2.1.

Instrumentation

An integrated OCT and Raman system was built to acquire the 3D structure of skin cells and their chemical compositions. As shown in Fig. 1, the sample arms of the OCT and Raman modules are kept in the same optical path so the two beams can fall in the same lateral position.

Fig. 1

Schematic of the integrated OCT (blue box) and Raman (red box) system. The OCT comprises the Ce:YAG crystal fiber light source, a Mirau-type interferometer, and a 2D CCD sensor for fast image acquisition. The Raman system includes an NIR narrow-band laser, a long-wavelength-pass filter, and an OSA. The optics in the Mirau module has AR coatings around the 780 nm wavelength. OBL, objective lens; PBS, polarizing beam splitter; AQWP, achromatic quarter-wave plate; DM, dichroic mirror; LF, line filter; BE, beam expander; LWPF, long-wavelength-pass filter.

The OCT’s light source uses a 445-nm blue laser diode as the pumping source, which is collimated by an aspheric lens and then reflected into a $40\times$ objective lens (PLN 40X, Olympus) by a dichroic mirror (LM01-466-25, Semrock) to focus the laser on the crystal fiber for the broadband light source generation. The crystal fiber with silver plating at one end was operated in a double-pass scheme. The residual laser light was filtered through the dichroic mirror to obtain a broadband light source having a center wavelength of 560 nm and a full width at half maximum of 99 nm, as shown in Fig. 2. The broadband emission was coupled into a multimode fiber through a $20\times$ objective (RMS20X, Olympus). A Mirau-based FF-OCT was employed as the backbone for Raman integration. The infinite-conjugate imaging system consists of a water immersion microscope objective (Umplanfl 20X/0.5, Olympus), a 45-deg mirror, a projection lens (AC254-150-A-ML, Thorlabs), and a two-dimensional (2D) CCD sensor (B0620M, IMPERX). The OCT has a penetration depth of nearly $200\mu \mathrm{m}$. The total power on the cell line was 3 mW, which falls in the IEC Class 3R range and is not hazardous for the skin. The lateral and axial resolutions of OCT are 0.8 and $0.75\mu \mathrm{m}$ (in tissue), respectively.

Fig. 2

Output spectrum of ${\mathrm{Ce}}^{3+}:\mathrm{YAG}$ crystal fiber.

The Raman module light source was a 780-nm LD with a volume holographic grating to reduce the line width to ${10}^{-3}$ to ${10}^{-4}\mathrm{nm}$. The near-infrared excitation has negligible fluorescence and a more than $400\text{-}\mu \mathrm{m}$ penetration depth. The spectral side lobes of the LD were eliminated by a laser line filter (LL01-780-12.5, Semrock). The beam then enters a beam expander system consisting of a lens pair with focal lengths of 35 mm (AC254-35-B-ML, Thorlabs) and 300 mm (AC254-300-B-ML, Thorlabs). The second dichroic mirror, ${\mathrm{DM}}_2$ (DMSP650R, Thorlabs), reflects the light into the Mirau objective. The Raman excitation beam is then focused to $1.7\mu \mathrm{m}$ on the sample with a power of 10.5 mW and an optical slice thickness of $5.6\mu \mathrm{m}$. In the Raman detection arm, the backscattered light from the sample is reflected by the third dichroic mirror, ${\mathrm{DM}}_3$ (DMSP805R, Thorlabs), followed by a long wavelength pass filter (LP02-785RS-25, Semrock) to block the residual excitation light and Rayleigh scattered light. Finally, the backscattered light is coupled into a fiber-coupled optical spectrum analyzer (WP785, Wasatch Photonics), which has a spectral resolution of $6{\mathrm{cm}}^{-1}$ at a wavelength range of 802 to 932 nm (350 to $2100{\mathrm{cm}}^{-1}$). The mechanical stage of OCT scanning is also used here, but the distance of each movement is only one-quarter of the field of view (FOV) of the OCT along the $x$ direction ($72.9\mu \mathrm{m}$) and one-third along the $y$ direction ($73.2\mu \mathrm{m}$). Each sample will sweep 12 points with an integration time of 4 s and an average of 100 measurements. It should be noted that the laser power of the Raman system is 10.5 mW at 780 nm, which falls in IEC Class 3B, and may be hazardous for the skin. In our experiments, no visible damage was observed after the measurements.

2.2.

Sample Preparation

The cell lines were cultured with DMEM-high glucose (GibcoTM) and 10% fetal bovine serum solution in a petri dish, then placed in an incubator at a constant 37°C and 5% carbon dioxide concentration to proliferate. After about 1 week of incubation, trypsin-EDTA is added to the dish to decompose the attached proteins between the cells and the culture dish. When the cells are detached from the culture dish, an appropriate serum is added to terminate the trypsin action. And the cell preparation was completed by adding 4% of a fixed solution (paraformaldehyde) to the culture dish. Previous studies have shown that using formalin for fixation affects cells’ lipids and protein components.^33,34 The 4% paraformaldehyde solution causes changes in the cells’ Raman spectrum, but the effect is smaller than that of other solutions.³⁵ A cell concentration of $4\times {10}^7\text{cells}/\mu \mathrm{l}$ was used, and there was no agarose as the cell base material in the Raman measurement. The cell line soaked in 4% fixative (trioxane) was directly aspirated by a burette and dropped on fused silica to complete the sample preparation.

2.3.

OCT Imaging and Raman Spectroscopy Protocol

The OCT scans the skin cells with a 2D translational stage to expand the FOV, and the images were stitched to form a larger 3D volume. A $z$-axis scan is performed for one FOV each time, then moved horizontally to the next FOV, and the same $z$-axis scan is completed. Finally, the 3D stitched volume is about $1166.4\mu \mathrm{m}\times 878.4\mu \mathrm{m}\times 200\mu \mathrm{m}$ ($z$-direction), composed of 16 pieces of FOV (4 by 4). The $z$-axis scan speed was $0.812\mu \mathrm{m}/\mathrm{s}$, so it took about 3 min to scan a 3D volume of a single FOV. The 3D volumetric data of one FOV ($291.6\mu \mathrm{m}\times 219.6\mu \mathrm{m}\times 200\mu \mathrm{m}$), as shown in Fig. 3, was analyzed quantitatively by the ImageJ^®, which is an open-source image processing and analysis program to segment the cells and calculate various 3D features.

Fig. 3

A 3D OCT image of the melanoma cell line. The scale bar is $20\mu \mathrm{m}$.

The image processing has four steps: pre-processing, Gaussian filtering, binarization operation, and measurement analysis. The 3D image is chopped in the pre-processing to remove the strong reflection interface and the bottom cell-free signal. The second step is to perform Gaussian filtering, which results in the convolution of the original image with the Gaussian functions shown in Eqs. (1) and (2). One is to perform Gaussian filtering in the horizontal plane (i.e., $x\text{-}y$ plane), and another is to perform Gaussian filtering in the depth direction ($z$-axis):

The third step is the binarization process, which can be subdivided into threshold setting, watershed, and erosion operations. The threshold setting is based on ImageJ^®’s built-in auto threshold function to differentiate the inter-cell and extra-cell regions.^36,37 The watershed algorithm segments the original cell clusters and the interference of cell debris. The structural element during the erosion process was 3 pixels by 3 pixels. After the mean threshold method, the image is slightly expanded, so erosion is used to retain cell images. Finally, skin cell capture and feature quantification were performed using 3D object counter and 3D ROI manager in ImageJ^®.

2.4.

Machine Learning Protocol

The cell numbers of keratinocytes (HaCaT), SCC (A431), BCC (BCC-1/KMC), primary melanocyte cells, and melanoma cancer cell lines (A375) are 119, 58, 74, 36, and 227, respectively. The cell volume, compactness, surface roughness, average intensity, and internal intensity standard deviation were extracted from the OCT images.³⁸ After removing the fluorescence components from the acquired Raman spectra, the entire spectra (600 to $2100{\mathrm{cm}}^{-1}$) were used. Three discriminating algorithms, i.e., linear discriminant analysis (LDA),³⁹ k-nearest neighbors (KNN) algorithm,⁴⁰ and decision tree (TREE),⁴¹ were utilized with three common architectures for ensemble learning, i.e., bagging, boosting, and subspace.⁴²

MATLAB was utilized for scripting the machine-learning algorithms with 10-fold cross-validation. About 10% data were used as the test set for each fold. The parameters of the LDA, KNN, and TREE are briefly described as follows:

3.1.

OCT Results

The skin cells of this experiment are mainly divided into two categories: the first is the keratinocyte-based cells, including the cell line of keratinocytes (HaCaT), the cell line of SCC (A431), and the cell line of BCC (BCC-1/KMC). The second category is the melanocyte-based cells, which contain primary melanin cells and melanoma cancer cell lines (A375).

3.1.1.

Images of keratinocyte-based cells

Keratinocytes are the most abundant cells in the epidermis. Both SCC and BCC are developed from keratinocytes. This study employs keratinocytes and cell lines of SCC and BCC. We used HaCaT (keratinocyte), A431 (SCC), and BCC-1/KMC (BCC) to analyze the 3D morphology and intensity of the original OCT images. 3D features were extracted from the isotropic and cellular-resolution OCT images, including volume, compactness, surface roughness, average intensity, and internal intensity standard deviation of cells. The “intensity” related features are average intensity and internal intensity standard deviation of cells. The “morphology” related features are volume, compactness, and surface roughness.

As shown in Fig. 4, HaCaT has a smooth surface, the cell shape is close to a circle in two dimensions, and the 3D shape approximates the spherical shape. On the contrary, cancer cells have a protrusion on the surface, which may be related to the aggressiveness of cancer cells.^11,24 Cancer cells tend to grow out of a tentacle-like shape to expand their cancerous areas. In terms of brightness, the HaCaT cell line generally has the lowest average intensity, but it depicts the most significant internal contrast. The low average intensity may be due to the relative homogeneity of keratinocyte structure, so the backscattering light intensity drops significantly.⁴³ Cancer cells are more irregular in appearance, so the backscattered light becomes more intense. Because of the above two characteristics, the machine-learning-based algorithm could distinguish cancer cells from normal cells.

Fig. 4

3D views (left column) and ImageJ processed images (right column) of (a) HaCaT, (b) A431, (c) BCC-1/KMC, (d) melanocyte, and (e) A375. The thick vertical and horizontal white lines in (a), (c), (d), and (e) of the left column represent the reflections from the glass slide in the Mirau module. The ImageJ delineates the boundary of the cells.

3.1.3.

3D features analysis of skin cells

Figure 5 summarizes the features extracted from the five types of skin cells. As shown in Fig. 5(a), our experiment’s volume mean and standard deviation are generally consistent with that from the literature.³⁸ The HaCaT cell line and the A431 cell line were cultured from the spinous layer cells and the granular layer, whereas the BCC cells were mainly cultured from the basal layer cells. The volume of BCC cells is larger than that of the spinous layer and the granular layer, which may be caused by different growth cycles and growth positions. The melanocyte-based cells all evolved from melanocytes, so there is no apparent volume variation. It is shown Fig. 5(b) that the normal cells (HaCaT and melanocyte) have slightly higher compactness than the cancer cells (BCC, SCC, and melanoma). It could be reflected that the formation of cancer cells in the form of protrusion,^11,24 resulting in a shape that is not spherical. Compactness makes it possible to distinguish whether it is a cancer cell. The degree of cell surface roughness in cytology can reflect that it may be cancer cells.^11,24 From Fig. 5(c), the surface roughness of cancer cells can be found to be much higher than that of normal cells. Therefore, surface roughness makes it possible to distinguish between normal and cancer cells.

Fig. 5

(a) Chart of skin cells’ volume compared with literature.⁴⁰ Charts of skin cells’ comparison on (b) compactness, (c) surface roughness, (d) average intensity, and (e) intensity standard deviation. (f) Summary of the effectiveness of the five features on the classification of five cells (H, HaCaT; M, melanocyte; MM, melanoma; B, BCC; S, SCC; O, △, and X represent significant difference ($p<0.001$), slight difference ($p<0.05$), and almost no difference ($p>0.05$), respectively.

In Fig. 5(d), it can be seen that the cells with a high gray value of keratinocyte-based cells are all cancerous cells. In contrast, the melanocyte-based cells exhibit an opposite trend, i.e., normal cells have a higher gray value than cancerous cells. The cytoplasm contains several micron-level organelles, such as the endoplasmic reticulum and mitochondria. The endoplasmic reticulum and peptide are layered, and the thickness of the single layer is less than the system resolution. But the mitochondria are ellipsoids, and the diameter⁴⁵ is close to the system resolution. The long axis is slightly larger than the resolution, and their number is significant. It is speculated that the backscattering of the mitochondria mainly contributes to the scattered signals detected in the cells.⁴⁶ In recent years, molecular biology studies have also observed variations in the mitochondria of normal cells and cancer cells, and mitochondria are the prominent organelles that provide energy for the cell. Therefore, when the cells are more active, the number of mitochondria may be higher, and there may be a clustering phenomenon. Thus, cancer cells can be detected with the results of the keratinocyte-based cell in Fig. 5(d) and the OCT image of Fig. 4. They have more strong backscattered light, so the overall average intensity of cancer cells is relatively high. However, the melanocyte-based cell is just the opposite. Although there are also differences in mitochondria, the high absorption rate and high refractive index of melanin⁴⁴ have more influence than the former, so the difference in average intensity is mainly due to the amount of melanin. The reason is that primary melanin cells are closer to melanocytes in the living body, so the average cell intensity is slightly higher than melanoma cell lines cultured in multiple generations.

The cells’ internal intensity standard deviation can show the intracellular intensity distribution. From Fig. 5(e), it can be found that normal cells generally have a higher tendency. The previous OCT diagram shows that the nucleus and cytoplasm of normal cells are significantly different in intensity from cancer cells, so the classification of normal cells and cancer cells can clearly distinguish the differences. Therefore, this feature is also used as one of the parameters for determining cell types. As shown in Fig. 5(f), the average gray value, surface roughness, and internal intensity standard deviation can distinguish cancer cells from normal cells. In addition, the normal cells, i.e., HaCaT and melanocyte, can be differentiated by the average intensity feature. However, due to the significant standard deviation between cancer cells’ volume and compactness, the OCT image features may not effectively discriminate among the cancer cells.

3.2.

Analysis of Raman Spectra

Figure 6 shows the Raman spectra obtained from the five kinds of cells where the signal from the Mirau objective is removed. The spectra are similar in a macroscopic view, and it is found that the standard deviation of the Raman spectra of normal cells (i.e., HaCaT, Melanocyte) are two to five times that of the cancer cells (i.e., SCC and BCC, and A375). The standard deviation of normal cells’ spectra is high because they may have pathological changes due to poor survival and reproduction ability.

Fig. 6

Raman spectra of (a) HaCaT (keratinocyte cell line), (b) SCC (SCC cell line), (c) BCC (BCC cell line), (d) primary melanocytes, and (e) melanoma cell line. Black lines and pink lines represent the average value and the standard deviation, respectively.

After removal of the fluorescent backgrounds, it is shown in Fig. 7(a) that below the Raman shift of $1200{\mathrm{cm}}^{-1}$, the keratinocyte-based cell cancers (BCC and SCC), in general, have higher Raman signals, which are also depicted in the literature.⁴⁷ Above $1220{\mathrm{cm}}^{-1}$, melanoma and keratinocyte-based cancer cells are comparable in signal strength. The Raman signal decreases above $1350{\mathrm{cm}}^{-1}$ (i.e., 872.18 nm), mainly because the detector material is germanium. Its quantum efficiency sharply declines above 875 nm, so the spectra are difficult to compare above $1350{\mathrm{cm}}^{-1}$. The detailed biochemical characteristics of the molecules associated with the spectral ranges are shown in Table S1 in the Supplementary Material.^47,48 Some spectral components such as $780{\mathrm{cm}}^{-1}$, 925 to $946{\mathrm{cm}}^{-1}$, 990 to $1010{\mathrm{cm}}^{-1}$, 1281 to $1302{\mathrm{cm}}^{-1}$ have peak shifts ranging from 3 to $12{\mathrm{cm}}^{-1}$ as compared with those in the literature. The A375 cell line results show that nearly 80% of the spectrum calculated in 600 to $1350{\mathrm{cm}}^{-1}$ is comparable to the Raman spectrum from the literature.^47,48

Fig. 7

Comparison of Raman spectra. (a) Melanoma cell line (A375) versus keratinocyte-based skin cancer cell lines (BCC and SCC) and (b) keratinocyte-based skin cancer cell lines (BCC versus SCC).

The Raman spectra comparison of BCC and SCC cell lines is shown in Fig. 7(b). It can be found that BCC has a higher Raman signal below $1250{\mathrm{cm}}^{-1}$, which is consistent with that in the literature.⁴⁷ Above $1250{\mathrm{cm}}^{-1}$, the SCC Raman signal prevails from BCC.⁴⁷ However, the spectra are difficult to compare because the quantum efficiency of the photodetector has a sharp decline beyond $1350{\mathrm{cm}}^{-1}$ (872.18 nm). The detailed biochemical characteristics of the molecules associated with the spectral ranges are shown in Table S2 in the Supplementary Material.^47–50 The Raman spectra of the two cell lines have nearly 88% that can be corresponded to those in the literature between 900 and $1350{\mathrm{cm}}^{-1}$.⁴⁷

3.3.

Results of Machine Learning on OCT Image Features and Raman Spectra

Figure 8 shows the discrimination accuracies of applying machine learning algorithms to the 3D OCT image features. The three classifiers are enhanced by the ensemble architectures. Six algorithms were attempted, i.e., boosting + LDA, bagging + LDA, subspace + LDA, subspace + KNN, boosting + TREE, and bagging + TREE. For each of the classifiers, only the best result is chosen in Fig. 8. The classifications of cancer cells with normal cells [Fig. 8(a)] and normal keratinocytes with melanocyte cells [Fig. 8(b)] both have good discrimination results. The mean intracellular intensity (MI) and the standard deviation of intensity (IS) have results with higher accuracy, and each can achieve an accuracy of more than 85%, mainly due to the slight difference in the 3D characteristics between cancer cells and normal cells. In particular, the mean intensity and standard deviation of intensity distribution may differ depending on whether or not it is a cancer cell or a normal cell. When combining the best features of these two categories, the accuracy is also quite high in the keratinocyte cell lines compared with melanocyte cells. The main reason is that the morphological differences in cell size and shape of these two kinds of cells are quite different, so the melanocytes and keratinocyte cells can be better discriminated. However, the classification ability of the OCT images is relatively low for the classification among cancer cells, only reaching nearly 70% accuracy, as shown in Figs. 8(c) and 8(d). It is speculated that although cancer cells can be clearly distinguished from normal cells, there is no significant difference in the morphological features among the cancer cells. Since there are significant chemical composition differences in cancer cells,^47,48 RS is used as an aid. The Raman spectra show a weaker classification ability between normal cells (i.e., large standard deviation in the Raman spectra of normal cells in Fig. 6). It could be superior in the classification of cancer cells with its chemical composition specificity capability.

Fig. 8

Discrimination accuracies using 3D OCT cell image features by machine learning algorithms (TREE, decision tree; KNN, $K$-nearest neighbors; LDA, linear discriminant analysis) on (a) cancerous versus normal cell lines, (b) keratinocyte cell line (HaCaT) versus melanocyte, (c) melanoma versus keratinocyte-based skin cancer cell lines, (d) BCC versus SCC. (V, volume; C, compactness; SR, surface roughness; IS, intensity sigma; AI, average intensity; 1&2, first two top discriminant features; 1&2&3, first three top discriminant features; ALL, all features.).

The full spectra (600 to $2100{\mathrm{cm}}^{-1}$) of the three cancer cell lines are used. Table 1 summarizes the result using the wavenumber-dependent signal intensities. All three cancer cells can be discriminated against using the Raman spectra, and the classification accuracy is near 100%. However, the time for calculating Raman spectral feature is 3 to 10 times that of the 3D OCT image feature [i.e., 2 to 5 s (OCT) versus 6 to 50 s (Raman spectra)]. The advantages and disadvantages of OCT 3D image features and Raman spectra are complementary. OCT is good at distinguishing between normal cell lines, and normal and cancer cell lines. Raman is suitable for discriminating cancer cell lines. Integrating the two can help improve the differentiation between skin cancer and normal cell lines.

Table 1

Discrimination accuracies using all Raman spectra by machine learning algorithms (LDA, linear discriminant analysis; KNN, K-nearest neighbors; TREE, decision tree) on (a) melanoma versus keratinocyte-based skin cancer cell lines and (b) BCC versus SCC.