Fluorescence spectroscopy as a tool to in vivo discrimination of distinctive skin disorders

Accepted Manuscript Title: Fluorescence Spectroscopy as a tool to in vivo discrimination of distinctive skin disorders Authors: Vitoria Helena Maciel, Wagner Rafael Correr, Cristina Kurachi, Vanderlei Salvador Bagnato, Cacilda da Silva Souza PII: DOI: Reference:

S1572-1000(16)30234-4 http://dx.doi.org/doi:10.1016/j.pdpdt.2017.03.014 PDPDT 930

To appear in:

Photodiagnosis and Photodynamic Therapy

Received date: Revised date: Accepted date:

15-11-2016 19-2-2017 26-3-2017

Please cite this article as: Maciel Vitoria Helena, Correr Wagner Rafael, Kurachi Cristina, Bagnato Vanderlei Salvador, da Silva Souza Cacilda.Fluorescence Spectroscopy as a tool to in vivo discrimination of distinctive skin disorders.Photodiagnosis and Photodynamic Therapy http://dx.doi.org/10.1016/j.pdpdt.2017.03.014 This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

1 TITLE PAGE Fluorescence Spectroscopy as a tool to in vivo discrimination of distinctive skin disorders Vitoria Helena Maciel a, b Wagner Rafael Correr a Cristina Kurachi a Vanderlei Salvador Bagnato a Cacilda da Silva Souza b

a

Institute of Physics of São Carlos, University of São Paulo

Av. Trabalhador São-Carlense, 400 - Parque Arnold Schimidt, São Carlos - SP, 13566-590 – Brasil b

Medical School of Ribeirão Preto, University of São Paulo

Av. Bandeirantes, 3900 - Monte Alegre, Ribeirão Preto- SP, 14048-900 - Brasil

Corresponding author: Cacilda da Silva Souza1 Faculdade de Medicina de Ribeirão Preto da Universidade de São Paulo Divisão de Dermatologia Av. Bandeirantes, 3900 Monte Alegre, Ribeirão Preto, São Paulo, Brasil CEP: 14048-900 E-mail: [email protected]

Word count: 3683 Abstract: 221

Tables: 04

Figures: 04

2

Fluorescence Spectroscopy as a tool to in vivo discrimination of distinctive skin disorders Vitoria Helena Maciel a, b Wagner Rafael Correr a Cristina Kurachi a Vanderlei Salvador Bagnato a Cacilda da Silva Souza b a

Institute of Physics of São Carlos, University of São Paulo

b

Medical School of Ribeirão Preto, University of São Paulo

HIGHLIGHTS 

Fluorescence spectroscopy has potential application to detect modifications of malignant tissues.



k-Nearest Neighbour computational classification is useful to analyze the fluorescence spectra.



One supposedly unknown sample is classified per most common category of k nearest samples



k-NN classifier could differentiate normal or abnormal skin conditions of distinctive etiologies and similar morphology.



Neoplastic or inflammatory conditions could be discriminated with high specificity and sensitivity.



Mainly basal cell subtypes and psoriasis lesions could be distinguished among them.



Discrimination of distinctive erythematous-squamous disorders by k-NN classifier, also showed high specificity and sensitivity.

3 

Fluorescence spectroscopy with k-NN analysis appears to be a promising approach for skin disorders.

Abstract Background: Fast and non-invasive analytical methods, as

fluorescence

spectroscopy, have potential applications to detect modifications of biochemical and morphologic properties of malignant tissues. In this study, we propose to analyze the fluorescence spectra using k-Nearest Neighbours algorithm (k-NN) and ratio of the fluorescence intensity (FI) to differentiate skin disorders of distinctive etiologies and morphologies. Materials and methods: Laser-induced autofluorescence spectra upon excitation at 408 nm were collected from basal cell carcinoma (BCC) subtypes (n=45 /212 spectra), psoriasis (PS) (n=37 /193 spectra) and Bowen’s disease (BD) (n=04 /19 spectra) lesions and respective normal skin at sun-exposed (EXP) and nonexposed (NEXP) sites of the same patient. Results: The mean ratios of FI values at selected wavelengths of emission (FI600nm /FI500nm) were significantly lower in BCC and PS lesions compared to EXP [P=0.0001; P=0.0002, respectively]; but there were no significant differences between abnormal conditions. The analysis of fluorescence spectra using k-NN can discriminate normal or abnormal skin conditions (EXP, BCC, BD, PS) of distinctive etiology, neoplastic or inflammatory (BCC, BD and PS) and morphologies (nodular and superficial BCC, BD and PS) as high as 88% and 93% sensitivity and specificity means, respectively; also, similar erythematous-squamous features (superficial BCC, BD and PS) with 98% and 97% sensitivity and specificity means, respectively. The k-NN computational analysis appears to be a promising approach for distinguish skin disorders. Key words: Fluorescence spectroscopy; Basal cell carcinoma; Psoriasis; Bowen’s disease.

4

Introduction

Non-melanoma skin cancer (NMSC) is the most common type of cancer worldwide, mostly in light-skinned people. Like in other cancer types, the critical factor is the early diagnosis in order to provide proper treatment [1]. It depends on prompt recognition of NMSC, or pre-malignant lesions, which is confirmed through histopathological analysis, the gold standard procedure. Fast and non-invasive analytical methods, as fluorescence spectroscopy (FS), have potential applications to detect modifications of biochemical and morphologic properties of pre- or malignant tissues. FS comprises several techniques which could be auxiliary tools for the conventional diagnostic approaches because of its advantages. In the FS, light is used to excite tissues and the emitted fluorescence provides information about the structure and function of target tissues without disrupting them [2, 3]. The onset of pathological conditions results in changes of physiological processes and in quantity, distribution, and the biochemical properties of fluorophores [3, 4]. Skin autofluorescence (AF) derivatives from native fluorophores have been used as a method for differentiation between normal and unhealthy skin [3, 5-7]. AF from tissues is attributed to biomolecules such as tryptophan (Trp), tyrosine (Tyr), phenylalanine (Phe), reduced form of nicotinamide adenine dinucleotide (NADH), flavin adenine dinucleotide (FAD), endogenous porphyrins and other fluorophores as pyridoxine derivatives, which are linked to metabolic processes; and to collagen, elastin, and more generally proteins, that are related to the structural arrangement of cells and tissues [2, 3, 8]. Variations of techniques based on optical spectroscopy have been proposed for discrimination between normal and malignant cells in colorectal [9, 10], oral [11], cervical [12, 13] and breast [14, 15] cancers. Previously, controversial findings in the discrimination between melanoma and nevi were remarked [16-18]. However, endogenous fluorescence of NMSC, excited in the UV region of the

5 spectrum, showed distinctive features from normal tissue and seemed promising to be exploited for non-invasive diagnostics [19]. Exploratory analysis methods are used to extract information and detect trends in the data set, based on the multivariate measurements of the samples. For spectroscopic diagnosis of cervical intraepithelial neoplasia in vivo, using laserinduced fluorescence spectra at multiple excitation wavelengths, Ramanujam et al. [20] used a multivariate statistical algorithm based on Principal Component Analysis (PCA). It is an unsupervised method, in which the class separation occurs without the need for initial information on the nature of the samples and the goal is to identify natural clusters between the samples.[20]. In the supervised methods, such as k-Nearest Neighbours (k-NN) algorithm, it is necessary to have some initial information on the identity of the samples for the formation of the classes and the objective is to develop a model based on the information contained in the samples. It is a lazy learning algorithm that makes decisions based on the entire training data set. In other words, k-NN is a nonparametric method used for classification, or so identifying which of a set of categories a new sample belongs, based on a data training set containing samples whose category membership is known. k-NN classification finds a group of k samples in the training data set that are closest to the new sample, and bases the assignment of a label on the predominance of a specific class in this neighborhood. There are three key elements of this approach: i) a set of labeled samples, e.g., a set of stored records; ii) a distance or similarity metric to compute distance between samples; and iii) the value of k, the number of nearest neighbors. To classify an unlabeled sample, the distance of this sample to the labeled samples is computed, its k-NN are identified, and the class labels of these nearest neighbors are then used to determine the class label of the new sample [21]. Thus, FS provides the possibility for real-time, non-invasive diagnosis of tumors and other tissue disorders [3] and different exploratory analysis methods are available for using. The purpose of the present study is to analyze laserinduced autofluorescence (LIAF) spectra using ratio analysis and computational

6 classifiers, mainly k-NN, to discriminate some categories of skin conditions, mainly neoplastic (NMSC) from inflammatory conditions represented by psoriasis (PS).

Material and Methods This study was approved by the Institutional Ethics Committee for Human Research of the University Hospital, School of Medicine of Ribeirão Preto, University of São Paulo (USP); the Declaration of Helsinki Principles was followed and patients gave written, informed consent. A spectroscopy system coupled and to a portable computer was assembled with commercial devices in the CePOF, Institute of Physics of São Carlos, USP, only for research purposes. A diode laser (Nichia NDV 4313, Nichia Co., Tokushima, Japan) emitting at 408 nm was used as an excitation source. The system has a Y-type probe (Ocean Optics, Dunedin, FL, USA) with two 600-m optical fibers, one delivering the excitation laser and the other one collecting the reemitted light from the target tissue. A USB2000 spectrometer (Ocean Optics, Dunedin, FL, USA) allowed fluorescence evaluation in a spectral range between 372 nm up to 1046 nm wavelengths, with 1 nm resolution. A long-pass filter at 435 nm (GG435 Schott, Lebanon, PA, USA) was used to remove the backscattered light before entering in the spectrometer; and only the fluorescence emission was evaluated. The fluorescence signal collected by the detector was displayed and stored in a computer using the software OOIBase32 (Ocean Optics, Dunedin, FL, USA). The LIAF spectra were measured in vivo on average from five distinct points from each lesion, including basal cell carcinoma (BCC) subtypes (n=45 lesions/ 212 spectra): nodular (n= 33 lesions/ 153 spectra), superficial (n= 07 lesions/ 36 spectra) and mixed BCC (n=05 lesions/ 23 spectra); Bowen’s disease (BD) (n=04 lesions/ 19 spectra) and psoriasis (PS) (n=37 lesions/ 193 spectra). There was exclusion of the spectra measurements with excessive saturation. In general, it was related to the irregular surface of lesion and the lack of contact of probe tip, mainly of BCC nodular. On the other hand, lesions with several morphological aspects, like PS (erythema, desquamation and infiltration) additional spectral measurements were collected.

7 All BCC and BD lesions included in the study have histopathological confirmation. Every lesion was comparatively analyzed with respective normal sunexposed skin (EXP) and non-exposed skin (NEXP) of the same patient, in total 336 and 311 spectra, respectively. The ratio analysis was performed to average ratio of fluorescence intensity (FI) at selected wavelengths at 600 nm and 500nm (FI600nm/ FI500nm) for each sample of BCC and PS lesions; the second wavelength at 635nm (r 635/500 nm) was used for PS lesions and every lesion was compared to normal sun-exposed skin (EXP). Scripts were written on Matlab 7.5 software (MathWorks, Natick, Massachusetts, USA) to process the spectra. Computational classifiers were used to achieve a method to differentiate skin lesions of distinctive etiology. Before the use of computational classifier, the spectra were normalized by their maxima (or each spectrum was divided by its maximum intensity value). This normalization results in all spectra having a maximum intensity equal to one arbitrary unity and allows the direct comparison among distinct skin phototypes. Normalization allows an effective comparison across a heterogeneous set of samples. It was utilized in order to amplify and compare the spectral characteristics of normal EXP and NEXP skin of different phototypes, and malignant or inflammatory tissues. The resulting normalized intensity for each spectrum has a dimensionless value, and consequently becomes less dependent on the intensity factor. [22] Computational classification was performed using the k-NN. In this algorithm, one unknown sample is classified per most frequent category of k nearest samples, based on similarity calculations. A variation of k-NN considering the weighted-distance contributions of each neighbor was used. Thus, nearer neighbors contribute more to decision than more distant ones. In order to validate the results, 10-fold cross validation technique was used. Performance parameters such as sensitivity and specificity mean values were evaluated for each classifier. Furthermore, the number of neighbors was varied to achieve optimal results. The k-NN classification method relies on comparing an unknown spectrum to a "database" of known spectra. This is performed by calculating the Euclidean

8 distance in a N-dimensional space, which is done in the range of milliseconds to seconds. Since, the acquisition time is in the order of milliseconds, thus, given the operation time, one can acquire several spectra from a lesion in a minute or so.

Results

There was wide variation of the FI values as observed in the spectra data collected from all BCC subtypes lesions (n=45 /212 spectra) (Fig. 1a). The mean spectral curve of BCC lesions in all phototypes was under the EXP and NEXP skin spectral curves (Fig. 1b). Whereas, the mean spectral curve differences between BCC and normal skin at NEXP and EXP skin sites seem higher in phototype I compared to phototypes III/IV (Figs.1c-d).

Likewise, there was wide variation of the FI values as observed in the spectra data collected from PS lesions (n=37/193 spectra) (Fig. 2a). Additionally, the mean spectral curve was variable, it be seemed related to predominance of the clinical aspect of silvery coarse scaly (n=7) or red infiltrative plaques (n=6), or a combination of both patterns (n=24) (Data not shown). In overall comparison, the mean spectral curve from PS lesions was between NEXP skin and EXP skin sites (Fig. 2b).

However, the normalized spectral mean curves based on collected data from all BCC subtypes (n=45 /212 spectra), PS (n=37 /193 spectra) and BD (n=04 /19 spectra) lesions and respective normal skin at EXP sites were quite similar and no differences can be noticed from one spectrum to another. There was not any evident feature in spectra that allows direct discrimination among them.

9 The FI values of all BCC subtypes (n=45) in 600nm and 500nm resulted in a mean ratio (FI600nm/ FI500nm) significantly lower than those of normal skin at EXP (P = 0.0001) and NEXP sites (P = 0.0008). No differences between normal skin at EXP and NEXP sites (P = 0.13) were observed (Fig. 3a). Also, the mean ratios (FI600nm/ FI500nm) and (FI635nm/ FI500nm) obtained with FI values from PS lesions (n=37) were significantly lower compared to those of normal skin at EXP (P = 0.0002; P = 0.0001) e NEXP (P = 0.0003; P = 0.0002) sites, respectively (Fig. 3b).

The k-NN algorithm was used to categorize the spectra. Table 1 shows the results of weighted k-NN classifier for discrimination among spectra from all categories: BCC, BD, PS and normal sun-exposed skin (EXP). These are mean values from the ones obtained from each fold of cross validation.

In the sequence, the k-NN classifier was used to discriminate inflammatory (PS) and neoplastic conditions, BCC subtypes (nodular BCC and superficial BCC) (Table 2); and in another analysis, squamous-erythematous skin conditions (PS, BD, superficial BCC) and nodular BCC subtype were compared to each other (Table 3).

Finally, the k-NN classifier was used to discrimination only among the squamouserythematous skin conditions, inflammatory (PS) or neoplastic (BD, superficial BCC), respectively (Table 4). Even increasing the number of neighbors, the values of sensitivity and specificity have little variation, demonstrating the stability of the classifier.

10

Discussion A non-invasive screening technique is desired to potentially detect lesions at an earlier stage before they can be detected using standard techniques, namely the biopsy procedure that provides the histopathological analysis. AF spectroscopy is a non-invasive optical technique that has been exploited for detection of malignant and premalignant lesions in different tissue types without the need for exogenous compounds. The use of drug-induced fluorescence techniques raises some troubles related to photosensitizer administration, besides side effects and higher costs [23]. Most of measuring systems for spectroscopic diagnosis, both ex-vivo and in-vivo tissues, utilize lasers, filtered lamps or light emitting diode as the light source [23]. Despite

of

controversies,

method

variations

achieved

by

fluorescence

spectroscopy or fluorescence imaging could distinguish normal from malignant cells in different types of cancer [9-15]. The overall decrease and the shift of AF intensity to longer wavelengths have been detected for colorectal malignant tissue samples. [9, 10] Whereas in malignant oral lesions additional peaks at 635, 685, and 705 nm due to the accumulation of porphyrins [11], or significant decrease in spectral intensity for leukoplakia can be observed in the unprocessed spectra [24]. FS proved to be effective in revealing the biochemical composition of tissue and potential useful for skin cancer diagnosis [23]. However, early studies based in fluorescence images upon excitation at 365 nm and normalized fluorescence excitation of in vivo melanomas and naevi did not allow accurate discrimination between them. This proposed technique did not improve the melanomas diagnosis, also naevi and melanomas were misclassified [17]. For early detection of malignant melanoma, only stepwise two-photon excitation of melanin seems to provide some encouraging results [23]. There are conflicting results related to the effectiveness of drug-induced fluorescence technique for discrimination between malignant and non-malignant

11 skin lesions, also to determine the BCC margins, which seems less accurate than clinical examination [23, 25]. Bispectral method combining AF with protoporphyrin IX fluorescence could yield good or partial agreement with the histopathological boundaries of the removed tumours by Mohs micrographic surgery, and this agreement was better compared with using only the PpIX fluorescence [26]. Brancaleon et al. [19] examined in vivo FI of BCC and squamous cell carcinoma (SCC). Both NMSC types, upon excitation at 295 nm or 350 nm wavelength, showed a larger and weaker fluorescence signal, respectively, from the tumor lesion compared to the surrounding normal tissue [19]. At 295 nm wavelength, the stronger emission intensity is mainly due to increased fluorescence of tryptophan moieties in the tumor, which can be attributed to its hyperactivity and/or cell proliferation. On the other hand, the decreasing in FI at 350 nm wavelength was possibly induced by destruction of collagen and elastin crosslinks surrounding the tumor nests in dermis, but not in normal skin [19]. Similar emission and excitation properties of collagen and elastin result in spectra that could not be distinguished. In addition, 410 nm is strongly absorbed by hemoglobin; and the increased blood supply within tumors could explain differences of the emission intensity between malignant and normal tissues [27]. Unfortunately, unlike internal mucosal surfaces, there is a wide variation in the skin color among individuals or even different skin locations from the same patient [27]. The intensity level strongly depends on the melanin content, in various anatomic areas and skin phototype [8]. Since melanin also absorbs the fluorescence emission strongly, this might be responsible for lower accuracy in differentiating normal and cancer tissues in patients with skin type III (darker skin). Lower fluorescence intensities resulted in an extensive overlap between the normal and cancer data sets [27]. In concordance, our results showed as a rule that lighter normal skin type exhibited higher intensity of spectrum. Additionally, we observed marked differences in signal intensities from healthy skin, and both BCC and PS lesions through the ratio analysis, however when used the method of smoothing (normalization) of the

12 spectra, the comparison among all spectra samples were quite similar and only a few differences could be noticed from one spectrum to another. The computational classification k-NN algorithms used in this study is based on similarity calculations between the samples. Therefore, one supposedly unknown sample is classified per most common category of k nearest samples, and a variation of k-NN considers the weighted-distance contributions of each neighbor. Therefore, k-NN classification strongly relies on the representativeness of the database. If it is too particular, or not representative - which means not all categories phototypes, lesions types etc. were represented in the database, the classification would be misled, thus changing the sensitivity and specificity. Additionally, new spectra can be compared with the stablished database and its sensitivity and specificity should be somewhat similar with the obtained ones, since 10-fold cross-validation was used to test the algorithm. Furthermore, the representativeness of the database is crucial in this sort of algorithm. The k-NN classification is somewhat in an ad hoc manner, however the database can be fed or constructed with thousands or hundreds of thousands of spectra and then its representativeness would be more comparable to the universe of data. The self-consistent classification would be achieved with another sort of algorithms as Decision trees, Bayesian algorithm, etc. Indeed, we had tried those selfconsistent methods, given their universality; nevertheless, their yields are poorer when compared to k-NN, mainly due to the high variability in spectral behavior in small numbers, because we had access to circa of 240 neoplastic and inflammatory lesions or normal skin, and 1100 spectra. The performance values (sensitivity and specificity) could be improved by increasing the number of spectra. Our results showed a high degree of discrimination between all categories (BCC, BD, Ps and EXP), by using fluorescence based k-NN algorithms. Despite of the varying the number of k, high average values of the sensitivity and specificity obtained with k (3), 94.1% and 94.7%, were maintained with k (15), 92.1% and 92.5%, respectively. It is essentially different discriminating all lesions from sun-exposed skin and discriminating each lesion among them and among sun-exposed skin, i.e., BCC vs

13 PS vs BD vs sun-exposed skin. If we put in same set all lesions, inflammatory or neoplastic, and try to classify them against sun-exposed normal skin, we would observe a lower sensibility (0.891) and specificity (0.878) for k=11 (data not shown). Notably, the sequential analysis the k-NN classifiers can discriminate the squamous-erythematous lesions of distinctive etiologies, with even better results, such as sensitivity and specificity of 97.8% with k (3) and 99.3% and 99.9% with k (11), respectively. Likewise, the comparative analysis that included lesions of the different etiologies (inflammatory and neoplastic) and morphology features, PS, BD and BCC subtypes, superficial (squamous-erythematous) and nodular BCC, can discriminate each one of these skin conditions. High values of the sensitivity and specificity, above 92% with k (3), indicate this promising technique for the prompt and noninvasive tool for diagnosis of these distinctive skin conditions. Principal component analysis (PCA)-based non-parametric k-NN algorithm was applied in the spectral analysis for discrimination between normal, benign, and malignant ovarian tissues by Kamath et al. (2009) [28]. AF spectra were recorded with 325-nm pulsed laser excitation in vitro specimens. This classification technique can achieve clear discrimination between normal, benign, and malignant ovarian samples, with a specificity, sensitivity, and accuracy of about 100%, 90.9%, and 94.2%, respectively. Our high values of the sensitivity and specificity obtained from clinically similar skin conditions indicate valuable using of the computational analyses through the method purposed. The computational classification k-NN algorithm seems a potential tool application to discriminate some skin conditions. In addition, as it is a supervised non-parametric classifier, the chance of obtaining tendentious results is lower than that of non-supervised. Also, compared to other classification techniques, such as the statistical analysis for excitation emission measurements, measured fluorescence intensities, area under each spectrum, PCA (data not shown), the computational classification k-NN

14 algorithm was successful and efficient to distinguish for example some clinically similar squamous-erythematous lesions, but of different etiologies.

Conclusion In conclusion, the results of the present study suggest that the analysis fluorescence spectra using k-NN computational classifiers can discriminate skin conditions from distinctive etiologies and morphology features, neoplastic and inflammatory conditions, mainly superficial and nodular BCC subtypes, BD and PS, with high specificity and sensitivity. Furthermore, this classification method can be used as a tool for more rigorous methodology to the problem of analyzing the results using fluorescence as auxiliary tool for diagnosis. Based on our observations, the k-NN analyses appears to be a promising approach for distinguish skin disorders.

Financial Support: Optics and Photonics Research Center (CePOF) of Institute of Physics of São Carlos, University of São Paulo, a Research, Innovation and Dissemination Center of São Paulo Research Foundation. No Conflict of interest Acknowledgments This study received support from CePOF (Centro de Pesquisas em Óptica e Fotônica) of Institute of Physics of São Carlos, University of São Paulo, a Research, Innovation and Dissemination Center of São Paulo Research Foundation.

References 1. Mogensen M, Jemec GB. Diagnosis of nonmelanoma skin cancer/keratinocyte carcinoma: a review of diagnostic accuracy of nonmelanoma skin cancer diagnostic tests and technologies. Dermatol Surg. 2007;33(10):1158-74. 2. Ramanujam N. Fluorescence spectroscopy of neoplastic and non-neoplastic tissues. Neoplasia. 2000;2(1-2):89-117.

15 3. Croce AC, Bottiroli G. Autofluorescence spectroscopy and imaging: a tool for biomedical research and diagnosis. Eur J Histochem. 2014;58(4):2461. 4. Richards-Kortum R, Sevick-Muraca E. Quantitative optical spectroscopy for tissue diagnosis. Annu Rev Phys Chem. 1996;47:555-606. 5. Anderson RR, Parrish JA. The optics of human skin. J Invest Dermatol. 1981;77(1):13-9. 6. Brancaleon L, Lin G, Kollias N. The in vivo fluorescence of tryptophan moieties in human skin increases with UV exposure and is a marker for epidermal proliferation. J Invest Dermatol. 1999;113(6):977-82. 7. Gillies R, Zonios G, Anderson RR, Kollias N. Fluorescence excitation spectroscopy provides information about human skin in vivo. J Invest Dermatol. 2000;115(4):704-7. 8. Drakaki E, Borisova E, Makropoulou M, Avramov L, Serafetinides AA, Angelov I. Laser induced autofluorescence studies of animal skin used in modeling of human cutaneous tissue spectroscopic measurements. Skin Res Technol. 2007;13(4):350-9. 9. Sato R, Fujiya M, Watari J, Ueno N, Moriichi K, Kashima S, et al. The diagnostic accuracy of high-resolution endoscopy, autofluorescence imaging and narrow-band imaging for differentially diagnosing colon adenoma. Endoscopy. 2011;43(10):862-8. 10. Liu L, Nie Y, Lin L, Li W, Huang Z, Xie S, et al. Pattern recognition of multiple excitation autofluorescence spectra for colon tissue classification. Photodiagnosis Photodyn Ther. 2013;10(2):111-9. 11. Jayanthi JL, Subhash N, Stephen M, Philip EK, Beena VT. Comparative evaluation of the diagnostic performance of autofluorescence and diffuse reflectance in oral cancer detection: a clinical study. J Biophotonics. 2011;4(10):696-706. 12. Chidananda SM, Satyamoorthy K, Rai L, Manjunath AP, Kartha VB. Optical diagnosis of cervical cancer by fluorescence spectroscopy technique. Int J Cancer. 2006;119(1):139-45. 13. Murali Krishna C, Sockalingum GD, Vidyasagar MS, Manfait M, Fernanades DJ, Vadhiraja BM, et al. An overview on applications of optical spectroscopy in cervical cancers. J Cancer Res Ther. 2008;4(1):26-36. 14. Alfano RR, Tang G, Pradhan A, Lam W, Choy D, Opher E. Fluorescence spectra from cancerous and normal human breast and lung tissues. Quantum Electronics, IEEE Journal of. 1987;23(10):1806-11. 15. Breslin TM, Xu F, Palmer GM, Zhu C, Gilchrist KW, Ramanujam N. Autofluorescence and diffuse reflectance properties of malignant and benign breast tissues. Ann Surg Oncol. 2004;11(1):65-70. 16. Lohmann W, Paul E. In situ detection of melanomas by fluorescence measurements. Naturwissenschaften. 1988;75(4):201-2. 17. Ebert B, Kohl M, Sukowski U, Rinneberg H, Winter H, Bellmann KP, et al. Fluorescence Imaging of Cutaneous Malignant Melanomas and Naevi. Lasers in Medical Science. 1998;13(3):204-8. 18. Lohmann W, Nilles M, Bodeker RH. In situ differentiation between nevi and malignant melanomas by fluorescence measurements. Naturwissenschaften. 1991;78(10):456-7.

16 19. Brancaleon L, Durkin AJ, Tu JH, Menaker G, Fallon JD, Kollias N. In vivo fluorescence spectroscopy of nonmelanoma skin cancer. Photochem Photobiol. 2001;73(2):178-83. 20. Ramanujam N, Mitchell MF, Mahadevan A, Thomsen S, Malpica A, Wright T, et al. Spectroscopic diagnosis of cervical intraepithelial neoplasia (CIN) in vivo using laser‐induced fluorescence spectra at multiple excitation wavlengths. Lasers in Surgery and Medicine. 1996;19(1):63-74. 21. Wu X, Kumar V, Ross Quinlan J, Ghosh J, Yang Q, Motoda H, et al. Top 10 algorithms in data mining. Knowledge and Information Systems. 2008;14(1):137. 22. Drakaki E, Kaselouris E, Makropoulou M, Serafetinides AA, Tsenga A, Stratigos AJ, et al. Laser-induced fluorescence and reflectance spectroscopy for the discrimination of basal cell carcinoma from the surrounding normal skin tissue. Skin Pharmacol Physiol. 2009;22(3):158-65. 23. Calin MA, Parasca SV, Savastru R, Calin MR, Dontu S. Optical techniques for the noninvasive diagnosis of skin cancer. J Cancer Res Clin Oncol. 2013;139(7):1083-104. 24. Venugopal C, Nazeer SS, Balan A, Jayasree RS. Autofluorescence spectroscopy augmented by multivariate analysis as a potential noninvasive tool for early diagnosis of oral cavity disorders. Photomed Laser Surg. 2013;31(12):605-12. 25. Na R, Stender IM, Wulf HC. Can autofluorescence demarcate basal cell carcinoma from normal skin? A comparison with protoporphyrin IX fluorescence. Acta Derm Venereol. 2001;81(4):246-9. 26. Stenquist B, Ericson MB, Strandeberg C, Molne L, Rosen A, Larko O, et al. Bispectral fluorescence imaging of aggressive basal cell carcinoma combined with histopathological mapping: a preliminary study indicating a possible adjunct to Mohs micrographic surgery. Br J Dermatol. 2006;154(2):305-9. 27. Panjehpour M, Julius CE, Phan MN, Vo-Dinh T, Overholt S. Laser-induced fluorescence spectroscopy for in vivo diagnosis of non-melanoma skin cancers. Lasers Surg Med. 2002;31(5):367-73. 28. Kamath SD, Bhat RA, Ray S, Mahato KK. Autofluorescence of normal, benign, and malignant ovarian tissues: a pilot study. Photomed Laser Surg. 2009;27(2):325-35.

17

1a

1b

Figure 1a) Spectra data collected from all BCC subtypes lesions (n=45 /212 spectra); 1b) Mean spectral curve of basal cell carcinoma (BCC) (red line), sun-exposed (EXP) normal skin (green line) and non-exposed (NEXP) normal skin (blue line) sites in overall sample.

1c 1a

1d

18

Figure 1c) Spectra data collected from all BCC subtypes lesions in phototype I patient (n=1); 1d) in phototype III patients (n=17) examples. BCC (red line); Sun-exposed (EXP) normal skin (green line); Non-exposed (NEXP) normal skin (blue line).

2a

2b

Figure 2a) Spectra data collected from PS lesions (n=37 /193 spectra); 2b) Mean spectral curve from PS lesions (black line), normal skin at sun-exposed (EXP) (green line) and non-exposed (NEXP) (blue line) sites.

19

3a

3b

Figure 3a) Ratio analysis: average values of the ratio between wavelengths 600nm and 500nm (FI600nm/ FI500 nm) in the basal cell carcinoma (BCC; n=45); 3b) and between wavelengths 600nm, 635 nm and 500nm (FI600nm/ FI500nm; FI635nm/ FI500nm) in the psoriasis (PS; n=37) lesions compared to respective normal skin at sun-exposed (EXP) and nonexposed (NEXP) sites. Error bars correspond to +/-1 SD

Table 1- k-NN classifier for discrimination among all categories of lesions, basal cell carcinoma, Bowen’s disease, psoriasis, and normal sun-exposed skin

BCC | BD| PS | EXP Mean Values categories k- Nearest Neighbors (n) Sensitivity Specificity 0.941 0.947 k-NN (3) 0.931 0.936 k-NN (7) 0.924 0.929 k-NN (11) 0.921 0.925 k-NN (15) BCC: basal cell carcinoma; BD: Bowen’s disease; PS: psoriasis; EXP: normal sun-exposed skin; n: numbers of neighbors range 3 up to 15

20 Table 2- k-NN classifier for discrimination among the categories of psoriasis, superficial BCC and nodular BCC lesions. PS | sBCC | nBCC categories k- Nearest Neighbors(n)

Mean values Sensitivity

Specificity

0.926 0.961 k-NN (3) 0.906 0.945 k-NN (7) 0.872 0.921 k-NN (11) 0.852 0.908 k-NN (15) PS: psoriasis; sBCC: superficial basal cell carcinoma; nBCC: nodular basal cell carcinoma; n: numbers of neighbors range 3 up to 15

Table 3- k-NN classifier for discrimination among the squamous-erythematous skin conditions (psoriasis, Bowen’s disease and superficial BCC) and nodular BCC.

PS | BD | sBCC |nBCC categories k- Nearest Neighbors(n)

Mean values Sensitivity

Specificity

0.925 0.965 k-NN (3) 0.92 0.957 k-NN (7) 0.877 0.93 k-NN (11) 0.851 0.914 k-NN (15) PS: psoriasis; BD: Bowen’s disease; sBCC: superficial basal cell carcinoma; nBCC: nodular basal cell carcinoma; n: numbers of neighbors range 3 up to 15

21 Table 4- k-NN classifier for discrimination among the three squamouserythematous skin conditions: psoriasis, Bowen’s disease and superficial BCC. PS| BD |superficial BCC Mean values categories k- Nearest Neighbors (n) Sensitivity Specificity 0.978 0.978 k-NN (3) 0.978 0.978 k-NN (7) 0.993 0.999 k-NN (11) 0.982 0.962 k-NN (15) PS: psoriasis; BD: Bowen’s disease; sBCC: superficial basal cell carcinoma; n: numbers of neighbors range 3 up to 15