2.1.

Model

A modified BLL model for diffusely reflected light was tested experimentally in this work. Reflected intensity from the surrounding healthy skin was always taken as a reference in analysis of the malformation-reflected light intensity;¹ therefore, all presented skin chromophore concentration values in pathologies are relative compared with the volunteers healthy skin. Exact values for mean path lengths of the skin-remitted photons at particular wavelengths are needed for processing the measured data. Here, we used experimentally determined values from our previous study of the photon time-of-flight in skin.¹⁰ In this study, picosecond laser pulses at eight narrow wavelength bands were launched into healthy forearm skin of volunteers via an optical fiber, and the diffusely reflected signals were detected at various distances by means of the time-correlated single-photon counting method. By comparing the shapes of input and output pulses [$a(t)$ and $b(t)$], the temporal distribution function $f(t)$ of photon arrivals following infinitely narrow $\delta$-pulse input was found by deconvolution of the integral,

The corresponding distribution of the backscattered photon path lengths in skin was found as

It was assumed that only skin melanin, oxyhemoglobin, and deoxyhemoglobin are absorbing the incident light in this model. Scattering was excluded because diffusely reflected light intensities from the skin pathology and the healthy skin were compared, assuming that scattering properties in the healthy skin and the pathology region are very similar. Reduced scattering coefficient ${{\mu }_s}^{\prime }$ values can change up to two times¹¹ which does not contribute much to the backscattered light intensity, if compared with chromophore absorption. Modified BLL model for diffusely reflected light:

2.2.

Measurement Setup

Our previously created device¹ was used for clinical measurements; the setup scheme is presented in Fig. 1. Skin was uniformly illuminated by three narrow laser-emitted spectral lines using a specially designed disc-shaped light diffuser. Three couples of laser modules with equal emission wavelengths were placed at opposite sides of the light-shielding cylindrical wall of the device, so six radial laser beams were scattered by the diffuser. Three laser lines (448, 532, and 659 nm) via the diffusive reflector simultaneously illuminated the skin region of interest comprising the skin malformation – e.g., nevus, hemangioma, or seborrheic keratosis. RGB camera of a Nexus 5 smartphone was used for spectral image data collection using AZ Camera software. Two orthogonally oriented polarizers were exploited to minimize detection of skin specular reflection – one in front of illumination source and another in front of the camera. The single snapshot approach with exposure time in the millisecond range excluded the motion artifacts in three spectral line images.

Fig. 1

Illumination and image capturing scheme of the experimental device.¹⁴

2.3.

Image Processing

Image processing scheme is illustrated in Fig. 2. When transforming a single RGB image data set (taken under three spectral line illumination) into three spectral line images – one image for each laser line, the RGB crosstalk correction is necessary. The correction algorithm Ref. 15 exploits three spectral sensitivity curves (R, G, and B) of the image sensor. Crosstalk means that each wavelength can be detected at two or three detection bands of the sensor, but with different probabilities – depending on the detection band sensitivity for this wavelength. The crosstalk correction algorithm extracts the contribution of each working wavelength in the output signal of each detection band (R, G, and B). In our case, skin was uniformly illuminated by three laser lines (448, 532, and 659 nm) and we used the provided manufacturer spectral sensitivity values at the R, G, and B detection bands of the sensor for these wavelengths for crosstalk correction. Sets of R, G, and B channel values were collected from each pixel, then the crosstalk correction was performed with further calculations of the corresponding three spectral line images. These images were segmented to separate the pathology from healthy skin area using k-means clustering function in MATLAB.¹⁶ The green 532-nm image was used for segmentation. After that, the mean reflected intensity values from the surrounding healthy skin area at each of the used wavelengths ($I_0(\lambda )$) were calculated; they represented reference values when there were no tissue absorption changes due to skin pathologies. The reflected spectral line intensities from image pixels related to the area of pathology were divided by the reference values to obtain three attenuation coefficients $k(\lambda )=I(\lambda )/I_0(\lambda )$ (Fig. 2), where $k_B$ is related to 448 nm, $k_G$ to 532 nm, and $k_R$ to 659 nm wavelength.

Fig. 2

Image processing scheme.

2.4.

Clinical Measurement Conditions

Clinical in-vivo measurements were taken under permission of the local ethics committee having written consent of all 77 volunteers with skin photo-types I or II (Fitzpatrick classification), aged between 20 and 68, in cooperation with certified dermatologist Anna Berzina (The Clinic of Laser Plastics, Riga). In total, 99 skin pathologies were examined in this study: three basal cell carcinomas, 27 dermal nevi, 12 hemangiomas, 16 combined nevi, one melanoma, 17 junctional nevi, 22 seborrheic keratoses, and one blue nevus. Data for dermal, combined, and junctional nevi, as well as those for hemangiomas and seborrheic keratoses, may be considered as sufficient for primary conclusions while the few cases of skin cancers (basal cell carcinoma and melanoma), and blue nevus can serve only for general illustration.

3.1.

3D Graphs for Spectral Attenuation Coefficients

The measured attenuation coefficients $k_i$ (reflected intensity ratios: pathology versus healthy skin) for the above-mentioned eight groups of pathologies are presented as 3D graphs in Fig. 3; all $k_i$ values are calculated in percentages. Each point in the graph represents one clinical data pixel value from a segmented pathology. Points are arranged more densely where pixels have similar values, and more scattered where only few pixels have these values. For some pathology types, like dermal nevi, we had more samples (27); therefore, 3D graph cloud consists of 400,000 nonoverlapping points. But for other pathologies, like blue nevi and melanoma, we had only one sample each, therefore, 3D graph cloud consists just of 6000 nonoverlapping points.

Fig. 3

Attenuation coefficient 3D graphs in % for different groups of skin pathologies: (a) dermal nevi, seborrheic keratoses, and hemangiomas; (b) blue nevus, melanoma, and basal cell carcinomas; (c) dermal, combined, and junctional nevi; (d) dermal nevi, melanoma, and basal cell carcinomas.

Figure 3(a) compares the $k$-value clouds for three different benign pathologies – dermal nevi, seborrheic keratoses, and hemangiomas. Parts of them are nonoverlapping, e.g., the specific volume related to hemangiomas and that related to nevi can be easily distinguished. Even better separation between malformations can be observed in Fig. 3(b) where spectral attenuations of two malignant pathologies – melanoma and basal cell carcinomas – are compared with those of a benign pathology – blue nevus. All three pathologies here can easily be distinguished; the $k_R$ values for blue nevus and melanoma are lower than those for basal cell carcinomas while melanoma exhibits lower $k_G$ values than the blue nevus. Spectral attenuations of three different nevi types – dermal, combined, and junctional – are compared in Fig. 3(c); they form a compact cloud but still, each type mainly covers a specific volume in the 3D graph. Data for malignant pathologies (basal cell carcinomas and melanoma) and typical benign pathology (dermal nevi) are compared in Fig. 3(d). Again, some values of all three malformations are slightly overlapping but the $k_R$ values for nevi are clearly higher than those for melanoma and lower than those for basal cell carcinomas.

To summarize, the graphs in Fig. 3 exhibit specific volume-shape features for each of the examined eight groups of skin pathologies. This kind of image data representation may find further application in quantitative diagnostics of skin malformations, e.g., for characterizing and identifying of specific skin pathologies by the spatial location of the $k_B-k_G-k_R$ data points determined from the spectral line images.

3.2.

Attenuation Coefficient Cubes and Limiting Values for Skin Chromophore Content Changes

All possible values of chromophore concentration variations for melanin, oxyhemoglobin, and deoxyhemoglobin in moles ($M$) were calculated for all possible attenuation coefficient values from 1% to 100% (full cubes) in frame of the above-regarded modified BLL model (Sec. 2.1.) using the 3D data representation (Fig. 4). Those values can be both positive (increased chromophore concentration) and negative (decreased concentration) since they represent chromophore concentration changes in pathology compared with surrounding healthy skin (shown in the color bars of Fig. 4). Melanin concentration increase in pathology is inversely correlated to the $k_R$ values [Fig. 4(a)], as extinction coefficient of melanin at 659 nm is considerably lower than that at the two other used wavelengths. The color changes (representing concentration increase/decrease changes) in the oxyhemoglobin and deoxyhemoglobin cubes are almost reverse: where oxyhemoglobin values are higher, deoxyhemoglobin values are lower, and vice versa [Figs. 4(b) and 4(c)].

Fig. 4

Full 3D cubes representing possible concentration variations of three main skin chromophores, calculated in frame of the used modified BLL model: (a) melanin, (b) oxyhemoglobin, and (c) deoxyhemoglobin. Color bar, the chromophore concentration increase/decrease in the malformation, expressed in moles $·{10}^{-5}$.

The model-limited chromophore content variation ranges were calculated, as well (Table 1). Variation range for melanin was the largest, from $-23.67$ to 338.93 $M·{10}^{-5}$. Oxy- and deoxy-hemoglobin concentration variations were model-limited in much narrower ranges.

Table 1

Minimal, maximal values, and value ranges for melanin, oxyhemoglobin, and deoxyhemoglobin concentration changes in skin malformations according to the used modified BLL model, ki∈[1%,100%].

3.3.

3D Representation of the Clinical Data

Four groups of nevi were analyzed using the above-described 3D representation: blue, combined, dermal, and junctional (Fig. 5). As well as four other groups of pathologies: basal cell carcinomas, hemangiomas, melanoma, and seborrheic keratoses (Fig. 6). Only points representing the pixel values in clinical data images were left in the 3D graphs of spectral attenuation coefficients. Green color in all graphs represents chromophore zero changes. If relative chromophore content is higher in the pathology than in the adjacent healthy skin, the points in the graph are colored yellow or red; if the chromophore content has decreased, the points are colored blue. The melanin, oxyhemoglobin, and deoxyhemoglobin relative concentration scales are leveled equally for all included pathologies.

Fig. 5

3D representation of chromophore content changes in nevi: (a)–(c) blue nevus, (d)–(f) combined nevi, (g)–(i) dermal nevi, and (j)–(l) junctional nevi. Color bar, the chromophore concentration increase/decrease in the pathologies, moles $·{10}^{-5}$.

Fig. 6

3D representation of chromophore content changes in four examined skin pathologies: (a)–(c) basal cell carcinomas, (d)–(f) hemangiomas, (g)–(i) melanoma, and (j)–(l) seborrheic keratoses. Color bar, the chromophore concentration increase/decrease in the pathologies, moles $·{10}^{-5}$.

Different specific spatial cloud shapes in the presented 3D graphs are related to the examined skin pathologies, along with different $k$-value distributions. Nevi and melanoma exhibit the highest melanin concentration increase values [Figs. 5(a), 5(d), 5(g), 5(j), and 6(g)] while the oxyhemoglobin concentration increase values are highest for hemangiomas [Fig. 6(e)], as could be expected from the lesion anatomy considerations.