Differentiation of basal cell carcinoma and healthy skin using multispectral modulation autofluorescence imaging: A pilot study

An approach for differentiating basal cell carcinoma (BCC) and healthy skin by combining a multispectral modulation autofluorescence imaging with the linear discriminant analysis has been proposed. The experimental setup, which employs a 365-nm narrowband excitation, 4 replaceable bandpass filters and a digital camera, has been assembled and applied to study freshly excised samples of BCC. In the experimental setup, modulation of the UV-excitation and demodulation of the visible light images allow for both increasing a signal-to-noise ratio and suppressing a nonfluorescence background in the autofluorescence images of tissues. The observed results demonstrate an ability for distinguishing both ordinary and keratinized BCC from healthy skin justifying the perspectives of the multispectral modulation autofluorescence imaging use for non-invasive and intraoperative diagnosis of BCC and other low-pigmented malignancies of the skin. © 2019 Journal of Biomedical Photonics & Engineering.

One of the most promising applications of fluorescence spectroscopy and imaging is non-invasive and least-invasive diagnosis of skin cancers based on excitation either endogenous or exogenous fluorophores [1,4,38]. The methods of exogenous fluorescence diagnosis [39] employ various fluorescent dyes, which are injected into the skin, accumulated in cancer cells, fluoresce under a proper illumination, and, in some cases, could provide the photodynamic effect during the subsequent therapy [40][41][42][43][44][45][46]. This technique yields high sensitivity; however, it is also characterized by numerous disadvantages significantly limiting its reliability: (i) the procedure of dye delivery is rather time-consuming; (ii) it requires invasive delivery protocols (injection) or minimally-invasive topical or per-oral administration, and its further elimination from patient body; (iii) exogenous fluorescent dyes cannot provide satisfactory differentiation between similar pathologies (i.e. the specificity of the method is not high [4]).
The autofluorescence (AF) spectroscopy and imaging relying on endogenous fluorophores represent non-invasive and fast approach for skin cancers diagnosis. However, due to high pigmentation of pigmentary skin nevi [47] and early malignant melanomas of the skin [48,49], this method has a significant limitation because of low light penetration depth [4]. The non-melanoma cancers of the skin [50], such as basal cell carcinoma (BCC) and squamous cell carcinoma (SCC), are usually less pigmented and, as a consequence, appear to be promising goal for the AF diagnosis [4,[51][52][53][54][55][56][57][58]. Despite the significant progress in this field of fluorescence biomedical technologies, the development of novel techniques for skin diagnosis based on AF remains a challenging problem.
In this paper, we propose an approach for differentiating BCC and healthy skin using the multispectral modulation AF imaging combined with linear discriminant analysis (LDA). We assemble an experimental setup employing a 365-nm narrowband excitation of the tissue sample AF, 4 replaceable bandpass filters and a digital camera. It allows for both increasing a signal-to-noise ratio and suppressing a nonfluorescent background in the AF images of tissues using modulation of the UV-excitation and demodulation of the AF visible light images. We use the experimental setup to study in vitro samples of freshly excised BCC from 6 patients. The results of this pilot study demonstrate ability for distinguishing ordinary and keratinized BCC from healthy skin using the proposed imaging approach. We analyze the accuracy of the tissue differentiation using this multispectral modulation AF imaging. We qualitatively compare the proposed technique with the existing developments in the area of multispectral AF imaging of BCC, which allows us to highlight the potentials of its use for noninvasive and intraoperative diagnosis of BCC and other low-pigmented skin malignancies.

Experimental setup
To study the images of non-melanoma skin cancer AF, we assembled a setup for the modulated multi-spectral fluorescence imaging [59,60]. It includes excitation and detection channels (see Fig. 1).
In the excitation channel, a polychromatic light of a mercury lamp radiates the sample of interest after passing a 365-nm narrowband excitation filter, a rectangular beam homogenizer, and a ! !"# = 5 Hz mechanical chopper. The homogenizer shapes the UVbeam to rectangular form leading to the spatiallyhomogeneous irradiance of the sample, which is placed behind a non-fluorescent reference window (see Fig. 1(c)). Despite the polychromatic mercury lump with a total power of 100 W is employed in our study, a monochromatic UV-irradiance of the sample is only about 10 mW/cm 2 fulfilling the guidelines limiting an exposure of the skin in vivo to the electromagnetic waves [61], as specifically, the permissible UVirradiance of the skin is of about 100 mW/cm 2 for a typical measurement time of 10 sec.
The UV-radiation being interacted with tissue via scattering and emittance of fluorescence [62] is comprised of both the sharp line of the residual UVexcitation light and the broadband fluorescence varying for different regions of the sample under study. In the detection channel, a set of the bandpass interference filters and a digital camera are used for detection of the quasi-monochromatic visible light images ! ! (!, !) of the object fluorescence, where index ! defines a certain bandpass filter, ! is the radius vector at the image plane, and ! is the time. Spectral transparency curves !(!) of the bandpass filters with different central wavelengths ! ≃ 400, 450, 500, and 550 nm, respectively, and similar full-width at half-maximum (FWHM) of ∆! ≃ 40 nm for all filters are shown in Fig. 1(b). We select the described set of the bandpass filters from commercially available ones to both almost continuously cover the most intense part of healthy skin and BCC AF spectra [4] and divide AF spectra into four initially-orthogonal ranges -i.e. separate channels of the multispectral imaging. We use a digital camera with complementary metal-oxide-semiconductor (CMOS) sensor and the frame rate of ! !"# =25 Hz.  In the experimental setup (see Fig. 1), the average angle of the UV radiation incidence on the sample surface (i.e. the angle between the optical axis of the UV-excitation channel and the perpendicular to the sample surface) and the angle of the image detection (i.e. the angle between the optical axis of the imaging channel and the perpendicular to the sample surface) are ! ≃ 17° and !′ ≃ 0°, respectively. Numerical apertures of the UV-excitation and the fluorescence imaging system are NA = 0.09 ( 2! ≃ 10°) and NA′ = 0.16 (2!′ ≃ 19°), respectively. Along with simplification of the experimental setup and minimization of its dimensions, the use of conventional low-aperture optics in excitation and detection channels allows for providing both a homogeneous irradiance of the sample surface and a spatial resolution of about 0.1 mm at the object plane for the entire field of view.
We perform fluorescence imaging in non-polarized light and assume isotropic character of the sample fluorescence indicatrix. This assumption is correct in case of the small apertures along with the close-tonormal angles of incidence and imaging [63]. At the same time, numerous researches, who are studying the fluorescence of turbid media (including, the fluorescence of tissues), have shown complex polarization-dependent and anisotropic character of this phenomenon [64][65][66][67][68][69][70]. By combining optimized geometry of fluorescence excitation and collection with modern fluorescence data processing techniques, one could mitigate an impact of the light scattering in a tissue, reconstruct its intrinsic fluorescence [71][72][73][74][75] and improve an efficiency of fluorescence collection [76][77][78]. Furthermore, a polarization-dependent character and an anisotropy of tissue fluorescence can serve as a source of additional information for differentiating healthy and pathological tissues [79][80][81].
During the experimental study, the in vitro tissue sample is rigidly fixed behind the non-fluorescent reference window (see Fig. 1(a)). By rigidly fixing the sample surface towards the reference window, we eliminate the influence of various negative factors on the images: (i) an accuracy of the sample positioning; (ii) fluctuations of an incident angle of the UVradiation; and (iii) non-normal geometry of imaging (i.e. non-orthogonality of the imaging optical axis towards the sample surface). In our work, we consider lowpigmented malignancies of the skin -pigmentation of BCC is insignificant compared to highly-pigmented nevi and melanomas of the skin [4]. Therefore, we neglect the influence of a laterally-inhomogeneous tissue pigmentation on the data of the fluorescence imaging [82].

Visible light image processing
Using Fourier-domain analysis, we reconstruct the fluorescence images ! ! (x) corresponding to various bandpass filters λ [83]: This technique employs modulation of the UVradiation intensity and demodulation of the image sequences in the visible range. It yields detecting a set of the sample AF images ! ! (!) with both an increased signal-to-noise ratio and a reduced impact of the nonfluorescent background [59,60]. In order to provide accurate demodulation of the data, a frame rate of the imaging system ! !"# should be much lower compared to the frame rate of the camera ! !"# and the modulation frequency ! !"# . In the described experimental setup (see Fig. 1), the frame rate of the multispectral modulation AF imaging is about ! !"# = 0.5 Hz; it is primarily limited with the inertness of the CMOS sensor and could be further increased (even to real-time operation) by substituting the CMOS camera with the advanced one characterized with improved performance.
In order to study an ability for the differentiation of healthy and pathological tissues using the proposed multispectral AF imaging technique, we examined a set of freshly excised BCC samples provided by the Sechenov University (Moscow, Russia). The tissue samples came from 1 male and 5 female patients aged between 57 and 78 were investigated. The abnormal tissue specimens were investigated within 3 hours after the surgical removal. Classification of the normal skin and BCC tissues was done on the basis of initial clinical diagnosis and proved by a further histological examination of the biopsies. Table 1 summarizes the data on the BCC samples, including the final diagnosis, histological image, patient gender and age, side and size of excision.  Figure 2 shows results of the multispectral modulation AF imaging of the 1 st BCC sample from Table 1. Panel (a) shows digital image of the sample, where the borders mark the healthy skin (I), BCC (II), flake of keratinized BCC (III) and imaged area (IV). Panel (b) visualizes reverse contrast frames from AF image ! ! (!, !) at ! = 500 nm. Panel (c) shows a set of reverse contrast AF images ! ! (!) of BCC extracted from the modulated AF sequences (see Eq. (1)). In complete accordance with the Refs. [53,84], the experimental results show lower AF intensity for the ordinary BCC tissues. At the same time, the highest fluorescence intensity corresponds to the flake of keratinized BCC (region III in Fig. 2(a)).
The theoretical background for distinctive AF changes in cancer tissue was investigated earlier [4,24,85,86]. The altered metabolic state of cancer cells and resulting abnormal concentration of nicotinamide adenine dinucleotide (NADH) were suggested to be the origin of cancer AF changes [86,87], which allows using the AF phenomenon for discriminating malignancies from normal tissues [87][88][89][90][91]. The changes of the AF spectrum in malignant skin tissues (under the UV-excitation in the spectral range of 260 to 400 nm) were reported to originate from the decreased collagen and elastin content in tissues, as well as from the decreased NADH levels [84,[92][93][94][95][96][97][98]. Furthermore, the differences in optical properties of healthy and pathological tissues (i.e. absorption and scattering of the UV-radiation and the fluorescence component) provide the contrast between the healthy skin and BCC in the AF spectra and images, which is originated owing to the listed features of malignant tissues.
A set of the four fluorescence images ! ! (!) (see Fig. 2(c)) represents information about the sample AF intensity in different spectral channels. Such data could not be directly used for differentiation of tissues, for example, using simple intensity threshold, since AF intensity depends on a number of factors (the shape of the skin surface, its pigmentation and hydration, and others). For instance, according to the data of Fig. 2(c), the ordinary BCC tissue fluoresces less intensively comparing to the healthy skin; but, oppositely, the keratinized BCC fluoresces much intensively. Therefore, novel approach for tissue segmentation, which does not rely on simple AF intensity analysis should be introduced.

LDA-based differentiation of tissues
For differentiation of healthy and abnormal tissues, we perform statistical analysis of normalized fluorescence intensities in all of the four spectral channels using the LDA approach [93,99] -i.e. a linear transformation technique, which yields reduction of the experimental data dimensionality and maximization of the separation between classes.  Table 1), where characters I to IV define the healthy skin, BCC, flake of keratinized BCC and imaging areas, respectively; (b) reverse contrast frames of the multispectral AF image ! ! (!, !) at ! =500 nm; (c) reverse contrast demodulated multispectral AF images ! ! (!) at ! =400, 450, 500, 550 nm.   Table 1); (b) LDA space for the 1 st BCC sample (see Fig. 2 and Table 1). Classes of the healthy skin, BCC and keratinized BCC are represented with blue, red and green areas, respectively. Light blue 4D plane in (a) and line in (b) show the threshold for discriminating healthy skin and BCC.
We introduce the principal components based on normalized intensities of the tissue AF. The intensities of each ! ! -pixel in the AF images ! ! (!) define a radiusvector ! ! in a four-dimensional (4D) principal component space where is a normalized intensity corresponding to the ! !! -pixel fluorescence in a certain spectral channel ! =400, 450, 500 or 550 nm, and ! = 4 is a number of the spectral channels. These principal components are basically orthogonal due to initial orthogonality of the spectral transparency curves !(!) of the bandpass filters (see Fig. 1(b)). The full size of the AF image ! ! is !×! points; thus, it allows for calculating !×! radius vectors in the principal component space. Fig. 3(a) demonstrates an example of the principal component space, which is calculated based on the data of the 1 st BCC sample imaging (see Fig. 2 and Table 1). Namely, it shows three-dimensional (3D) projections of the four-dimensional principal component space, where various points are colored according to areas of the healthy skin (blue), BCC (red) and keratinized BCC (green). This data set represents a typical character of the principal component plot, which is observed for all BCC samples. The sets of points in the principal component space by BCC and healthy skin tissues form classes, which are principally distinguishable.
We have implemented the Fisher LDA approach [99] to highlight an ability for the differentiation of healthy skin and BCC based on the proposed AF imaging technique. In order to find the procedure of the data projection into the 1D Fisher LDA subspace and to estimate the threshold value, we perform the following data analysis.
• We perform multispectral modulation AF imaging of all six BCC samples (see Table 1).
• We manually choose a number of healthy skin and BCC points in the AF images ! ! (!) in order to form healthy skin and BCC classes in the principal component space for further threshold selection. The amounts of healthy skin and BCC points, which are assigned a priori basing on the data of histological analysis and expert visual examination, varies from sample to sample. However, this amounts account at least 30 points for each tissue sample and each class.
• Using all the manually selected points, we form two groups of radius vectors ! !,!"#$ , ! !,!"" representing healthy skin and BCC classes in the principal space. For these two classes, we calculate mean vectors ! !"#$ , ! !"" and scatterers ! !"#$ , ! !"" : where ! !"#$ and ! !"" are numbers of priory selected points of healthy and abnormal tissues, respectively. Since LDA employs linear data projection into 1D subspace ! ! = ! ! ! ! (here, ! ! is a projection operator), in 1D space we define the points of healthy skin and BCC classes ! !,!"#$ , ! !,!"" , the related mean values ! !"#$ , ! !"" and the scatterers ! !"#$ , ! !"" : The LDA should maximize separation of healthy skin and BCC classes in 1D subspace, and this is equal to maximizing the following criterion where are between-class and within-class scatterers, respectively. By maximizing the criterion (4), we obtain the following solution for the projection operator Thereby, by solving Eq. (5) with the priorycharacterize healthy skin and BCC points, we find the projection of the data to 1D Fisher LDA subspace.
• Knowing the mean values and the scatterers for the healthy and abnormal tissues, we set a threshold for the discrimination of these two classes in the LDA space. This threshold divides the space between the mean values ! !"#$ , ! !"" into two parts, in direct ratio to the values of scatterers ! !"#$ , ! !"" .  Table 1). The observed distributions of the data sets in the LDA space could be well-described by normal (Gaussian) probability density functions. Light blue line in Fig. 3(b) illustrates the value of the LDA threshold, which was determined using the described procedure based on the measurements of all six samples. In Fig. 3(a), equal linear threshold is shown as a light blue 4D plane in the principal component space. This threshold allows for differentiating both ordinary and keratinized BCC from healthy tissues of the skin.
After determining the projection operator ! ! and the threshold value, we could apply the following procedure to discriminate healthy and BCC tissue.
• Calculation of the experimental data representation in the principal component space (see Eqs. (2)-(3)) for the entire number of the image pixels.
• Projection of the data from the principal component space to the LDA space using the operator ! ! (see Eq. (5)).
• Differentiation of healthy skin and BCC tissues in the LDA space using the prior defined threshold value.
• Back-projection of the discriminated LDA data onto the initial visual image (clinical photo) of the sample.
In the LDA space (see Fig. 3(b)), both ordinary and keratinized BCC data appear to be shifted in one direction towards healthy skin data. The separation of tissues becomes possible because the LDA of the multispectral AF images allows studying spectrallydependent data on the tissue AF instead of analyzing only the AF intensity.

LDA-Based Mapping of the Skin
In order to highlight an ability for discrimination of BCC from healthy skin tissue using the proposed AF imaging approach, we implement equal projection operator ! ! and LDA threshold value for differentiating healthy and cancer tissues of all six BCC samples from Table 1. Figure 4 shows results of the tissue mapping: semi-transparent green masks represent the BCC area, which is estimated via the proposed multispectral AF imaging; red lines show the actual area of lesions. Since all of the imaged BCC samples correspond to the clinically identified type of BCC lesions (i.e. these samples do not contain any subsurface or clinically undetected BCC fragments), in our research, we estimate the actual area of the lesions by combining the visual examination of tissues by medical expert with the further histological study of tissue fragments.
Considering the results of the 1 st BCC sample mapping (see Fig. 4(a) and Table 1), we could observe high efficiency of the proposed technique. For this sample, we correctly identified 96.8% of BCC area and 85.7% of healthy skin area, while the mistakes of BCC and healthy skin identification were 3.2% and 14.3%, respectively. For other BCC samples (panels from (b) to (f) in Fig. 4 corresponding to the BCC samples 2 to 6 in Table 1), we observed lower accuracy of tissue discrimination; Table 2 summarizes an analysis of the proposed technique accuracy, as well as reports averaged values -76.4 % and 82.5 % of BCC and healthy skin tissues were accurately characterized. The minimal amount of imaged BCC samples is enough to highlight the ability for tissue differentiation using the proposed multispectral modulation AF imaging technique. However, this amount is not enough to perform in-depth statistical analysis of experimental data -i.e. to analyze sensitivity and specificity, type I and type II errors of tissue differentiation.
The accuracy of the tissue mapping could suffer from a number of factors, including laterallyinhomogeneous absorption [4,82] and anisotropic polarization-dependent scattering [64,73] of both the UV-radiation and the AF component in tissues. For further increase of imaging efficiency and tissue discrimination accuracy, the listed features of tissue should be taken into account during the analysis of the multispectral AF imaging data [74,75,100].
Furthermore, by increasing the number of tissue samples at the stage of algorithm calibration, we could also significantly improve the proposed technique performance.

Discussions
The results of this pilot study involving in vitro tissue examination allow us to perform preliminary analysis, i.e. to perform feasibility test, which aims to objectively uncover strengths and weaknesses of the proposed AF imaging approach before committing to a full-blown study. Since physical properties of in vitro and in vivo tissues, including their AF spectra, could differ significantly depending on the time passed between surgery excision and experimental study, obviously, our future work should be dedicated to extensive clinical research focused on in vivo measurements. Nothing but the clinical research of in vivo samples, which involves the stage of method calibration based on large amount of samples and the stage of extensive statistical analysis of method performance, could form a basis for development of novel techniques for skin cancer diagnosis based on the multispectral modulation AF imaging. Fig. 4 Mapping healthy and BCC tissue using the multi-spectral AF imaging: semi-transparent green masks represent the BCC area estimated via the proposed approach; red lines show the actual BCC area estimated via the visual examination of tissues by medical expert combined with the further histological study of tissues. In our future research, we would also focus on the development of a portable experimental setup with improved operation rate, up to real-time imaging. For this purpose, we would: (i) increase the sensitivity of digital camera; (ii) optimize the UV-radiation parameters -i.e. select the optimal wavelength and increase the excitation power to the maximal permissible value; (iii) optimize a set of the bandpass optical filters for accommodating the spectrum of tissue AF and improving separability of classes of healthy and pathological tissues in the LDA-space; (iv) manage the LDA threshold value to maximize the tissue discrimination accuracy, and (v) use the advance methods of image processing to eliminate an impact of background tissue properties on differentiation accuracy. Other technical improvements, such as (i) substitution of the mercury lamp with UV-diode, (ii) increase of the number of UV-excitation channels to combine the AF imaging with the concept of fluorescent matrix [4], and (iii) improvement of the performance of optical systems in both excitation and imaging branches, can bring the system operation to a much higher level.
From a viewpoint of the spatial resolution and the examined object dimensions, all the existing fluorescence imaging methods can be divided into two classes -the methods of microscopic and macroscopic fluorescence imaging [3]. The microscopic techniques of the fluorescence imaging [84,[101][102][103] are characterized with very high lateral and depth spatial resolution, as well as advanced sensitivity, which yields visualizing the internal structure of the skin. These methods allow for studying separate cells and analyzing intravascular and interstitial cell processes. An AF microscopy was combined with Raman microspectroscopy for diagnosis of tissues and, in particular, for detection of BCC in tissue specimens excised during Mohs micrographic surgery [104,105]. Despite the tremendous resolution and sensitivity, microscopic fluorescence imaging techniques are characterized with limited field of view and excessive laboriousness, which makes them unreliable for use in a clinical practice.
The macroscopic fluorescence imaging techniques aim distinguishing between different types of tissues and finding their margins in a large fields of view and with a smaller (compared to microscopy) spatial resolution [58,103,106]. A modern trend in development of novel instruments for fluorescence diagnosis of the skin is associated with the use of multimodal imaging technologies -i.e combining endogenous and exogenous fluorescence phenomena [107] or combining fluorescence imaging with Raman micro-spectroscopy [108]. Furthermore, approaches of multispectral fluorescence imaging, which combine either several excitation and/or several detection channels with advanced methods of signal processing, are extensively developed during the past decades based on exogenous and endogenous fluorescence phenomena [106,[109][110][111].
The multispectral AF imaging approach proposed in this paper belongs to the group of the macroscopic fluorescence imaging techniques. Despite its technical simplicity, the proposed method provides relatively high performance, sensitivity and specificity of tissue differentiation compared to the existing methods of macroscopic AF diagnosis of low-pigmented skin malignancies [106,109,111]. The developed method allows for distinguishing both ordinary and keratinized BCC lesions from healthy skin thanks to the favorable combination of the narrow-line-width UV-excitation, the four orthogonal channels of image detection and the LDA-based approach for tissue discrimination, which relies on the use of normalized spectral AF intensities. The latter problem cannot be solved by simple analysis of the AF intensity since the AF intensity of ordinary BCC is lower, and the AF intensity of keratinized BCC is higher, compared to healthy skin [58]. The developed method provides discrimination of tissues even in case of simple low-aperture optical elements and standard uncooled CMOS-cameras are employed. This makes the proposed technique rather cheap; thus, more attractive and reliable for practical use in a clinic, when compared to the complex and expensive methods of the AF microscopy [84,[101][102][103] and the techniques, which combine fluorescence and Raman scattering phenomena [104,105,108].
In our opinion, the most promising application of the proposed multispectral AF imaging technique is associated with accomplishing the complete removal of the tumor and maximal preservation of the normal skin [112] by complementing such well-known techniques as Mohs micrographic surgery [113] and micrographic Doppler velocimetry [114], which are widely applied nowadays for removing BCC and SCC of the skin. In Mohs micrographic surgery, microscopic examination of all excised tissues occurs during the surgery, thereby providing tremendous accuracy and eliminating the need to estimate how far laterally or deep the cancer spreads [115]. In case of further increase of the described technique sensitivity, it could be applied in order to limit the amount of the tissue to be examined via microscopy during the cancer removal, and, as a consequence, to reduce the laboriousness and duration of the Mohs surgical procedure. Furthermore, this technique could be generalized for studying other types of biological objects, such as bacteria, or even for applications in pharmaceutical and material sciences, in chemical industry and food inspection.

Conclusions
In conclusion, in this paper, we proposed an approach for differentiating BCC and healthy skin, which relies on the use of the multispectral modulation AF imaging combined with the LDA-based processing of the normalized tissue AF intensities in four spectrallyorthogonal imaging channels. We enhanced the signalto-noise ratio and suppress the non-fluorescent background in the AF images by employing the modulation of the UV-radiation and the Fourier-domain demodulation of the AF visible images. We assembled the experimental setup and applied it for studying the samples of freshly excised BCC in vitro. The observed results demonstrate an ability for distinguishing both ordinary and keratinized BCC from healthy skin justifying the perspectives of the proposed technique use for diagnosis of BCC and other low-pigmented malignancies of the skin.

Disclosures
All authors declare that there is no conflict of interests in this paper.