{"paper_id":"a142bbdc-0353-4531-b978-15d8936f5950","body_text":"1Scientific  RepoRtS  |         (2019) 9:19795  | https://doi.org/10.1038/s41598-019-56308-y\nwww.nature.com/scientificreports\nRaman spectroscopy as a non-\ninvasive diagnostic technique for \nendometriosis\nUgur parlatan  1,6*, Medine Tuna inanc1,6, Bahar Yuksel ozgor 2, Engin oral 3, Ercan Bastu4, \nMehmet Burcin Unlu1 & Gunay Basar5\nEndometriosis is a condition in which the endometrium, the layer of tissue that usually covers the inside \nof the uterus, grows outside the uterus. One of its severe effects is sub-fertility. The exact reason for \nendometriosis is still unknown and under investigation. Tracking the symptoms is not sufficient for \ndiagnosing the disease. A successful diagnosis can only be made using laparoscopy. During the disease, \nthe amount of some molecules (i.e., proteins, antigens) changes in the blood. Raman spectroscopy \nprovides information about biochemicals without using dyes or external labels. In this study, Raman \nspectroscopy is used as a non-invasive diagnostic method for endometriosis. The Raman spectra of \n94 serum samples acquired from 49 patients and 45 healthy individuals were compared for this study. \nPrincipal Component Analysis (PCA), k- Nearest Neighbors (kNN), and Support Vector Machines \n(SVM) were used in the analysis. According to the results (using 80 measurements for training and 14 \nmeasurements for the test set), it was found that kNN-weighted gave the best classification model \nwith sensitivity and specificity values of 80.5% and 89.7%, respectively. Testing the model with unseen \ndata yielded a sensitivity value of 100% and a specificity value of 100%. To the best of our knowledge, \nthis is the first study in which Raman spectroscopy was used in combination with PCA and classification \nalgorithms as a non-invasive method applied on blood sera for the diagnosis of endometriosis.\nEndometriosis is defined as the growth of endometrial gland and stroma outside the endometrial cavity, which is \ncaused by an outflow into the peritoneal cavity. Previous reports demonstrated that one in ten women all around \nthe world sought medical support due to endometriosis and endometriosis-related symptoms including pelvic \npain (38.7%), dyspareunia (29.5%), and infertility (11.6%)\n1. Given that the diagnosis of endometriosis depends on \nhistopathologic examination after surgical excision, this approach requires anesthesia induction and hospitaliza-\ntion. Therefore, it significantly affects the quality of life of patients. Thus, researchers focus on new non-invasive \nmethods for the diagnosis of endometriosis, including transvaginal ultrasonography, analysis of blood biomark-\ners, and genetic predispositions.\nRaman spectroscopy provides information about molecular structures and chemical bonds of substances via \nthe detection of inelastically scattered photons\n2. In Raman spectroscopy, the sample is illuminated by a laser \nbeam and inelastically scattered light, which is composed of different frequencies, is observed. The scattered light \ncontains two types of scattering, namely Rayleigh and Raman scattering. The intensity of the light in Rayleigh \nscattering is strong and the frequencies of the scattered and the incident light are the same, whereas in Raman \nscattering, the intensity is very weak (about 10\n− 6 of the incident beam intensity) and the frequency of the scattered \nlight is different from the frequency of the incident light. The difference between the frequencies of Rayleigh scat-\ntering and the inelastically scattered photons can be defined as the Raman shift. The Raman shifts correspond to \nthe vibrational frequencies of the molecules in a targeted sample.\nThe vibrational frequencies of each chemical bond within a molecule (e.g., O-H, C-O) are different, hence \ntheir fingerprints can be uniquely seen in the spectrum. It was reported in a study that during the disease, the \namount of protein biomarkers in the blood varied, and these variations could be identified using multiplex and \n1Bogazici University, Physics Department, Istanbul, 34470, Turkey. 2Esenler Maternity and Children’s Hospital \nObstetrics and Gynecology Department, Istanbul, 34230, Turkey. 3Istanbul University Cerrahpasa School of \nMedicine, Reproductive Endocrinology and Infertility Division, Obstetrics and Gynecology Department, Istanbul, \n34301, Turkey. 4Acibadem University School of Medicine, Department of Obstetrics and Gynecology, Istanbul, \n34752, Turkey. 5Istanbul Technical University, Physics Engineering Department, Istanbul, 34469, Turkey. 6These \nauthors contributed equally: Ugur Parlatan and Medine Tuna Inanc. *email: ugur.parlatan@boun.edu.tr\nopen\n\n2Scientific  RepoRtS  |         (2019) 9:19795  | https://doi.org/10.1038/s41598-019-56308-y\nwww.nature.com/scientificreportswww.nature.com/scientificreports/\nsingle immunologic testing technologies 3. In this context, Raman spectroscopy is a useful tool for detecting \nthe chemical content of a sample. Biologic samples such as tissue, blood, and serum are well-suited measure-\nment samples for Raman spectroscopy because chemical changes accompany progressions of most diseases. \nTherefore, Raman spectroscopy has significant potential to provide valuable information to physicians in medical \ndiagnostics\n4.\nStudies have shown that disease diagnostics with Raman spectroscopy is possible for both tissue and blood \nserum samples. Raman spectra of blood serum samples were used to diagnose many types of diseases, including \nAlzheimer’s disease\n5, oral cancer6, nasopharyngeal cancer7, colorectal cancer8, dengue infection9, lung cancer10, \nhepatitis B11, and breast cancer12. Endometriosis, however, has thus far only been studied using Raman spectros-\ncopy through tissue. Lieber et al ., indicated that Raman spectroscopy could differentiate tissues diagnosed as \nnormal or endometriotic from tissues that were diagnosed as benign-cystic or cancerous13. Patel et al. showed that \nstages of endometrial cancer could be distinguished using Raman imaging14. In another study by Notarstefano et \nal., luteinized granulosa cells were measured using Raman micro-spectroscopy to separate ovarian endometriosis \nfrom control samples15.\nRecently, k-Nearest Neighbor (kNN) and Support Vector Machines (SVM) combined with Principal \nComponent Analysis (PCA) have frequently been used together with spectroscopy in disease diagnostics. kNN is \na classification method based on the commonality within groups; every single spectrum can be treated as a point \nin a multidimensional space. This method calculates the Euclidean distance between each pair of spectra points. \nThen, by regarding the majority vote of its nearest neighbors, the class assignment of a sample is performed\n16.\nSupport Vector Machine algorithm is a powerful, supervised learning algorithms, which were introduced by \nVapnik17. It is used as a classification method in which every data element is viewed as a point in n-dimensional \nspace (n is the number of features) with the value of each feature being the value of an individual coordinate. \nClassification of the data is achieved by determination of the hyperplane that maximizes the margin between the \ngroups. It is an elegant approach for the classification of spectral data\n18–21.\nIn some recent studies, classification methods and Raman spectroscopy were used together for disease diag-\nnostics. Dingari et al. reported that Raman spectroscopy and multivariate classification could discriminate lesions \nin stereotactic breast biopsies, irrespective of microcalcification status 22. Li et al . developed a method for the \nnon-invasive detection of colon cancer using Raman spectroscopy together with PCA and kNN23.\nIn this article, we report the first Raman spectroscopy-based classification model that can be used as a \nnon-invasive diagnostic technique for endometriosis. This new approach requires only blood serum from a \npatient with endometriosis for the diagnosis of the disease. Therefore, the diagnosis of endometriosis could be \nachievable without laparoscopy.\nResults and Discussion\nThe mean Raman spectra of the two groups are demonstrated in Fig.  1b. Although the intensity difference \nbetween the groups in the spectral range of 500–750 cm− 1 is apparent, this spectral interval was not used in the \nclassification processes because the signal variance is high in that region. The appropriate region was chosen for \nthe classification using the variable selection procedure, which is described in the methods section. For this pro-\ncedure, the mean accuracy values of the classification models with the standard deviations (given in parentheses) \nwere calculated and are given in Table 1. The final feature selection was decided by considering the region with \nthe highest mean accuracy value, which was found as 790–1729 cm\n− 1 spectral interval. Then, PCA was applied on \nRaman Shift (cm-1)\nRaman Intensity (arb. u.)\n2505 00 7501 0001 2501 5001 750\n0\n0.02\n0.04\n0.06\n0.08\n0.1\nPatient\nControl\n250 500 750 1000 1250 1500 1750\n0\n0.5\n1\n1.5\n2\n2.5 104\nSerum\nWater\nBG (Serum-Water)\nBC (BG-Base Curve)\n(a) (b)\nFigure 1. (a) Background (BG) and baseline-corrected (BC) Raman spectra of a serum sample. (b) Normalized \nBC mean Raman spectra of the control and patient groups. Standard deviations of each group were plotted and \noverlaid as shaded curves.\n\n3Scientific  RepoRtS  |         (2019) 9:19795  | https://doi.org/10.1038/s41598-019-56308-y\nwww.nature.com/scientificreportswww.nature.com/scientificreports/\nthe normalized and baseline corrected Raman spectral data to extract the relevant features for the selected region \n(790–1729 cm− 1). The number of PCs was set in the 95% of the total variance explained (TVE). The percentage \nTVE values for PCs were calculated as 48.3, 17.2, 13.6, 5.2, 4.3, 2.9, 2.1, and 1.6, respectively. This condition \nrequires 8 PCs for this model. All 8 PCs were included in the model. Figure 2a shows the PCA scores of the first \nagainst the third PC to visualize the discrimination of the two groups on the orthogonal feature plane.\nSome of the peaks labeled on the loading graph, given in Fig. 2b, demonstrate shifts and variations, which can \nbe interpreted as changes in structure and the amount of some chemicals in serum during the disease. Among \nthese bands, 1005 cm\n− 1 was tentatively assigned to phenyl ring angular vibrations due to phenylalanine content or \nC=CH bending vibration due to the ground state beta carotene content24. The presence of beta carotene also con-\ntributed to the 1156 and 1520 cm− 1 bands, which are C-C and C=C stretching bond vibrations, respectively25. The \npeak at 1450 cm−1  was assigned to CH2 bending vibration, which exists in lipids, phospholipids, and some amino \nacids26. Besides these, the peaks around 1239 and 1650 cm−1  were assigned to amide III (parallel beta sheet) and \namide I, respectively. They are related to the secondary structure of proteins such as alpha helix (1657 cm − 1), \nparallel β-sheet (1630 cm− 1), and turn (1670 cm− 1)12. The importance of these bands in the diagnosis of endome-\ntriosis is not yet clear, and it is to be investigated in the future. On the other hand, the alteration of the bands at \n1156 and 1520 cm\n− 1 may refer to a change in the amount of beta carotene in the patient group. One explanation \nfor this change could be that the alteration of retinoic acid metabolism in patients with endometriosis27. Taylor et \nal. reported a decrease in carotenoids in endometriotic tissues, which may be provide hope for medical therapies \nas adjuvants or alternatives to the surgical excision28. Therefore, beta carotene, which is an important member of \nthe carotenoid family, may have a protective role against endometriosis.\nAfter PCA, the study was carried a step further to examine the performance of the machine learning algo -\nrithms on the classification of the Raman spectral data. For this purpose kNN (fine and weighted) and SVM \n(cubic and quadratic) were used. The data set included measurements from 49 patients and 45 healthy individu-\nals. The training and the cross-validation (5-fold) data sets were separated by selecting 85% of the total data (con-\ntaining 41 patient and 39 control measurements) randomly. The remaining 15% (including 8 patient and 6 control \nmeasurements) of the data was used as unseen data to assess the predictive power of the classification models.\nThe performance of the applied classification methods in terms of sensitivity, specificity, positive predictive \nvalue (PPV), and negative predictive value (NPV) is presented in Table 2. Sensitivity and specificity are measures \nof classification success in predicting diseased and control specimens, respectively. Detailed explanations of these \nterms are given in Table 3. The results indicated that application of the kNN-weighted algorithm on the spectral \ndata exhibited the highest classification model accuracy among the others. Using this algorithm in the training \nprocedure, 33 of 41 patients and 35 of 39 control samples were correctly classified. During the testing phase, the \nmodel was allowed to guess the correct label (“patient” or “control”) of the unseen datum one by one. The results \nindicated that the model correctly classified 8 of the 8 patients and 6 of the 6 control samples. In short, this result \nindicates a promising potential for the use of Raman spectroscopy together with the kNN-w classification algo-\nrithm for non-invasive diagnostics of endometriosis.\nFeature Selection Mean Accuracy (%)\nRegion (cm−1 ) kNN-f kNN-w SVM-c SVM-q\n450–1729 76.2 (2.9) 78.0 (3.6) 73.8 (4.1) 76.9 (4.3)\n790–1729 79.4 (3.8) 82.1 (2.5) 80.0 (2.3) 82.5 (2.9)\n1140–1729 72.8 (5.1) 77.3 (2.2) 77.5 (3.4) 78.5 (2.2)\n1368–1729 63.3 (1.9) 65.8 (4.2) 68.5 (5.8) 64.5 (1.8)\nTable 1. Comparison of the mean accuracy results of kNN and SVM classification models for the four selected \nregions after 10 repetitions of calculations.\nFigure 2. PCA performance on the training data set, which includes normalized BC data from 41 patients and \n39 healthy individuals. (a) PCA score plot (PC1 vs. PC3) (b) Loading 1 and Loading 3 spectra.\n\n4Scientific  RepoRtS  |         (2019) 9:19795  | https://doi.org/10.1038/s41598-019-56308-y\nwww.nature.com/scientificreportswww.nature.com/scientificreports/\nconclusion\nDeveloping a non-invasive method for endometriosis is challenging and currently under investigation. There \nare new strategies for improving transvaginal ultrasonography skills to diagnose mostly deep infiltrating endo-\nmetriosis. Biomarker or genetic predisposition studies are being published in a growing manner. Laparoscopy is \nthe most secure way to diagnose endometriosis, but it is an invasive method requiring such that patients should \nundergo a kind of surgery. Instead, a non-invasive method would be more economical and patient-friendly for \nthe diagnosis of endometriosis. In this respect, as it was demonstrated for the first time in this article, Raman \nspectroscopy technique together with PCA and the classification algorithms could be a good candidate as a \nnon-invasive diagnostic method for endometriosis.\nTo further improve this study, one might classify particular spectral bands of the serum spectrum that cor -\nrespond to the suspected biomarkers of endometriosis. However, because there are insufficient literature data \nfor reference Raman signals of all biomarkers of endometriosis (i.e., annexin V , VEGF , CA-125, slCAM-1\n3), the \nRaman spectrum of each suspected biomarker should be measured as the reference spectrum to make more reli-\nable inferences about the disease.\nMethods\nPatient Selection. Forty-nine patients who had a surgical diagnoses of endometriosis and 45 healthy women \nwith no history of pelvic pain or infertility were enrolled in this study after ethical approval was granted by the \nEthics Committee of the Faculty of Medicine, Acibadem University. Each participant gave written informed con-\nsent. All experiments were performed in accordance with relevant guidelines and regulations. Student’s t-test was \napplied on the data of volunteers who joined the study. There were no statistically significant differences between \nthe patient and control groups in terms of age, BMI (body mass index), presence of uterine myomas, and adeno-\nmyosis, as given in Table 4. The patients were not divided into subgroups for the investigation because there is no \nknown account to determine whether the main presenting symptom has a different underlying pathophysiology. \nFour patients had uterine myomas in the patient group, and three women had uterine myomas in the control \ngroup; all were asymptomatic. In the patient group, two patients had adenomyosis. Women with comorbidities, \ndrug users, and patients with pelvic pain that was not proven to be endometriosis and who were not on their \nsecretory phase (16–28th day) of the menstrual cycle were excluded.\nSample Preparation. Blood samples were taken in 10-mL serum separator tube (Vacusera) and centrifuged \nat 1500 g for 10 minutes to isolate the serum. All the serum samples were stored at 4 °C and measured a maximum \nof two days after the collection. For the measurement, approximately 0.5 mL of the serum sample was prepared \nin a quartz cuvette.\nExperimental Setup. The experimental arrangement was built around a home-built microscope that \nincluded a water immersion microscope objective (60X, NA, Olympus). A single mode diode laser (CrystaLaser) \nwith wavelength 785 nm and power 100 mW was used for Raman excitation. The unwanted back-reflected beams \nwere filtered using a Faraday isolator (FI, EOTech), which was placed in front of the diode laser. A laser line \nTraining kNN-f(a) kNN-w(b) SVM-c(c) SVM-q(d)\nSpecificity 84.6 (33/39) 89.7 (35/39) 84.6 (33/39) 87.1 (34/39)\nSensitivity 78.0 (32/41) 80.5 (33/41) 75.6 (31/41) 75.6 (31/41)\nPPV 84.2 (32/38) 89.2 (33/37) 75.6 (31/35) 83.8 (31/37)\nNPV 78.6 (33/42) 81.4 (35/43) 83.8 (34/45) 76.7 (33/43)\nTest kNN-f(a) kNN-w(b) SVM-c(c) SVM-q(d)\nSpecificity 100 (6/6) 100 (6/6) 100 (6/6) 100 (6/6)\nSensitivity 87.5 (7/8) 100 (8/8) 87.5 (7/8) 87.5 (7/8)\nTable 2. Comparison of the predictive ability of kNN and SVM classification models. All results are given \nin percentages. Information given in parentheses represents the ratio of number of correct predictions to the \nnumber of true class measurements. \n(a)fine, (b)weighted, (c)cubic, (d)quadratic.\nActual Positive\n(P)\nActual Negative\n(N)\nPredicted Positive True Positive\n(TP)\nFalse Positive\n(FP)\nPPV\nTP/(TP + FP)\nPredicted Negative False Negative\n(FN)\nTrue Negative\n(TN)\nNPV\nTN/(TN + FN)\nSensitivity\nTP/(TP + FN)\nSpecificity\nTN/(TN + FP)\nAccuracy\n(TP + TN)/(P + N)\nTable 3. The definitions of sensitivity, specificity, positive predictive value (PPV), negative predictive value \n(NPV), and accuracy.\n\n5Scientific  RepoRtS  |         (2019) 9:19795  | https://doi.org/10.1038/s41598-019-56308-y\nwww.nature.com/scientificreportswww.nature.com/scientificreports/\nfilter (LF) was employed to obtain a clean laser profile around 785 nm (Semrock, LL01-780-12.5). The sample \nwas illuminated through a focusing lens and the back-scattered light at 180 °C geometry was collected using \nthe same lens. The laser power on the sample was detected around 70 mW . The Rayleigh scattered photons were \nfiltered using two sequentially located Raman edge filters (Semrock). The Raman scattered beam was focused on \na 100 μm slit of a spectrometer (f = 303 mm, f#4.3, Andor) using an achromatic lens with a focal length 50 mm. \nThe spectrometer was equipped with a 600 lines/mm grating and with a thermoelectric-cooled CCD camera (at \n−90 °C, Andor iDus DU420A-OE). A schematic view of the equipment can be seen in Fig. 3.\nExperiment and Analysis. The measurement and spectral analysis scheme is given in Fig. 4 . According \nto this scheme, first, the toluene spectrum was measured using an exposure time of 0.2 s for wavenumber cali-\nbration. Secondly, the distilled water spectrum was acquired using an exposure time of 30 s with 14 successive \nscans. The average of the water spectra was used for the background subtraction. Next, the Raman spectra of the \nserum samples were measured using the same integration parameters as with the water measurements. Each \nserum sample was measured twice sequentially. After cosmic-ray removal from the spectral data, 14 scans were \ndecreased to 10 scans by excluding those with higher variance, and then these ten scans were averaged for each \nmeasurement. The spectra, which belonged to the same volunteer, were then averaged. Thereby, the data under-\nwent pre-processing through a graphical user interface (GUI) that we wrote on the MATLAB platform. The \nGUI performs the pre-processing steps, namely calibration, background (BG), and baseline correction (BC), as \ndemonstrated in Fig.  1a. The developed wavenumber calibration method, which uses the Raman spectrum of \ntoluene, was applied\n26. The reference bands of the toluene spectrum were used to calibrate the distilled water and \nserum spectra 2. The distilled water spectra were subtracted from the corresponding serum spectra to exclude \nsignals coming from the water and cuvette. This step makes the spectrum background-corrected (BG). After the \nBG correction, there still remain auto-florescence signals coming from the serum sample. To further exclude \nthese unwanted signals, baseline correction was applied for each spectrum by fitting a cubic spline curve on the \nselected 12 wavenumber points on the spectrum. To perform baseline subtraction, the selected wavenumbers \n(corresponding to the data points) were identical for each spline curve to ensure objectivity for each sample. \nAfterwards, the spline curve was subtracted from the BG spectrum to obtain the baseline-corrected (BC) spec-\ntrum (Fig. 1a). Then, vector normalization was applied for each BC spectrum. The mean spectra of the normal-\nized BC data of the two groups can be viewed in Fig. 1b.\nTo further explore the data, PCA was applied to the vector-normalized BC data. This is a method for data \ndescription and compression, which is useful for reducing the dimension of large data sets while preserving most \nof the information. Its discriminating power for grouping data into clusters makes PCA noteworthy for diagnos-\ntic studies. After PCA analysis, built-in MATLAB functions were used to apply kNN (fine and weighted) and \nSVM (cubic and quadratic) classification methods to construct classification models. The feature selection was \nperformed by re-constructing all the models 10 times for the selected regions because the 5-fold cross-validation \nalgorithm of MATLAB’s classification software is a random process. The standard deviation and the mean \naccuracy values for each model determined and the best interval of the spectrum, on which the accuracy of \nControl group Patient group p-value\n# of Volunteers 45 49\nAdenomyosis (n) 0 2 (4.08%) 0.290\nUterine myoma (n) 3 (6.60%) 5 (10.20%) 0.561\nBMI 25.53.3 24.63.6 0.179\nMean Age (years) 27.17.8 29.45.4 0.315\nTable 4. Demographic data for the patient and control groups. (n): number of patients with myomas/\nadenomyosis. BMI: body mass index. Confidence level: 0.95.\nFigure 3. The experimental arrangement for Raman spectroscopy.\n\n6Scientific  RepoRtS  |         (2019) 9:19795  | https://doi.org/10.1038/s41598-019-56308-y\nwww.nature.com/scientificreportswww.nature.com/scientificreports/\nclassification methods were the highest, were calculated. The average and the standard deviation values of the \naccuracy calculations are shown in Table 1. After feature selection, 85% of the total spectral data was selected as \na training group, which included 41 patient and 39 control measurements. Then, the remaining 15% was set as \ntest data, which contained 8 patient and 6 control measurements. By concerning the training and test results, the \nspecificity, sensitivity, PPV , NPV , and the accuracy of the classification models were calculated according to the \nequations given in Table 3.\nData availability\nThe corresponding author can provide the datasets of this study upon reasonable request.\nReceived: 10 April 2019; Accepted: 10 December 2019;\nPublished: xx xx xxxx\nReferences\n 1. Fuldeore, M. J. & Soliman, A. M. Prevalence and symptomatic burden of diagnosed endometriosis in the united states: national \nestimates from a cross-sectional survey of 59,411 women. Gynecologic and obstetric investigation 82, 453–461 (2017).\n 2. Ferraro, J. R. Introductory Raman spectroscopy (Elsevier, 2003).\n 3. Vodolazkaia, A. et al. Evaluation of a panel of 28 biomarkers for the non-invasive diagnosis of endometriosis. Human reproduction \n27, 2698–2711 (2012).\n 4. Choo-Smith, L.-P . et al . Medical applications of raman spectroscopy: from proof of principle to clinical implementation. \nBiopolymers: Original Research on Biomolecules 67, 1–9 (2002).\n 5. Ryzhikova, E. et al. Raman spectroscopy of blood serum for alzheimer’s disease diagnostics: specificity relative to other types of \ndementia. Journal of biophotonics 8, 584–596 (2015).\n 6. Sahu, A., Sawant, S., Mamgain, H. & Krishna, C. M. Raman spectroscopy of serum: an exploratory study for detection of oral \ncancers. Analyst 138, 4161–4174 (2013).\n 7. Feng, S. et al. Nasopharyngeal cancer detection based on blood plasma surface-enhanced raman spectroscopy and multivariate \nanalysis. Biosensors and Bioelectronics 25, 2414–2419 (2010).\n 8. Lin, D. et al. Colorectal cancer detection by gold nanoparticle based surface-enhanced raman spectroscopy of blood serum and \nstatistical analysis. Optics express 19, 13565–13577 (2011).\n 9. Mahmood, T. et al. Raman spectral analysis for rapid screening of dengue infection. Spectrochimica Acta Part A: Molecular and \nBiomolecular Spectroscopy 200, 136–142 (2018).\n 10. Huang, Z. et al . Near-infrared raman spectroscopy for optical diagnosis of lung cancer. International journal of cancer  107, \n1047–1052 (2003).\n 11. Kamińska, A. et al. Detection of hepatitis b virus antigen from human blood: Sers immunoassay in a microfluidic system. Biosensors \nand Bioelectronics 66, 461–467 (2015).\n 12. Pichardo-Molina, J. et al. Raman spectroscopy and multivariate analysis of serum samples from breast cancer patients. Lasers in \nmedical science 22, 229–236 (2007).\n 13. Lieber, C. A., Molpus, K., Brader, K. & Mahadevan-Jansen, A. Diagnostic tool for early detection of ovarian cancers using raman \nspectroscopy. In Biomedical Spectroscopy: Vibrational Spectroscopy and Other Novel Techniques, vol. 3918, 129–135 (International \nSociety for Optics and Photonics, 2000).\n 14. Patel, I. I. et al. High contrast images of uterine tissue derived using raman microspectroscopy with the empty modelling approach \nof multivariate curve resolution-alternating least squares. Analyst 136, 4950–4959 (2011).\n 15. Notarstefano, V . et al. Vibrational characterization of granulosa cells from patients affected by unilateral ovarian endometriosis: New \ninsights from infrared and raman microspectroscopy. Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy 212, \n206–214 (2019).\nFigure 4. Infographic for pre-processing and the data analysis.\n\n\n7Scientific  RepoRtS  |         (2019) 9:19795  | https://doi.org/10.1038/s41598-019-56308-y\nwww.nature.com/scientificreportswww.nature.com/scientificreports/\n 16. Cover, T. & Hart, P . Nearest neighbor pattern classification. IEEE transactions on information theory 13, 21–27 (1967).\n 17. Vapnik, V . N. The Nature of Statistical Learning Theory. (Springer-Verlag, Berlin, Heidelberg, 1995).\n 18. Lin, W . et al. Classification of in vivo  autofluorescence spectra using support vector machines. Journal of Biomedical Optics 9, \n180–187 (2004).\n 19. Palmer, G. M. et al. Comparison of multiexcitation fluorescence and diffuse reflectance spectroscopy for the diagnosis of breast \ncancer (march 2003). IEEE Transactions on Biomedical engineering 50, 1233–1242 (2003).\n 20. Majumder, S., Ghosh, N. & Gupta, P . K. Support vector machine for optical diagnosis of cancer. Journal of Biomedical Optics 10, \n024034 (2005).\n 21. Zomer, S., Brereton, R. G., Carter, J. F . & Eckers, C. Support vector machines for the discrimination of analytical chemical data: \napplication to the determination of tablet production by pyrolysis-gas chromatography-mass spectrometry. Analyst 129, 175–181 \n(2004).\n 22. Dingari, N. C. et al . Development and comparative assessment of raman spectroscopic classification algorithms for lesion \ndiscrimination in stereotactic breast biopsies with microcalcifications. Journal of biophotonics 6, 371–381 (2013).\n 23. Li, X. et al. Raman spectroscopy combined with principal component analysis and k nearest neighbour analysis for non-invasive \ndetection of colon cancer. Laser Physics 26, 035702 (2016).\n 24. Hernández, B., Pflüger, F ., Kruglik, S. G. & Ghomi, M. Characteristic raman lines of phenylalanine analyzed by a multiconformational \napproach. Journal of Raman Spectroscopy 44, 827–833 (2013).\n 25. Vítek, P ., Osterrothová, K. & Jehlička, J. Beta-carotene—a possible biomarker in the martian evaporitic environment: Raman micro-\nspectroscopic study. Planetary and Space Science 57, 454–459 (2009).\n 26. Başar, G. et al. Investigation of preeclampsia using raman spectroscopy. Journal of Spectroscopy 27, 239–252 (2012).\n 27. Y amagata, Y . et al. Retinoic acid has the potential to suppress endometriosis development. Journal of ovarian research 8, 49 (2015).\n 28. Taylor, R. N., Kane, M. A. & Sidell, N. Pathogenesis of endometriosis: roles of retinoids and inflammatory pathways. In Seminars in \nreproductive medicine, vol. 33, 246–256 (Thieme Medical Publishers, 2015).\nAcknowledgements\nThis study was supported by the grants from The Council of Higher Education (Ph.D. fellowship) and Scientific \nand Technological Research Council of Turkey (Project number: 118S113).\nAuthor contributions\nG.B., M.B.U. and U.P . conceived the study. G.B., together with U.P ., directed the study. M.T.I. conducted the \nexperiments. U.P . and M.T.I. analysed the results. E.O., E.B., B.Y .Ö. and M.T.I. organized the sample collection. \nAll authors reviewed the manuscript.\ncompeting interests\nThe authors declare no competing interests.\nAdditional information\nCorrespondence and requests for materials should be addressed to U.P .\nReprints and permissions information is available at www.nature.com/reprints.\nPublisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and \ninstitutional affiliations.\nOpen Access This article is licensed under a Creative Commons Attribution 4.0 International \nLicense, which permits use, sharing, adaptation, distribution and reproduction in any medium or \nformat, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Cre-\native Commons license, and indicate if changes were made. The images or other third party material in this \narticle are included in the article’s Creative Commons license, unless indicated otherwise in a credit line to the \nmaterial. If material is not included in the article’s Creative Commons license and your intended use is not per-\nmitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the \ncopyright holder. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/.\n \n© The Author(s) 2019","source_license":"CC0","license_restricted":false}