nucleic acid, protein, carboydrate, and lipid [5]. Tis biocemical information is closely tied wit to te carcinogenesis of biological tissues. Tus, FTIR spectroscopy as been widely applied to detect various kinds of tumor, including leukemia [6], breast cancer [7], gastric can-
Table. Basic information of te investigated patients Information Sex cer [8] and lymp node metastases in gastric cancer [9]. Several investigations sow tat FTIR spectroscopy can serve as a reliable metod of cancer detection, a diagnostic tool for differentiating cancerous from normal tissues [0], and even can be used to definite surgical boundary boundaries of cancer resection []. However, te way to identify FTIR spectra between benign and malignant biomedical tissues is difficult and trivial, for te spectral data are extremely complex and complicated. Wile support vector macine (SVM) classification is a powerful binary model tat can represent data as points in te ig dimensional space and separate different categories by maximizing distances to its closest data points [2]. Tis classification model as te capability to identify and learn igly nonlinear relationsips from input data. SVM-based classifier approaces ave been employed to predict uman epidermal growt factor receptor (EGFR/ErbB-) inibitors [3] and ave been proposed for detecting Piwi-interacting RNAs (pirnas) wit te accuracy of 98% [4]. Te support vector macine classification is one kind type of supervised macine learnings, and as been widely applied in large datasets analysis among te biological and oter scientific fields during last decades [5]. Te purpose of tis study was to investigate te possibility of employing FTIR spectroscopy as a potential approac for classifying metastatic and non-metastatic lymp nodes of colangiocarcinoma in combination wit support vector macine (SVM) classification, wit te intent to get a ig classification accuracy. Materials and metods Tissue samples Number Male 25 Female 38 Lymp nodes Metastatic 20 Non-metastatic 43 A total of 63 lymp nodes were obtained from colangiocarcinoma patients wo underwent surgery for colangiocarcinoma ectomy in te Surgery Department of Peking University Tird Hospital, Cina. At room temperature, all samples were cut into cubes (.0 cm.0 cm.0 cm) and later measured on an attenuated total reflection (ATR) detection accessory wic ad attaced to spectrometer in te sortest possible time. After caracterization by te FTIR spectroscopy, tese samples were sent for definitive istopatology by serial sectioning immediately. Patologic diagnosis was performed by two independent patologists. All procedures were approved by te Peking University Biomedical Etics Committee and Institutional Review Board of Peking University Tird Hospital (IRB0000052-034), and written informed consent was obtained from eac patient. Collection of FTIR spectra Te FTIR spectra of te specimens were measured by means of a WQF-660 FTIR spectrometer (Beijing Rayleig Analytical Instrument Corporation, Beijing, Cina) wit ATR accessory. A Spectra-Tec mid-ir optical fiber was utilized to connect ATR wit FTIR spectrometer. Scan wave number srange is from 800 cm - to 000 cm - at a resolution of 8 cm -. 32 spectrums were collected from eac specimen and ten were combined to one for acieving an acceptable signal-to-noise ratio. Statistical analysis metod Spa Pro version 2.2 software (College of Cemistry and Molecular Engineering, Peking University, Beijing, Cina) was used to baseline correct, smoot te spectra, and measure te peak intensity of eac spectra band. SPSS version 20 software (IBM, Armonk, New York, USA) was used for statistical analysis. Tests of normal distribution and variance of omogeneity were performed for all parameters. Normally distributed data were analyzed wit Student s t test. In all instances P<0.050 was considered to be statistically significant. Support vector macine (SVM) classification were carried out to distinguis te FTIR spectra between metastatic and non-metastaticlymp nodes. Before performing SVM, standard normal variate (SNV) metod was adopted to preprocess spectroscopic data sets for reducing effects of baseline sift and non-specific scatter at te surface of te samples [6]. MATLAB 2926 Int J Clin Exp Med 207;0(2):2925-293
Table 2. Preliminary assignments of caracteristic bands of FTIR spectra Peak position (cm - ) Vibrations of te groups Reference substances 743 V (C=O) Lipid 640 Amide I band Protein 550 Amide II band Protein 460 Δ (C-H) Lipid 400 Δ (C-H), δ (C-O-H) Lipid 250 ν as PO 2 - Nucleic acid 60 Ν (C-O), δ (C-O-H), δ (C-O-C) Carboydrate 080 ν s PO 2 - Nucleic acid ν as, asymmetric stretcing vibration, ν s, symmetric stretcing vibration, δ, bending vibration. Table 3. Spectral parameters comparison between metastatic and non-metastatic lymp nodes Parameters Metastatic Non-metastatic N Mean ± SD N Mean ± SD t value I 743 20 0.08±0.008 43 0.034±0.024-2.807 0.023 I 640 20 0.29±0.084 43 0.257±0.07.249 0.27 I 550 20 0.05±0.035 43 0.02±0.026 0.336 0.738 I 460 20 0.02±0.005 43 0.0±0.006 0.049 0.96 I 400 20 0.02±0.006 43 0.0±0.005 0.520 0.605 I 250 20 0.00±0.004 43 0.008±0.005.34 0.94 I 60 20 0.005±0.004 43 0.007±0.005 -.584 0.8 I 080 20 0.08±0.009 43 0.007±0.005 3.242 0.002 I: peak intensity. Figure. Average Fourier transform infrared spectrum from metastatic and non-metastatic lymp nodes (wave number between 800-000 cm - ). Te intensity of non-metastatic at wave number 743 cm - (related to lipid) is significantly iger tan metastatic s. Wile te intensity of non-metastatic at wave number 640 cm - (related to protein) and 080 cm - (related to nucleic acid) is muc lower tan metastatic s. Tese differences strongly contribute to distinguis non-metastatic and metastatic lymp nodes. P ple i at wavenumber k could be standard normalized as (): x ik, SNV = - _ xi, k xi 2 / p - p 2 _ x - () / i, k xi k = Were _ xi means te average of spectroscopic data of sample i, wile p is te number of te wavelengt, and (p-) is te freedom degrees. Ten SVM algoritm, wic employs a non-linear mapping to transform te original training data into iger dimensional data, was carried out to distinguis metastatic and non-metastatic groups [7]. Te principle is explained as following [8]. For labeled training data of te form (x i, y i ) i {,,n} were x i is an n-dimensional Feature vector and y {-,} te labels, a decision function is found representing a separating yperplane defined as (2): fx = ~z, xi + b (2) were ω is te weigt vector, b is te bias value, and Φ(x) is te kernel function. By projecting te data using a mapping Φ(x), nonlinear decision boundaries in te in-put data space can be obtained. Finding te yperplane wile maximizing te margin is formulated as te following optimization problem: subject to: min 2 N ~ S~ + C/ pi i = y i ~z, x + i b $ - pi, pi $ 0, i =,..., N (3) R203a (MatWorks, Inc., Natick, Mass., USA) was used for SNV pre-processing and SVM model building. Te spectroscopic data of samwere C is te cost parameter constant, i is parameter for andling non-separable data, and te index i labels te N training cases. Note 2927 Int J Clin Exp Med 207;0(2):2925-293
dard, and were compared wit te classification from te SVM. Te sensitivity, specificity and accuracy of SVM metod were calculated. Te formulas were as follows: Sensitivity = TP/(TP+FP) Specificity = TN/(TN+FN) Accuracy = (TP+TN)/(TP+TN+ FP+FN) Figure 2. Average Fourier transform infrared spectrum from metastatic and non-metastatic lymp nodes (wave number between 800-350 cm - ). Te differences of intensity between metastatic and non-metastatic are more obvious, especially at wave number 743 cm -, 640 cm - and 550 cm -. were TP means te true positive ones, FP means te false positive ones, TN means te true negative ones, FN means te false negative ones. Figure 3. Average Fourier transform infrared spectrum from metastatic and non-metastatic lymp nodes (wave number between 350-000 cm - ). Te differences of intensity between metastatic and non-metastatic at wave number 60 cm - (related to carboydrate) and 080 cm - (related to nucleic acid) are more visible. tat y {-,} is te class labels, and x i is te independent variables. Nonlinear classification using te Gaussian kernel as following equation (4) was investigated: K x, xi = exp - x - xi v 2 c m (4) were v is te kernel widt, wic controls te amount of te local influence of support vectors on te decision boundary. Results and discussion Te basic information of patients participated is summarized in Table. Eigt peaks were identified and given a preliminary assignment reated to biomolecules including nucleic acid, protein, lipid or carboydrate, as reported in Table 2. Eigt peak intensity (I) parameters, wic could be used to rougly measure te relative content of a biomolecule, were calculated. Te basic teory of FTIR spectroscopy differentiating cells or tissues is tat, FTIR spec- A retrospective validation and leave-one-out cross-validation were also used to evaluate te discriminatory power of SVM metod. Te results of patologic diagnosis served as stan- trum as its unique band spectral properties, wic could indirectly reflect te alterations about te conformation of functional groups, te order of cemical bonds, te amount of ydrogen bonding and te secondary protein structure in cells. Two parameters, including I 743 (related to Lipid) and I 080 (related to Nucleic acid), are founded statistically different between metastatic and nonmetastatic lymp nodes, as is presented in Table 3. Te average FTIR spectrum of metastatic and non-metastatic groups of lymp nodes after preprocessing of SNV is sown in Figures -3, in wic te differences between FTIR spectra of metastatic and non-metastatic groups become pronounced. Particularly, peak intensities of metastatic group at te wave 2928 Int J Clin Exp Med 207;0(2):2925-293
Table 4. SVM discrimination compared wit patologic results of lymp node metastases in colangiocarcinoma SVM classifier results Metastatic Patologic results Non-metastatic Sensitivity Specificity Accuracy Metastatic 5 2 75.0% 95.3% 88.9% Non-metastatic 5 4 Figure 4. ROC analysis of SVM model: te value of AUC is 0.889, 95% CI [0.826, 0.953], P<0.000. numbers related to lipid, protein and nucleic acid, are iger tan tat of non-metastatic group. All of tese above findings indicate tat, tere could be a faster cellular metabolism, enanced cell proliferation wit more nucleic acid produced and an increased expression of proteins in cancer cell, as we ad reported in previous works [9, 20]. SVM discriminant model was applied to classify metastatic and non-metastatic groups of lymp nodes and LOOCV was utilized to evaluate te efficiency of SVM model. Comparison between FTIR spectroscopy tecnique and standard patologic diagnosis is presented in Table 4. As is presented, of te 43 known cases non-metastatic samples, 4 cases were correctly classified, wit 2 samples being misjudged; among from te 20 known cases of metastatic samples, 5 cases were correctly metastases in breast carcinoma by Fourier transform infrared spectroscopy [2]. However, te ultimate goal of developing tis novel approac is not to replace istopatology wit te FTIR tecnique, but to offer anoter effective way of distinguising metastatic from nonmetastatic lymp nodes. Additionally, as ortage of tis pilot researc is tat, te total number of lymp nodes studied is not large. Tus, te next step we ave planned to do is to increase te number of specimen in a prospective multicenter study. Toug tis step, te SVM classification model would become more powerful and furter improvements in sensitivity and accuracy. Conclusions classified, wit 5 samples being misjudged. Te accuracy, sensitivity, and specificity of diagnostic model were 88.9%, 75.0%, 95.3%, respectively, wic clearly indicate tat tis novel approac could well differentiate te FTIR spectra of metastatic versus non-metastatic groups. Te ROC analysis of SVM model is presented in Figure 4. Terefore, FTIR spectroscopy combined wit te SVM classifiers system as te potential to offer a new, safe and effective approac in for te diagnosis and classification of metastatic and non-metastatic lymp nodes and elping pysicians to better to value patient s clinical stage based on te number of metastatic lymp nodes and to practice corresponding treatments. Te results are in confirmation of sentinel lymp node In tis study, FTIR spectra caracteristics of metastatic and non-metastatic lymp nodes in 2929 Int J Clin Exp Med 207;0(2):2925-293
colangiocarcinoma ad been illustrated. A new diagnostic approac, FTIR spectroscopy combined wit support vector macine classification, can acieve a ig discrimination accuracy. Te results indicate tat FTIR spectroscopy wit SVM is practical, and could be applied in colangiocarcinoma clinical stage value in future clinical practice. Acknowledgements Tis study was supported by grants from Major Researc Project of Peking University Tird Hospital (BYSY20207). Disclosure of conflict of interest None. Address correspondence to: Zi Xu, Department of General Surgery, Peking University Tird Hospital, Beijing 009, Cina. E-mail: xuzi23456@sou. com References [] Siegel R, Ma J, Zou Z and Jemal A. Cancer statistics, 204. CA Cancer J Clin 204; 64: 9-29. [2] Luvira V, Nilprapa K, Budisawasdi V, Pugkem A, Camadol N and Kamsa-ard S. Colangiocarcinoma patient outcome in norteastern Tailand: single-center prospective study. Asian Pac J Cancer Prev 206; 7: 40-406. [3] In: Edge SB, Byrd DR, Compton CC, Fritz AG, Greene FL, Trotti A, editors. AJCC cancer staging manual. 7t edition. New York: Springer; 200. [4] Weiser MR, Montgomery LL, Susnik B, Tan LK, Borgen PI and Cody HS. Is routine intraoperative frozen-section examination of sentinel lymp nodes in breast cancer wortwile? Ann Surg Oncol 2000; 7: 65-655. [5] Henry H Mantsc, Dennis Capman. Infrared Spectroscopy of Biomolecules. Wiley-Liss: New York; 996; 24: 87-88. [6] Seng DP, Liu XC, Li WZ, Wang YC, Cen XL and Wang X. Distinction of leukemia patients and ealty persons serum using FTIR spectroscopy. Spectrocim Acta A Mol Biomol Spectrosc 203; 0: 228-232. [7] Tian PR, Zang WT, Zao HM, Lei YT, Cui L, Wang W, Li QB, Zu Q, Zang YF and Xu Z. Intraoperative diagnosis of benign and malignant breast tissues by fourier transform infrared spectroscopy and support vector macine classification. Int J Clin Exp Med 205; 8: 972-98. [8] Li QB, Wang W, Ling XF and Wu JG. Detection of gastric cancer wit fourier transform infrared spectroscopy and support vector macine classification. Biomed Res Int 203; 203: 609-63. [9] Bai YK, Yu LW, Zang L, Fu J, Leng H, Yang XJ, Ma JQ, Li XJ, Li XJ, Zu Q, Zang YF, Ling XF and Cao WL. Researc on application of fourier transform infrared spectrometry in te diagnosis of lymp node metastasis in gastric cancer. Guang Pu Xue Yu Guang Pu Fen Xi 205; 35: 599-602. [0] Sun X, Xu Y, Wu J, Zang Y and Sun K. Detection of lung cancer tissue by attenuated total reflection-fourier transform infrared spectroscopy-a pilot study of 60 samples. J Surg Res 203; 79: 33-38. [] Zang XQ, Xu YZ, Zang YF, Wang LX, Hou CS, Zou XS, Ling XF and Xu Z. Intraoperative detection of tyroid carcinoma by fourier transform infrared spectrometry. J Surg Res 20; 7: 650-656. [2] Wang ML, Xuan SY, Yan AX and Yu CY. Classification models of HCV NS3 protease inibitors based on support vector macine (SVM). Comb Cem Hig Trougput Screen 205; 8: 24-32. [3] Kong Y, Qu D, Cen XY, Gong YN and Yan AX. Self-organizing map (SOM) and support vector macine (SVM) models for te prediction of uman epidermal growt factor receptor (EGFR/ErbB-) inibitors. Comb Cem Hig Trougput Screen 206; 9: 400-4. [4] Seyeddokt A, Aslaminejad AA, Masoudi-Nejad A, Nassiri M, Zairi J, Sadegi B. Computational detection of pirna in uman using support vector macine. Avicenna J Med Biotecnol 206; 8: 36-4. [5] Tan MX, Pu JT and Zeng B. Optimization of breast mass classification using sequential forward floating selection (SFFS) and a support vector macine (SVM) model. Int J Comput Assist Radiol Surg 204; 9: 005-020. [6] Barnes RJ, Danoa MS and Lister SJ. Standard normal variate transformation and de-trending of near-infrared diffuse reflectance spectra. Applied Spectroscopy 989; 43: 772-777. [7] Zang Y, Ren JC and Jiang JM. Combining MLC and SVM classifiers for learning based decision making: analysis and evaluations. Comput Intell Neurosci 205; 205: 423-58. [8] Keerti SS and Lin CJ. Asymptotic beaviors of support vector macines wit Gaussian kernel. Neural Comput 2003; 5: 667-689. [9] Zang WT, Tian PR, Zu Q, Zang YF, Cui L and Xu Z. Noninvasive surface detection of papillary tyroid carcinoma by fourier transform in- 2930 Int J Clin Exp Med 207;0(2):2925-293
frared spectroscopy. Cemical Researc in Cinese Universities 205; 3: 98-202. [20] Liu Y, Xu Y, Liu Y, Zang Y, Wang D, Xiu D, Xu Z, Zou X, Wu J and Ling X. Detection of cervical metastatic lymp nodes in papillary tyroid carcinoma by Fourier transform infrared spectroscopy. Br J Surg 20; 98: 380-384. [2] Tian P, Zang W, Zao H, Lei Y, Cui L, Zang Y and Xu Z. Intraoperative detection of sentinel lymp node metastases in breast carcinoma by Fourier transform infrared spectroscopy. Br J Surg 205; 02: 372-379. 293 Int J Clin Exp Med 207;0(2):2925-293