4.1.

Image Acquisition Head

Raman imaging with IFS ideally requires a homogeneous laser illumination of the entire image field during the recording process. To verify the conditions of our setup, the illumination of a sample image was simulated by software (OpticStudio, Zemax LLC Delaware). The pseudocolor image of the simulation [Fig. 3(a)] shows that most of the excitation energy is contained in $50\text{-}\mu \mathrm{m}$ diameter spots at the focus positions of every microlens. Because of the filling-factor of 65%, spaces between the microlenses are also weakly illuminated. At least for nonscattering samples, the remaining 35% do not contribute to excitation, especially since Raman signals arising from the gaps between the lenses are not coupled into the fibers at the other end of the image acquisition head. Due to optical aberrations, the illumination geometrical spot sizes are not exactly constant across the focal plane of the MLA but increase slightly near the edge of the field of view. The simulation reveals that excitation energy within the spots drops from the MLA center to the penultimate rows by $\sim 20\%$. From the penultimate row to the last row, there is an additional drop of 40%.

Fig. 3

(a) Simulated excitation intensity distribution at the sample side. The gray scale is logarithmic. (b) unprocessed (linear) camera image of a scale paper placed on the top of the image acquisition head. The spots show the excitation laser light.

For comparison, Fig. 3(c) shows a real camera image of the excitation spot received by switching on the laser and putting a piece of scale paper on the top of the image acquisition head. Due to scattering of the paper, the spots appear larger than predicted by the simulation. To a certain extent, the intensity impression is an artifact of the camera optics. The actual intensities were measured using a laser power meter in combination with a shadow mask. The total power of the excitation laser at the exit of the square fiber was 400 mW. Except for the four border rows, intensities of $\sim 1\mathrm{mW}/0.25{\mathrm{mm}}^2$ were measured. Obviously no significant loss of laser light occurs within the excitation pathway. However, at the border rows, the intensities drop to $\sim 0.6\mathrm{mW}/{\mathrm{mm}}^2$, which is in accordance with the simulation. Obviously, the shaft of the MLA holder causes vignetting. In conclusion, an image field of $18\times 18\text{pixels}$ is approximately homogeneously illuminated to within $\pm 10\%$. It should be noted that the edges of the MLAs do not run exactly between mircrolenses. Thus, the central microlenses are not precisely in the center of the holder and hence, the intensity gradients are not perfectly symmetric.

4.2.

Spatial Resolution for Skin Samples

If a scattering sample such as skin is examined, homogeneous illumination of the surface causes a decrease in spatial resolution because multiple scattering in the sample leads to crosstalk between the pixels: photons originating from illumination of a specific pixel also contribute to the Raman signal at neighboring pixels. To estimate this effect, a Monte Carlo simulation using scattering and absorption coefficients of human skin ${\mu }_{\mathrm{s}}^{\prime }=1.46{\mathrm{mm}}^{-1}$ and ${\mu }_{\mathrm{a}}=0.044{\mathrm{mm}}^{-1}$³¹ has been performed; see Fig. 4. Each fiber of the matrix captures light arising from certain sections of the sample surface. From the magnification factor of the image acquisition head, the aberrations, and the fiber core diameters, it was estimated that these sections are spots with $\sim 100\mu \mathrm{m}$ diameters.

Fig. 4

Monte Carlo simulation of multiple scattering in skin. (a) XZ projection showing the traces of photons impinging the skin sample at zero position. (b) Contributions of signals arising at neighbored excitation spots.

Figure 4(b) shows the contributions of adjacent pixels in a 0.5-mm raster. A signal strength of 1.00 is assumed if only one pixel at detection position (pixel at center of the insert) would be illuminated. If also illuminating the neighboring pixels, additional Raman-scattered photons will reach the pixel in the center. Four neighboring pixels in 0.5-mm distance add to 0.23 signal strength. Four neighboring pixels in 0.71-mm distance add to 0.18 signal strength. At 1.1-mm distance, there are 8 pixels contributing in total to 0.25 signal strength. Adding up all neighboring contributions, it turns out that approximately half of the signal detected on a certain position comes from scattered light originating from adjacent pixels.

The resulting effect on the lateral resolution is shown in Fig. 5 for the same human-skin model, where a homogeneously distributed Raman active component $c(r)$ is restricted laterally to a half space [$c(r)=0$ for $x<0$]. This example is intended to correspond to an experiment where the boundary of a large-area tumor region is examined. It is assumed that the cancerous region has a specific Raman signal that is not present in the healthy part. This assumption is a simplification. In reality, cancer is not indicated by a clear presence or absence of a definite Raman peak, but rather by slight changes of the spectrum. Nevertheless, the assumption is meaningful to depict the consequences of channel crosstalk. Figure 5 shows two situations: (I) all spots are illuminated simultaneously, as it is the case for our setup (black circles) and (II) only one spot located at the measurement position is illuminated, as it is the case for a pointwise scanning Raman microscope. For our setup, the transition of the Raman signal at the boundary ($x=0$) is remarkably broadened to more than 1 mm, whereas a single spot setup would result in a resolution far better than 0.5 mm.

Fig. 5

Calculated Raman signal (normalized to maximum value) for a large-area Raman source restricted to $x>0$ for illuminating all pixels (fiber grid) and a stepwise illumination of only one pixel (single fiber). The abscissa is the position of the receiving fiber relative to the boundary of the Raman source.

The Monte Carlo simulation also provides information about the signal distribution from different depths. For measurements on skin tissue, it is important to know the spatial distribution of Raman-scattering processes detected by the instrument. The Raman signal can be described with an integral over the probe volume which contains a weighting function $w(r)$

Figure 6 shows the depth dependence of the weight function $w(r)$ for Raman signal detected at one spot. The thin curve shows the signal contributions from Raman scattering at different depths as they would be received if only the detection pixel (spot) would be illuminated. In this case, most of the signal would arise in a range from the surface to $\sim 0.3\mathrm{mm}$ depth. The bold curve shows the corresponding contributions if all pixels (spots) are illuminated. In comparison to single pixel illumination, the simultaneous illumination of all pixels results in the detection of Raman-scattering processes in larger depths due to crosstalk as discussed for Fig. 4. In the case of real skin, the most intense Raman signals arise from the epidermis and the top of the papillary dermis ($\sim 50\mu \mathrm{m}$) but can be enhanced down to $\sim 200\mu \mathrm{m}$ using the optical clearing technique.³³

Fig. 6

Monte Carlo simulation of the Raman signal contribution from different depths of skin. $w_z(z)$ is the laterally integrated weight function [see Eq. (1)]. Bold: Illumination of all excitation spots. Thin line: Illumination at the detection spot only. The functions are normalized to unit area.

4.3.

Homogeneity of the Raman Images

The epoxy phantoms were used to examine the spatial signal intensity distribution of the setup. The homogeneity of the samples, more precisely of the signals arising from the phantoms, was verified by using a scanning single channel Raman spectrometer (Laser- und Medizin-Technologie GmbH, Germany; excitation wavelength 785 nm).⁵ At various positions of a phantom, the spectra were measured within a square of 2-cm border length. The variation of all unprocessed spectra was less than $\pm 5\%$, proving the homogeneity of the phantoms (Fig. 7).

Fig. 7

Raw spectra of an epoxy phantom measured with a single-channel spectrometer at different positions within a square of 2-cm border length.

Although homogeneous phantoms are used as samples, the received signal distribution was not uniform. Figure 8 shows the average intensity distributions of the raw spectra of a phantom, which appears as a dome-like shape with center intensities that are approximately four times higher than at the corners.

Fig. 8

Intensity distribution when measuring a homogeneous phantom. Instead of a flat distribution, the result shows a dome shape.

The intensity differences can be associated with (I) the spectrograph’s properties, (II) the characteristics of the image acquisition measurement head, and (III) the scattering properties of the sample.

The spectrograph’s sensitivity (I) depends on the positions of the fibers located in the V-grooves. Due to the design of the spectrograph, light emerging from fibers at the border positions of the V-groove holder is vignetted to a certain amount. The flat-field correction uncovers this characteristic. To perform the flat-field measurement, the fiber bundle matrix (Fig. 2) is removed from the image acquisition head and homogeneously illuminated using an integrating sphere. The fibers in the matrix are sorted with regards to their positions in the V-groove. Fibers at the positions $(X,Y)=(\mathrm{1,1})$ and (20,20) are located at the borders of the V-groove holder, whereas the fibers at (10,20) and (11,1) are in the center. As expected, the spatial sensitivity results in a half-pipe shape showing intensity. Minima at the corners are probably caused by some vignetting due to the limited size of the available integrating sphere (data not shown). A further potential source of inhomogeneities is the distribution of the excitation intensity (II). As shown above, the excitation intensity declines from the center to the border rows. A further reason for the dome-shape is the scattering of the samples (III). As described above, scattering leads to significant signal contributions from neighboring spots. Indeed, fibers in the center of the matrix receive the largest contribution from neighbors. When moving from the center to the borders, the numbers of neighbors and, accordingly, the signal, decrease. Assuming homogeneous illumination and absorption and scattering coefficients of skin, a calculation shows that the relative Raman intensity drops from 1 to 0.6 when moving 10 fibers diagonally from the corner to a center. Finally, a combination of (I), (II), and (III) leads to the dome-like shape shown in Fig. 7. Since the actual absorption and scattering coefficients, and hence the intensity distribution, differ from sample to sample, a universal correction function cannot be applied. Therefore, it was decided to include normalization into the preprocessing of spectra. Fluorescence background was removed using a sixth-order polynomial fitted to each spectrum.³⁴ Subsequently, the standard normal variate normalization was applied.³⁵ It should be mentioned that the laser unit used includes two excitation lasers with 1-nm gap. Thus, the setup is capable of removing background using shifted excitation differential Raman spectroscopy (SERDS).^18,36 However, the SERDS option was not used in this work, since Raman spectra reconstructions based on SERDS curves considerably differ from common Raman spectra, making a comparison to previously published results difficult. Further drawbacks of SERDS are the negative impact of photobleaching and the extension of the measurement time. It was therefore decided to apply a “classic” polynomial-based background removal algorithm. As shown in Fig. 9, highly uniform spectra result from the preprocessing. From neighboring pixels, it was estimated that the variance between the spectra is approximately only twice the variance due to noise.

Fig. 9

Average and range of variation ($\pm \sigma$) for normalized Raman spectra of a homogeneous epoxy sample.

To visualize the variations between the spectra of the Raman image, a principal components analysis (PCA) of the set of 400 spectra of an image was performed. Each spectrum of the image is represented as a linear combination of the orthonormal set of principal components $p_{\mathrm{m}}(\lambda )$:

Fig. 10

Principal component analysis: First principal components (bottom) $p_1$, (center) $p_2$, and (top) $p_3$ describing the largest variance between spectra (the offsets are introduced for clarity).

Except for two small peaks at 825 and 900 nm, the principal components do not contain Raman bands of the epoxy sample, therefore it seems reasonable that they only describe the instrument’s variation of spectral sensitivity.

4.4.

Comparison of Multichannel Setup with Single Channel Spectrograph

The measurements of epoxy phantoms and the PCA described above were also used to evaluate the influence of the Raman-image inhomogeneities in a medical application. A single-channel spectrograph (constructed at the Laser- und Medizin-Technologie GmbH, Germany) has been successfully applied to examine biopsy tissue samples.⁵ For this setup, a Raman image was realized using a motorized XY stage. The spectral resolution was 0.25 nm, i.e., similar as for our instrument. Partial least squares discriminant analysis (PLS-DA) has been used to discriminate normal from precancerous tissue. PLS-DA results in a discriminant function that assigns each tissue spectrum $g_k$ a scalar value $d_k$ by means of a scalar product that contains a weight function $b$ and an average spectrum $\overline{g}$ (Fig. 11), the tissue type is inferred from the sign of $d_k$

Fig. 11

Average tissue spectrum $\overline{g}$ (thin line) and weight function $b$ (thick line) of the PLS-DA.

In principle, each tissue spectrum $g_k$ can occur on any position of the Raman-image. Therefore, the coefficient $s^{(1)}$ can attain any value found in the PCA of the Raman-image inhomogeneity. It follows that for each $d_k$ a distribution of values with a standard deviation given as

Fig. 12

Reduction of sensitivity and specificity for discrimination of two tissue classes (normal, precancerous).

5.1.

Nevus at Forearm In Vivo

The setup was already tested on porcine skin samples ex vivo.¹⁸ To verify the capability for human skin samples in vivo, a forearm of a volunteer was placed on the top of the measurement head. Figure 13(a) delineates the test area as a skin imprint of the measurement head. A nevus with $\sim 3\text{-}\mathrm{mm}$ diameter can be seen in the upper left. The recording time was 2 min. Due to the fluorescence of the melanin, the position of the nevus can easily be seen on the basis of the signal strength of the raw data (not shown). However, the aim of the experiment was to identify the nevus only with the aid of Raman spectra without any fluorescence background. Thus, all fluorescence background was removed by a polynomial fit followed by a normalization of the spectra. To the resulting background-free Raman spectra PCA was applied. Figure 13(b) shows the Raman spectra inside and beyond the position of the nevus in vivo. Figure 13(c) shows the difference of the principal components 2 and 3 (PC2 and PC3) received from the PCA.

Fig. 13

(a) Nevus on a forearm. The circular skin imprint caused by the image acquisition head has a diameter of 20 mm. (b) Raman spectra at the position of the nevus (dotted curve) and beside it (continuous curve). (c) The difference of the principal components PC2 and PC 3 of the Raman spectra reflects the position of the nevus.

The nevus can be clearly seen as peak. Its diameter of the peak’s footprint is $\sim 7\text{pixels}$ diameter, which corresponds to 3.5 mm; i.e., the Raman image reflects the actual size of the nevus at the camera image. Thus, the achieved resolution of 0.5 mm matches the distance of the fibers in the matrix.

5.2.

Discrimination of Skin Regions In Vivo

As a test case, Raman-spectral maps of six skin regions (forearm, volar forearm, palm, thumb, leg, and foot) have been measured on four volunteers. The aim of this experiment was to verify whether the system is able to distinguish between different body parts. Following preprocessing as described above, PCA was applied to the total set of spectra.³⁸ The scores of the first two principal components are shown in Fig. 14.

Fig. 14

Scores of two principal components of skin Raman spectra from four persons in vivo. Colors according to analyzed skin of different body sites.

As can be seen from Fig. 14, the skin regions differ in their spectral features. For example, leg (cyan points) and foot (light brown points) are well separated and can be easily discriminated.

To investigate whether one can discriminate the skin regions using more spectral information, an unsupervised clustering algorithm was applied.³⁹ Each spectrum was reduced to the scores of four principal components. Inclusion of more than four components did not improve the result. The distribution of spectra on six clusters show that the spectra of forearm, palm, and thumb and spectra of forearm and leg are not clearly separated by the cluster algorithm (Table 1). Only the spectra of foot skin are clearly distinct, probably due to the very large thickness of the stratum corneum in this case. The main reason for the missing discriminability of the skin of different body sites is the interindividual variance of the skin. The thickness of skin layers as well as microcirculation of blood in the papillary dermis and the melanin concentration varies for different body sites. The concentration of carotenoids, which are most concentrated in the stratum corneum, is an individual parameter, depending on the skin area⁴⁰ and volunteers’ lifestyle.⁴¹ This is apparent when the cluster algorithm is applied to all spectra of a single person in vivo, as shown in Table 2 for an example. Here most of the skin spectra of a body site are assigned to a single cluster.

Table 1

Partition of skin spectra in vivo from four persons into six clusters obtained by hierarchical clustering. Unsupervised hierarchical clustering using a minimum variance method.39

Table 2

Partition of skin spectra in vivo into six clusters obtained by hierarchical clustering. Spectra from one person.

For discrimination of skin of different body sites without reference to individual properties, a more refined discrimination analysis including, for example, supervised learning and feature extraction is necessary.⁴² Eventually, interarea differences (Fig. 14) exist and should be considered by comparison between cancerous and healthy skin samples. However, this is beyond the scope of this pilot study and would require the investigation of many more samples: clearly the objective of a detailed follow-up study.

5.3.

Validation of Multiplex Raman for Cancer Diagnostics

Raman spectra of human skin cancerous and healthy tissue were recorded ex vivo and investigated as a possible future diagnostic tool for skin cancer in vivo. Pairs of healthy and cancerous skin biopsy samples were supplied from the Department of Dermatology, Charité. The samples were taken from various patients after surgery and are listed in Table 3. The experiments were approved by the Ethics committee of the Charité-Universitätsmedizin, Berlin (EA1/340/16) and conducted according to the declaration of Helsinki. All volunteers gave their written informed consent. Most of the samples were cylindrical since taken by core biopsy. After removal from the patient, samples 4a and 4b were stored at 6°C and measured one day later. All other samples were at first frozen at $-30°\mathrm{C}$ and stored. For the transportation from the hospital to the location of the Raman spectrograph, an ice-filled styrofoam box was used. Before starting the Raman measurements, the samples were slowly defrozen in a refrigerator at 6°C.

Table 3

Biopsy skin samples. BCC, basal cell carcinoma; SCC, squamous cell carcinoma; AK, actinic keratosis. Two consecutive samples originate from one volunteer, where the first sample is normal and the second sample is cancerous.

Figure 15(a) shows the biopsy sample 6b (Table 3) placed at the top of the image acquisition head. Two parallel needles were used to hold the sample by clamping or skewering. Before and during the measurement, the sample was covered by a lid containing a wet paper tissue to avoid drying. For some of the samples, it was necessary to increase the excitation intensity. The maximum output power of the used laser source is 500 mW. Taking into account the MLA fill factor and losses in the excitation pathway limits the excitation intensity to $\sim 0.8\mathrm{mW}/\text{pixel}$ when illuminating all 400 lenses, which is safe for in vivo application. For the samples replacing the $600\text{-}\mu \mathrm{m}$ square core fiber by a $300\text{-}\mu \mathrm{m}$ square core fiber, the excitation power is concentrated to 100 lenses, i.e., the excitation power increases fourfold.

Since the number of available biopsy samples was too low to perform any reliable refined discrimination analysis, it was decided to assess the results on the basis of the average spectra. Biopsy samples of normal skin from four patients and samples of basal-cell carcinoma (BCC) from three patients were measured, summarized, and normalized. The available samples were either entirely affected with cancer or entirely healthy. Thus, within a sample, the spectral patterns were quite similar at all positions. Figure 16 shows the resulting ranges of Raman-spectral values for both types of skin. The comparison of the Raman spectra reveals the following differences between normal skin and BCC:

Fig. 15

(a) Human skin biopsy sample from external ear at the top of the image acquisition head (sample 6b listed in Table 3). (b) Ex vivo Raman spectra of cancerous human skin tissue at position (8,10) (dotted line) in comparison with the corresponding healthy biopsy sample (continuous line). (c) Intensity of the Raman signal at $1448{\mathrm{cm}}^{-1}$ in color levels.

Fig. 16

Range of spectral values ($\text{average}\pm \sigma$) for normal skin (gray) and BCC biopsies (black). Wavenumber region marked with arrows are discussed in the text.

These differences are in accordance with published results received from a single channel Raman setup.⁴³ However, with our setup, far more pixels could be measured in a shorter time. In Ref. 43, a 10-min measurement time was applied to record the Raman spectrum of a single $100\text{-}\mu \mathrm{m}$ diameter spot. In this work, the Raman spectra of 100 spots were recorded within 2 min. Even though the experimental conditions differ to a certain extent (up-to-dateness of the hardware, excitation wavelength, and excitation power), the achieved measurement speed increase from $10\min /\text{spot}$ to $1.2\mathrm{s}/\text{spot}$ is striking and documents the capability of IFS even in the field of optical cancer diagnosis.