2.1.

Skin Tissue Samples

Skin tissue sections were obtained from the Nottingham University Hospitals National Health Service (NHS) Trust. Consent was obtained from the patients and ethical approval was granted from Nottingham Research Ethics Committee. Tissue sections were cut from blocks removed during MMS and standard BCC excision into $20\text{-}\mu \mathrm{m}$ sections for RMS investigations. After the RMS measurements, the analysed sections were stained using conventional hematoxylin and eosin (H&E) staining. Diagnosis was given by a consultant histopathologist.

2.2.

Raman Spectroscopy

A Raman microspectrometer was built based on an inverted microscope (IX71, Olympus, Essex, United Kingdom) equipped with an automated $XYZ$ translation stage (H117, Prior Ltd., Cambridge, United Kingdom), deep-depletion back-illuminated charge coupled device (CCD) detector (DU401A-BR-DD, Andor Ltd., Belfast, United Kingdom) and spectrograph (SR-303i, Andor Ltd., Belfast, United Kingdom) and a $785\text{-}\mathrm{nm}$ continuous wave GaAs diode laser (XTRA, Toptica Photonics, Munich, Germany). The laser power was set to $50\mathrm{mW}$ at the sample to avoid sample damage and ensure repeatable measurements. The objective lens of the microscope has a numerical aperture of 0.75 and a magnification of $\times 50$ . A microscopy camera (2-1C Infinity, Lumenera, Ottawa, Canada) was used to record images of the tissue sections. Figure 1 presents a schematic description of the instrument. The wave-number (vibrational frequency) axis was calibrated using Raman standard samples (ASTM E 1840), such as naphthalene and 1,4 bis (2-methylstyryl benzene) (Sigma, United Kingdom). The wave-number accuracy was found to be $\pm 0.5{\mathrm{cm}}^{-1}$ .

Fig. 1

Schematic diagram of the Raman micro-spectrometer: L785, laser; PF, plasmaline filter; m, mirror; M, microscope; ST, stage; s, tissue sample; NF, notch filter; CL, collecting lens; SP, Czerny-Turner spectrograph; CCD, detector.

2.3.

Building the RMS Database for BCC Discrimination

2.4.

Spectral Imaging of BCC

After the LDA model was built, the ability of RMS to detect and image BCC was tested on a set of six skin sections obtained from three new patients (no samples from these patients were included into the LDA classification model). Raman spectra from a selected region were acquired at $5\text{-}\mu \mathrm{m}$ intervals with $2\text{-}\mathrm{s}$ integration time at each position (total acquisition time was $3\text{to}5\mathrm{h}$ , depending on the size of the sample). Each spectrum was smoothed using the Savitsky-Golay algorithm (five points, second-order polynomial). Raman spectra of the $\mathrm{Mg}{\mathrm{F}}_2$ substrate were detected using a threshold filter in the $1370\text{to}1500\text{-}{\mathrm{cm}}^{-1}$ spectral range and eliminated from the classification model.

Spectra were binned over 10 or $15\mu \mathrm{m}$ to account for tissue heterogeneity, i.e., each new spectra was the average of the four or nine adjacent spectra to ensure that the acquired spectral data were representative of the tissue class. Thus, the spatial resolution of the biochemical images achieved was 10 or $15\mu \mathrm{m}$ , respectively. The LDA model was then applied to predict the class of each spectrum as BCC, epidermis, or dermis. An image was constructed based on the LDA model classification. An alternative method combining unsupervised $k$ -means clustering followed by LDA discrimination was also proposed ( $k$ -means-LDA). The aim of this method was to observe any loss in spectral images caused by the data compression methods required when large data sets are used. The $k$ -means clustering is a classification technique that groups the Raman spectra of the data set into a fixed number of clusters $k$ by minimizing the intragroup and maximizing intergroup differences. First, the initial data set is separated into an arbitrary number of $k$ clusters. Second, the vector of centroids of every cluster is calculated. The centroid is defined as the average expression value for all the samples in that cluster. Then, the distance between every sample and every centroid is calculated, each sample being reassigned to the closest cluster. The process is repeated iteratively until it converges.⁴⁸ To produce qualitative false-color images, the centroid Raman spectra corresponding to each group were obtained by $k$ -means clustering. The LDA model was then applied on the centroid spectra to classify these image groups into BCC, epidermis, or dermis, thus providing an automated objective diagnosis. Note that $k$ -means clustering is not part of the classifier training phase, but it is used only to apply the final classifier to the new data from the unseen tissue sections.

3.1.

Spectral Database

Mean Raman spectra of the 329 measured tissue specimens (127 BCCs, 92 epidermis, and 110 dermis) from 20 patients used to construct the multivariate classification model are presented in Fig. 2a . This figure shows that there exist spectral differences between BCC and epidermis or dermis, in agreement with previous reported works on skin cancer. ^{28, 30, 39} The differences between dermis and BCC appear to be mainly due to the presence of collagen I in dermis and not in BCC, as inferred from the computed spectrum difference shown in Fig. 2c. In addition, the main differences between BCC and epidermis can be explained by the higher amount of DNA in the tumor tissue, as shown in Fig. 2b. The higher amount of DNA is caused by the smaller amount of cytoplasm and higher density of cells present in the tumor, as can be seen in Fig. 3, where a typical H&E image of measured tissue regions of $50\times 50\mu \mathrm{m}$ diagnosed as BCC, epidermis, and dermis is presented. The H&E technique stains cell nuclei in purple/black (hematoxylin), depending on section thickness and the formulation of hematoxylin used, and most components of the cell cytoplasm in pink/red (eosin).⁴⁹ However, the images included in this study showed DNA in dark brown and dermis in pale orange as an effect of the CCD camera calibration. Regions with higher DNA will be darker, as it is the case of cancerous areas. On the contrary, regions with higher collagen will present a paler colour, as it is the case of dermis.

Fig. 2

(a) Mean Raman spectra of 329 tissue specimens—127 BCCs, 92 epidermis, and 110 dermis—from 20 patients used to construct the model. Spectra have been shifted vertically for clarity. Comparison between the Raman spectra of computed differences drmis minus BCC and BCC minus epidermis are shown in (b) and (c) along with Raman spectra of purified Collagen type I and DNA (Sigma, United Kingdom).

Fig. 3

H&E image of a typical skin tissue sections showing measured regions of $50\times 50\mu \mathrm{m}$ , being represented as empty squares. Color code: blue for epidermis, red for BCC, green for dermis. (Color online only.)

3.2.

LDA Classification Model

Based on the comparison among the mean Raman spectra of the three classes shown in Fig. 2, six Raman bands were chosen as “fingerprints” to discriminate between healthy skin and tumor regions. The selection criterion was to maximize differences among classes, thus the following band area ratios were chosen:

For building an automated detection and imaging method, a multivariate model based on LDA was developed, in which the ratios of the peak intensities, i.e., $r_1$ , $r_2$ , $r_3$ , $r_4$ , and $r_5$ , were used as input parameters. Two consecutive LDA routines were used.³⁰ The order of the LDAs takes into account the easiest discrimination among dermis and the other two classes elucidated in Fig. 2a. The model showed that RMS is able to discriminate nodular and morphoeic BCC from healthy tissue with $90\pm 9\%$ sensitivity and $85\pm 9\%$ specificity in a 70% to 30% split CV algorithm. The reported estimates are the means and standard deviations of sensitivity and specificity over randomly chosen partitions. The values obtained for fivefold CV were $89\pm 11\%$ sensitivity in BCC discrimination and $84\pm 10\%$ specificity, showing the stability of our estimates. LOOCV also reinforces this statement, achieving 94 and 84% for sensitivity and specificity, respectively.

A typical result for CV is shown in Fig. 4, where all 329 data from the spectral database are classified into three groups, circles for BCC, diamonds for epidermis, and triangles for dermis. Data used for training the model are represented as empty symbols in Fig. 4, while the symbol for data employed for validation of the model in Fig. 4 are color-filled. The employed algorithm consists of two consecutive LDAs. First, dermis is separated from the other two classes, $\text{BCC}+\text{epidermis}$ , which are considered one only group, by LDA (LDA1). Then, BCC is separated from epidermis by a new LDA (LDA2). The dashed line (boundary 1) and the solid line (boundary 2) represent the 95% target sensitivity discrimination lines of LDA1 and LDA2, respectively. The LDA score plot in Fig. 4 shows that there is a significant clustering of the spectra into three groups corresponding to BCC, epidermis, and dermis. Misclassifications occur mostly between BCC and epidermis, where clusters overlap over the 95% sensitivity boundary. Such overlap is expected considering the great similarities between the mean spectra of BCC and epidermis (see Fig. 2).

Fig. 4

Classification of 329 Raman spectra from BCC, epidermis, and dermis sampled into three groups by two consecutive LDAs. The dashed line (boundary 1) and the solid line (boundary 2) represent the 95% target sensitivity discrimination lines of the LDA1 and LDA2, respectively. Seventy percent of the data were used for training the model, and are represented in the figure as empty symbols. The other 30% were used for validation of the model, and their symbol in the figure are color-filled symbols. Symbol code: circles for BCC, diamonds for epidermis, and triangle for dermis. (Color online only.)

The mean Raman spectra of dermis and BCC in Fig. 2 show large differences in some of the selected peaks used in our model, such as those of $r_5$ , i.e., the main collagen peaks. Thus, we could expect that few dermis would misclassified as BCC and vice versa. However, Fig. 4 shows that several dermis spectra are located in the middle of the BCC cluster. Inspection of these spectra showed that they were more similar to the BCC mean spectrum of Fig. 2 than to the mean Raman spectra of the dermis presented also in Fig. 2. Correlation with the H&E images indicated that these spectra corresponded to inflamed dermis regions, which had a higher number of cell nuclei than normal dermis. The variability in Raman spectra of dermis depending on its distance to the tumor has already been reported.³⁰

Recall (cf. Sec. 2.3.2) that we are also interested in a reliable control of the classification regime. For example, an end-user might decide in real time to increase sensitivity of BCC detection to 99%. In our implementation, this can indeed be delivered immediately, i.e., without retraining the classification model. Certainly, sensitivity (or specificity) can be controlled with respect to any class, but here we illustrate this using BCC as the class of most interest. Thus, when the target sensitivity is 99%, we observe the following CV estimates of the delivered sensitivity: $98.6\pm 2.5\%$ (70% to 30%), $98.9\pm 2.6\%$ (80% to 20%), and 98.4 (leave-one-out); also, the corresponding (uncontrolled) specificity estimates obtained on the training subsets agree closely with those on the validation subsets under all three settings, and are around 76%.

3.3.

Raman Spectral Imaging

The LDA model was applied to create 2-D biochemical images of tissue sections using two different procedures. In the $k$ -means method, at least 11 classes were required to ensure discrimination between BCC and epidermis. As $k$ -means clustering is an unsupervised method, the presence of any skin-irrelevant element or alteration in the sample was detected and classified first as a new class, prior to BCC and epidermis discrimination. After splitting the spectra into 11 clusters, the Raman band ratios of the centroid spectra corresponding to each cluster were introduced as input in the LDA model. Therefore, each cluster, corresponding to a color in the pseudocolor image, was classified as BCC, epidermis, or dermis. The second method consisted of directly applying the LDA model to the individual Raman spectra measured at each location of the analyzed tissue region to classify each unknown spectra as BCC, epidermis, or dermis. The schemes of both procedures are presented in Fig. 5 . Pseudocolor images using both classification methods are presented in Figs. 6, 7, 8 . These figures contain the automated RMS diagnosis and are presented for comparison with the gold standard histopathology images. Typical RMS images of skin sections containing nodular BCC are shown in Fig. 6. Figures 6a, 6b, 6c show the ability of the RMS models to image tumor regions in a tissue section containing only BCC and dermis. Both dermis and tumor were accurately located within the tissue and were correctly classified. Only a few regions of inflamed dermis were misclassified as epidermis. The reason for this misclassification is likely to be due to the lower amount of DNA at those locations. In Figs. 6d, 6e, 6f, RMS was used to detect nodular BCC in a typical section containing all three regions included in the model: BCC, epidermis, and dermis. The correlation of the spectral images with the H&E image is extremely high. The dermis was correctly identified despite the presence of a large number of cells. The $k$ -means-LDA model had a better accuracy in classification of epidermis and fewer misclassifications of BCC as epidermis. Extremely close agreement with H&E–stained images was also obtained with morphoeic BCC in Fig. 7, where BCC regions as small as $30\text{to}40\mu \mathrm{m}$ were detected. In these cases, the $k$ -means-LDA method showed a higher number of epidermis misclassifications as BCC than direct application of the LDA model to each individual spectrum. From a clinical point of view, however, it is crucial that both classification methods were able to detect with high accuracy the presence of nodular and morphoeic BCC within the tissue sections, as well as the dermis regions. Nevertheless, due to the overlap of BCC and epidermis clusters shown in Fig. 4 some of the epidermis spectra were misclassified as BCC. However, this misclassification of epidermis as BCC has a lesser clinical significance, because if a region at the edge of the tissue is predicted as BCC but shows no BCC regions within the dermis, it is likely that it is real epidermis that has been misclassified as BCC. Also, areas located deep into the sample predicted as epidermis are more likely to be BCCs or very highly inflamed dermis than epidermis, unless they belong to hair follicles.

Fig. 5

Schematic diagram of the two different supervised procedures followed to build 2-D biochemical images of skin tissue sections.

Fig. 6

Imaging nodular BCC in skin sections. Comparison between pseudocolor Raman images (b) and (c) and (e) and (f) produced with the two supervised methods and corresponding hispathology (e) H&E images (a) and (d). Color code: yellow, dermis; light blue, epidermis; and dark blue, BCC. Images (b) and (e) were producd by the $k$ -means-LDA method, while (c) and (f) were produced by the direct LDA model. Tissue size: $500\times 500\mu \mathrm{m}$ . (Color online only.)

Fig. 7

Imaging morpheic BCC in skin sections. Comparison between pseudocolor Raman images (b) and (c) and (e) and (f) produced with the two supervised methods and corresponding hispathology (e) H&E images (a) and (d). Color code: brown, substrate; yellow, dermis; light blue, epidermis; and dark blue, BCC. Images (b) and (e) were producd by the $k$ -means-LDA method, while (c) and (f) are produced by the direct LDA model. Tissue size: (a) $240\times 720\mu \mathrm{m}$ ; (d) $240\times 840\mu \mathrm{m}$ . (Color online only.)

Fig. 8

Imaging of a skin tissue section clear of BCC. Comparison between pseudocolor Raman images (b) and (c) produced with the two supervised methods and corresponding hispathology H&E images (a). Color code: brown, substrate; yellow, dermis; light blue, epidermis; dark blue, BCC. Image (b) was produced by the $k$ -means-LDA method and (c) was produced by the direct LDA model. Tissue size: $240\times 540\mu \mathrm{m}$ . (Color online only.)

Finally, the technique was applied to skin tissue sections excised during MMS and declared clear of BCC by histopathology (Fig. 8). Both methods where able to detect dermis with high accuracy and no BCC regions were predicted within the dermis. Some epidermis regions were misclassified as BCC, particularly by the $k$ -means-LDA method, due to the higher spectral similarities between BCC and epidermis.

Therefore, due to the current lower specificity, we envisage that initially RMS may be used to image all tissue layers removed during MMS and only the sections declared clear of BCC or where BCC is detected at the edge of a tissue section would be evaluated by the surgeon using frozen sections to check that no BCC was missed. As the prediction accuracy and specificity would increase in time by inclusion of a large number of tissue specimens, the RMS may be used as an automated imaging method for BCC, eliminating the need of frozen section preparation and histopathology observation. In addition, the provision of automated objective diagnosis may also reduce interobserver variations during the histopathology evaluation of skin sections. These potential changes in surgery practice have the potential to improve the MMS efficiency, enabling more BCC patients to benefit from the best treatment available.