2.1.

Experimental Setup and Spectra Acquisition

The laboratory setup is presented in Fig. 1; it combines principles of RS and AF for human skin tissue studies. The setup includes a thermally stabilized semiconductor NIR diode laser module LML-785.0RB-04 ($785\pm 0.1\mathrm{nm}$ central wavelength, 150 mW) for excitation of Raman spectra and AF in the NIR region and a diode-pumped solid-state laser module (457 nm, 200 mW) for AF stimulation in the visible region. Laser radiation is delivered to the optical detector by excitation fibers ($100\mu \mathrm{m}$ diameter, 0.22 NA) and collimating lenses ${\mathrm{L}}_1$ and ${\mathrm{L}}_2$. NIR (785 nm) laser radiation passes through the bandpass filter (BPF), which cuts off the Raman component of the optical fiber. The first (${\mathrm{DM}}_1$) and second (${\mathrm{DM}}_2$) dichroic mirrors transmit 785- and 457-nm laser radiation to lens ${\mathrm{L}}_3$, which focuses exciting radiation onto the sample (7.5-mm focal length). The same lens ${\mathrm{L}}_3$ collects RS, AF, and backscattered radiation. Dichroic mirror ${\mathrm{DM}}_3$ splits collected radiation on VIS and NIR channels, which include an appropriate longpass filter (${\mathrm{LPF}}_1/{\mathrm{LPF}}_2$) to cut off exciting (NIR/VIS) laser radiation, matching lens (${\mathrm{L}}_4/{\mathrm{L}}_5$), and collection fibers ($200\mu \mathrm{m}$ diameter, 0.22 NA), connected to multichannel spectrograph Shamrock SR-500i-D1-R with an Andor iDus CCD digital camera DU416A-LDC-DD cooled up to $-65°\mathrm{C}$. The experimental setup was calibrated consequentially for RS and both NIR and VIS AF signals. Low-noise recording of RS and AF radiation was performed in spectral bands 790 to 930 nm and 560 to 750 nm with spectral resolution better than 0.05 nm.

Fig. 1

Experimental setup: ${\mathrm{L}}_1$, ${\mathrm{L}}_2$, ${\mathrm{L}}_4$, and ${\mathrm{L}}_5$: matching lenses, ${\mathrm{L}}_3$: focusing lens, BPF: bandpass filter, ${\mathrm{M}}_1$ and ${\mathrm{M}}_2$: mirrors, ${\mathrm{DM}}_1$, ${\mathrm{DM}}_2$, and ${\mathrm{DM}}_3$: dichroic mirrors, and ${\mathrm{LPF}}_1$ and ${\mathrm{LPF}}_2$: longpass filters.

The optical detector was positioned directly over the tissue sample at a distance of 7 to 8 mm. The beam diameter of the probing radiation on the tissue was 1.5 mm. Measurement depth was about 1 to 2 mm. The laser power on the skin was 30 mW for both lasers, which corresponds to the laser density $1.7\mathrm{W}/{\mathrm{cm}}^2$ on the skin. The acquisition time was 15 s for VIS AF and 30 s for RS and NIR AF analysis.

All tissue samples were irradiated by NIR laser during 2 to 3 min before spectra registration in order to increase RS signal-to-noise ratio due to decreasing NIR AF intensity by the effect of tissues photobleaching while the intensity of Raman scattering remains unchanged.²⁷ It is important to note that the shape of the AF spectral function remains the same during 2 to 3 min of the photobleaching process and only NIR AF intensity monotonically decreases. This monotonic decrease of AF spectra during the photobleaching process was also observed by Wang et al.²⁸

NIR and VIS signals were registered consequently from the same tissue areas. Raman and AF spectra were acquired for a tumorous region and normal skin closely adjacent to a lesion (within 2 to 4 cm). Three independent spectra measurements were made for each predefined by medical personnel spatial point, and then an averaged value was calculated for mean spectrum. It helps to calculate the changes in malignant tissue in comparison with normal tissue and compose relative criteria (see Sec. 2.3), which exclude the influence of individual variability of skin properties on obtained results.

2.2.

Tissue Samples

There were 79 patients with skin cancers (38 female and 41 male, all white, Caucasian, skin phenotype I and II) enrolled in this study. Ex vivo tissue samples were obtained after surgical resection at Samara Regional Clinical Oncology Dispensary under an approved protocol including patient agreement. The first cohort of 158 samples (79 healthy skins, 39 MM, and 40 BCC) underwent RS and NIR AF analysis. The second cohort of 74 samples (37 normal skins, 27 BCC, and 10 MM) underwent RS and AF spectra registration in visible and NIR regions. Tissue samples were stored in sterile boxes at $+4\pm 2°\mathrm{C}$ and were tested using an experimental setup no later than 4 h after resection. The skin tumor samples were $\sim 2\times 2\times 1{\mathrm{cm}}^3$ in size. All samples were divided into two pieces containing both a healthy tissue region and part of the tumor. One sample piece underwent experimental tests. The rest of the sample was fixed in formalin and prepared for histological analysis. The detailed distribution of human skin lesions, including information about tumor locations, is provided in Table 1.

Table 1

Summary of skin tissue samples.

Every tumor study was accompanied by histological analysis to make a final diagnosis. The protocols of ex vivo tissue diagnostics were approved by the ethical committee of Samara State Medical University.

2.3.

Spectra Processing and Data Analysis

NIR laser radiation simultaneously stimulates Raman scattering and AF response in the NIR spectral range. The collected spectra contain a wide exponentially decaying AF spectra curve with sharp Raman peaks, which are usually separated by the method of AF polynomial approximation.²⁹ In the current study, this method was upgraded with additional filtering of random noises and automatic selection of polynomial order.

All processed RS, NIR AF, and VIS AF spectra may be grouped according to histological data and further statistically processed using principal component analysis (PCA) or other multivariate analysis techniques.³⁰ It is well known that the number of tested samples should be greater or of the same order as a number of variables for mathematically correct covariance decomposition. In addition, the entire RS/AF spectrum analysis may provide incomplete understanding of chemical changes responsible for disease diagnosis.³¹ In this study, the analysis of obtained spectra was performed based on a number of a priori predefined spectral criteria described below. These criteria were associated with relative content of the main skin chromophores. Such an approach allowed for reduction in the number of variables from thousands (for the entire spectra) to a few variables and helps to include personal skin features in the analysis. It also helps us to register spectral information for understanding the differences between normal and cancerous tissues RS/AF spectra.

The RS analysis was performed with the help of a two-step phase-space analysis method,²⁶ which is based on alteration of ratio between RS spectral peaks $I_k$ in the tumor area and the healthy tissue near the lesion. Normalized intensity $I_k$ was defined as the ratio of maximum of RS intensity of $k$’th spectral band to the maximum RS intensity in the band 1440 to $1460{\mathrm{cm}}^{-1}$ as the most intensive RS band of typical Raman spectra for both cancer and normal skin (Fig. 2). Such an approach makes it possible to separate malignant and normal tissue in the first step. The differentiation of cancer types among selected cancerous samples was performed in the second step with the help of relative coefficients

Fig. 2

Normalized ex vivo Raman spectra of normal skin tissue, MM, and BCC.

Coefficients [Eq. (1)] for the two most intensive spectral ranges $k=1320$ and $1660{\mathrm{cm}}^{-1}$ form the phase plane which was used for tissue classification by discriminant analysis (DA). DA can separate two or more classes based on different statistical parameters of Gaussian distributions. The efficiency of the proposed approach is characterized by sensitivity and specificity and ability to select defined classes in different areas of phase plane. The analysis of skin tissues data allocation was performed using linear and quadratic DA classifiers.³²

Within the spectral range 870 to 920 nm, NIR AF intensity decreases with the growth of the wavelength (Fig. 3), and the most significant qualitative changes of the NIR AF spectrum for different tumors were observed within the range of 810 to 870 nm. For the comparative analysis of the experimental data, we approximate the NIR AF spectrum with an exponential function

Fig. 3

Typical normalized ex vivo AF NIR spectra of normal skin, MM, and BCC.

According to observed VIS AF spectra (Fig. 4), we propose to use a number of criteria to monitor the changes of flavins, lipids, and porphyrins content in skin tissues. The most useful among them are the ratios between intensities of VIS AF spectra maxima in 570 to 590 nm ($I_{570}$) and 610 to 690 nm ($I_{610}$) ranges ${FI}_{\mathrm{VIS}}=\frac{I_{610}}{I_{570}}$ and normalized spectral maxima position shift $F{\lambda }_{\mathrm{VIS}}=\frac{|{\lambda }_{\text{norm}}-{\lambda }_{\text{tumor}}|}{{\lambda }_{\text{norm}}+{\lambda }_{\text{tumor}}}$, where ${\lambda }_{\text{tumor}}$ and ${\lambda }_{\text{norm}}$ are the position of AF maximum for tumor and normal skin near lesion for the same sample.³⁵ The value of ${FI}_{\mathrm{VIS}}$ is proportional to the relative concentration of total flavins and lipopigments content. VIS AF spectrum shift $F{\lambda }_{\mathrm{VIS}}$ helps to estimate changes in different types of porphyrins appearing on the upper skin layers, as tumors contain more bacteria on their surface than a normal skin which results in different porphyrins composition.

Fig. 4

Normalized VIS AF spectra of normal skin tissue, MM, and BCC stimulated by 457 nm laser.

Such an approach makes it possible to define multidimensional space of different RS and AF criterial parameters. Each specific cancer type characterized by its own tissue components changes the equivalent to a specific set of criteria values. Thus, each cancer type fills in the exact area in multidimensional phase space, and, as a result, we have a mechanism for exact optical identification of tumor type. A correlation between the proposed criteria was estimated in order to find the groups of criteria with the lowest correlation for multidimensional space analysis. Pearson’s correlation coefficients were calculated for each pair of proposed criteria.³⁶ Further analysis of multidimensional phase spaces based on chosen criteria was performed with PCA. This type of analysis helps to reduce the dimensionality of the data while retaining most of the variation in the data set. It uses directions, called principal components, along which the variation in the data is maximal. Isolating a number of components, each sample can be represented by relatively few numbers instead of multiple variables. Samples can then be plotted, making it possible to visually assess similarities and differences between samples.³⁷

3.1.

Raman Spectroscopy Tissue Study

Absolute values of RS intensity may differ significantly from one skin tissue sample to another due to high variability of tissue components concentration. Each spectrum was normalized to maximum intensity in the whole spectral range of 1200 to $1800{\mathrm{cm}}^{-1}$ in order to overcome this effect. Typical registered Raman spectra of MM, BCC, and healthy skin are presented in Fig. 2. The most intensive RS band of typical normalized Raman spectra²⁶ of cancer and normal skin is located at $1450{\mathrm{cm}}^{-1}$ and associated with ${\mathrm{CH}}_2$ deformations of proteins and lipids. Other well observed bands are 1240 to $1280{\mathrm{cm}}^{-1}$ (stretching mode $\mathrm{C}═\mathrm{N}$), 1300 to $1340{\mathrm{cm}}^{-1}$ (twisting, wagging of bending mode ${\mathrm{CH}}_2$), 1540 to $1580{\mathrm{cm}}^{-1}$ (deformation mode $\mathrm{C}═\mathrm{C}$ and tryptophan), and 1640 to $1680{\mathrm{cm}}^{-1}$ (stretching mode $\mathrm{C}═\mathrm{O}$ amide I).

The two-step RS analysis of skin samples is shown in Fig. 5. Here, every tested sample is a single point on the phase plane with coordinates corresponding to RS spectral coefficients of the two-step algorithm described in Sec. 2.3. Figure 5(a) represents the first step of analysis where we separate normal skin tissues from the malignant pathologies. Such classification is possible with a total accuracy 54.7% using linear DA. Accuracy of MM separation versus all other skin types reaches 72.7%. One may see that the areas of normal skin and BCC are mainly overlapping in Fig. 5(a) and thus the separation of malignant and normal skin tissue is complicated. Therefore, coefficients $I_{1320}$ and $I_{1660}$ were considered low informative for skin tissue analysis.

Fig. 5

Skin tissues classification by two-step RS algorithm: (a) the first step—MM versus all other skin tissues and (b) the second step—MM versus BCC.

On the other hand, the problem of normal skin separation from malignant tissues is not so relevant in clinical practice as a problem of exact determination of malignant tissue type. It is more appropriate to separate MM from BCC. For this purpose, coefficients $I_{1320}$ and $I_{1660}$ were supplemented with coefficients ${\mathrm{RS}}_{1320}$ and ${\mathrm{RS}}_{1660}$, which were used in the second step [Fig. 5(b)]. The overall accuracy of BCC and MM separation rose up to 80.3% for the two-step method.

The achieved accuracy is 8% lower than the accuracy shown in our previous study.²⁶ The decrease in accuracy is caused by larger dispersion of tissue components content due to threefold increase in the number of tested samples.

3.2.

Near-Infrared Autofluorescence Tissue Study

As RS two-step analysis reveals the relative alteration of content of proteins and lipids in the tumor, the diagnostics accuracy may be improved by additional analysis of other tissue components. The melanin content proves to be a helpful additional criterion as MM are usually more pigmented lesions than BCC. Moreover, the content of pheomelanin and eumelanin may differ in malignant skin tumors.³⁸ Thus, measuring the melanin content may improve MM and BCC classification with optical methods.

Figure 3 demonstrates typical NIR AF spectra of normal skin and tumor stimulated by 785 nm laser. The dotted curve shows the exponential approximation [Eq. (2)] of NIR AF spectra. As the curvature (absolute value and its sign) and slope of approximation function depend on melanin content in the skin tissue, coefficients ${FI}_{\mathrm{NIR}}$ and $F{\lambda }_{\mathrm{NIR}}$ may be used for estimation of melanin value. And these coefficients may be measured simultaneously with Raman signal. Such an approach was first described in our study,³⁴ where only the curvature sign of the spectrum approximation function [Eq. (2)] was used for sample differentiation. In the current study, we additionally use absolute values of the curvature (criterion ${FI}_{\mathrm{NIR}}$) for MM and BCC separation. In general, analysis of coefficient ${FI}_{\mathrm{NIR}}$ showed a possibility to separate MM and BCC with a total accuracy of 60.8%.

The box-and-whisker plot (Fig. 6) demonstrates the values of coefficient $F{\lambda }_{\mathrm{NIR}}$ for MM, BCC, and normal skin classification. Note that median, first quartile, and minimum of $F{\lambda }_{\mathrm{NIR}}$ for BCC and normal skin have almost identical values. In the current study, it was assumed that the criterial norm $F{\lambda }_{\mathrm{NIR}}\ge 1$ for the maximum efficiency of MM selection among BCC and normal skin. The sensitivity and specificity for MM selection reached 69.2% and 85.0%, respectively, for this criterial norm.

Fig. 6

Classification of MM, BCC, and normal skin based on melanin content (curvature $F{\lambda }_{\mathrm{NIR}}$ of registered AF spectra stimulated by 785 nm laser).

Further increase in accuracy of malignant tumor detection with NIR AF may be achieved by combined analysis of both criteria (${FI}_{\mathrm{NIR}}$ and $F{\lambda }_{\mathrm{NIR}}$): the result was considered positive when one of the criteria showed that the neoplasm is MM. In this case, the sensitivity of MM detection reaches 92.3%, but the overall accuracy of MM and BCC separation was only 64.6%.

3.3.

VIS Autofluorescence Tissue Study

Unlike RS and NIR AF, the analysis of AF spectra stimulated in the visible region helps to reveal properties of lipids, flavins, and porphyrins in skin tissue. Skin lipids contribute to normal barrier function. Some lipids found on the skin’s surface make the skin unfriendly to fungi and bacteria.³⁹ Porphyrins play a significant role in metabolism processes, including metabolism of bacteria on the skin surface.^40,41 Flavins control photostimulated generation of melanin within the specialized cells.⁴² Thus, the synthesis of melanin in normal skin and tumors is related to the flavins present. Also, tumors are characterized by increased rate of metabolism processes and growing of MM or BCC causes the alteration of flavin and porphyrin concentrations, which are expressed in tumor color alteration in comparison with normal skin.

Typical VIS AF spectra have a local maximum in 570 to 590 nm for all skin types and maximum in spectral range of 610 to 690 nm (Fig. 4), which is a combination of two local maxima in 610 to 640 nm and 650 to 690 nm. Here, normal skin and MM have only one maximum in 629 and 623 nm, respectively, while BCC has two strong maxima of comparable intensity in 635 and 678 nm. Shown examples of spectra are common for studied skin tissues, but exact maxima positions and intensities may slightly differ from sample to sample. Redistribution of AF intensity between local maxima 610 to 640 nm and 650 to 690 nm is caused by changes in concentration of flavins and porphyrins and leads to the main maximum position shift. These changes may be defined by the spectral criteria pair ${FI}_{\mathrm{VIS}}$ and $F{\lambda }_{\mathrm{VIS}}$.

The possibility of MM and BCC separation by each VIS AF criteria is shown in Fig. 7. The sensitivity of MM detection equals 100% achieved for intensity threshold ${FI}_{\mathrm{VIS}}>0.91$ or for spectral shift $F{\lambda }_{\mathrm{VIS}}$ between 1 and 19 nm. But the accuracy of such detection does not exceed 51%. One may see insufficient accuracy of MM detection based on single VIS AF criterion analysis. The increase in accuracy up to 75% may be achieved by simultaneous monitoring of both criteria as it is shown in Fig. 8. Here, classification of MM and BCC was performed with linear DA on phase plane of porphyrins composition on tissue surface ($F{\lambda }_{\mathrm{VIS}}$) and relative content of porphyrins, lipids, and flavins (${FI}_{\mathrm{VIS}}$).

Fig. 7

Skin tissues classification based on VIS AF criteria: (a) porphyrins composition on tissue surface ($F{\lambda }_{\mathrm{VIS}}$) and (b) relative content of porphyrins, lipids, and flavins (${FI}_{\mathrm{VIS}}$).

Fig. 8

MM and BCC classification on the phase plane with joint VIS AF monitoring of porphyrins composition ($F{\lambda }_{\mathrm{VIS}}$) and relative content of porphyrins, lipids, and flavins (${FI}_{\mathrm{VIS}}$).

3.4.

Combined Raman Spectroscopy–Autofluorescence Study

As each proposed criterion refers to changes of different tumor components the enhancement of the classification accuracy may be achieved by combination of criteria from different spectroscopic techniques. For instance, Fig. 9 contains the phase plane for analysis of criteria pair for relative porphyrins, lipids, and flavins content monitoring and melanin level estimation (${FI}_{\mathrm{VIS}}$ and $F{\lambda }_{\mathrm{NIR}}$). The combination of AF criteria in VIS and NIR regions provides 8% to 10% more accurate separation of MM and BCC in comparison with single NIR or VIS AF analysis.

Fig. 9

MM and BCC classification on the phase plane with criteria ${FI}_{\mathrm{VIS}}$ and $F{\lambda }_{\mathrm{NIR}}$.

Instead of using separate criteria pairs, it is more efficient to simultaneously analyze as many criteria as possible, i.e., using the information about all chosen tissue components in a multidimensional criteria space. Such classification was performed in 4-D space (${FI}_{\mathrm{NIR}}$ and $F{\lambda }_{\mathrm{NIR}}$ for NIR AF, and ${\mathrm{RS}}_{1320}$ and ${\mathrm{RS}}_{1660}$ for RS analysis) for the first cohort of 158 studied samples and 6-D space (additionally measured ${FI}_{\mathrm{VIS}}$ and $F{\lambda }_{\mathrm{VIS}}$ for VIS AF) for the second cohort of 74 samples.

Figures 10 and 11 demonstrate 3-D representation of 4-D and 6-D space analysis, respectively. Complex optical methods show high efficiency of MM and BCC separation: the sensitivity and specificity of cancerous tissues classification are 94.9% and 92.5%, respectively, in 4-D criteria space, and 100% and 96.3% for 6-D criteria space.

Fig. 10

MM and BCC classification with PCA in 4-D space (${FI}_{785}$, $F{\lambda }_{785}$, $R{S}_{1320}$, and $R{S}_{1660}$): (a) PC1, PC2, and PC3 and (b) PC2, PC3, and PC4.

Fig. 11

MM and BCC classification with PCA in 6-D space (${FI}_{785}$, $F{\lambda }_{785}$, $R{S}_{1320}$, $R{S}_{1660}$, ${FI}_{457}$, and $F{\lambda }_{457}$): (a) PC1, PC2 and PC3 and (b) PC2, PC3 and PC4.

The accuracy of MM and BCC separation is slightly higher for the six criteria algorithm. This fact may be caused both by increasing the data analyzed (additional VIS AF criteria) and by decreasing the number of the samples studied. A correlation between the chosen criteria was calculated in order to ensure the comparative increase of classification accuracy and its independence on the number of the tested tissue samples. Table 2 shows correlation matrixes for 6-D criteria space. Here, the bold font indicates the significant correlations ($p\text{value}<0.01$).

Table 2

Correlation of criteria RS1320, RS1660, FINIR, FλNIR, FIVIS, and FλVIS.