Biophysical basis of skin cancer margin assessment using Raman spectroscopy

: Achieving adequate margins during tumor margin resection is critical to minimize the recurrence rate and maximize positive patient outcomes during skin cancer surgery. Although Mohs micrographic surgery is by far the most effective method to treat nonmelanoma skin cancer, it can be limited by its inherent required infrastructure, including time-consuming and expensive on-site histopathology. Previous studies have demonstrated that Raman spectroscopy can accurately detect basal cell carcinoma (BCC) from surrounding normal tissue; however, the biophysical basis of the detection remained unclear. Therefore, we aim to explore the relevant Raman biomarkers to guide BCC margin resection. Raman imaging was performed on skin tissue samples from 30 patients undergoing Mohs surgery. High correlations were found between the histopathology and Raman images for BCC and primary normal structures (including epidermis, dermis, inflamed dermis, hair follicle, hair shaft, sebaceous gland and fat). A previously developed model was used to extract the biochemical changes associated with malignancy. Our results showed that BCC had a significantly different concentration of nucleus, keratin, collagen, triolein and ceramide compared to normal structures. The nucleus accounted for most of the discriminant power (90% sensitivity, 92% specificity – balanced approach). Our findings suggest that Raman spectroscopy is a promising surgical guidance tool for identifying tumors in the resection margins.

staff training and physician training. As these requirements pose significant barriers to its use, Mohs is used in less than half (~40%) of cases, and less effective treatments such as standard surgical excision are more frequently employed [4]. Moreover, disparities in access to care, such as in among different geographic regions and racial and ethnic groups, leads to substantial underutilization of Mohs in underserved populations [4,5]. Therefore, a critical unmet need exists for low-infrastructure technologies that would enable general dermatologists to perform resections with high accuracy.
Raman spectroscopy is a nondestructive and label-free optical technique, and has demonstrated great clinical merits for tumor margin assessment in numerous types of cancer, including skin [6], brain [7,8], oral cavity [9,10], breast [11,12], and stomach [13]. Those studies either use fiber-optical Raman probe for single point sampling or Raman microscopy (also called Raman microspectroscopy) for optical imaging. One major advantage of Raman microscopy is that it has high resolution and sectioning comparable with that of conventional histology. Moreover, it provides an objective diagnosis with minimal tissue processing. Previous studies have demonstrated Raman microscopy can discriminate BCC from normal skin tissues with sensitivity of 90 -100% and specificity of 85 -93%. Nijssen et al. was among the first to discriminate BCC from surrounding normal tissue using Raman spectroscopy and reached 100%/93% sensitivity/specificity (15 patients, 59 spectra) [14]. Lieber et al. developed an in vivo Raman microscopy and achieved 100%/91% sensitivity/specificity (19 patients, 42 spectra) in classifying BCC and squamous cell carcinoma from paired normal skin tissues [15]. Larraona-Puy et al. demonstrated the ability of Raman microscopy in discriminating BCC from surrounding normal tissue with 90%/85% sensitivity/specificity (20 patients, 329 spectra) [16]. Kong et al. combined Raman microscopy with autofluorescence imaging to increase acquisition speed and achieved 100% sensitivity and 92.9% specificity for discriminating BCC [6].
Prior studies mostly utilized statistical algorithms to extract the spectral differences between BCC and normal tissue, such as principal component analysis [6,14], linear discriminant analysis [17] and maximum representation and discrimination feature [15]. Although such methods provided high diagnostic accuracy, they did not elucidate the nature and biochemical processes responsible for the spectral differences. Understanding the biophysical basis of the discriminatory power of Raman spectroscopy would increase fundamental knowledge of cancer processes as well as lay the groundwork for improving the diagnostic performance of the technology [18]. Therefore, our aim in this study is to obtain biophysically relevant markers from Raman spectra of BCC and surrounding normal tissue, and then build diagnostic model to guide BCC tumor margin delineation.
Our group recently proposed a biophysical human skin cancer model, an inverse model that infers the skin's biochemical makeup from its Raman spectrum [19]. Different from previous studies that selected a number of Raman bands as "fingerprints" to discriminate between healthy skin and tumor regions [16,20], our method is based on the fitting of pure spectral components. We validated the model using previous in vivo human skin cancer screening data [21], and demonstrated the feasibility of Raman spectroscopy to capture relevant biophysical changes accounting for the in vivo diagnosis [22]. Later, we presented a preliminary study of BCC tumor margin detection using the biophysical model based on a small data set from 14 patients [23]. This study demonstrated the feasibility of detecting biophysical changes between BCC and five primary normal structures (epidermis, dermis, hair follicle, sebaceous gland and fat), but has several limitations: firstly, the number of patients is small; secondly, inflamed dermis was not included in the study, which may be confused with BCC in histopathological diagnosis [6,16]; finally, a more comprehensive analysis is needed to link our biophysical approach with molecular vibrational approaches [24,25].
Here, we demonstrate that Raman spectroscopy is highly sensitive in capturing the biochemical differences between BCC and surrounding normal skin structures. Based on these biochemical differences, we can develop diagnostic algorithms to accurately discriminate BCC in Mohs excisions, which supports the future development of intraoperative assessment of tumor margins.

Patients and sample preparation
This study was approved by the Institutional Review Board (IRB) at The University of Texas at Austin and the Seton Healthcare Family. A total of 30 frozen tissue blocks were collected from 30 patients who had undergone Mohs at Austin Dermatologic Surgery Center. 18 samples were found to have both BCC tumors and surrounding normal tissue, and 12 samples contained only normal tissue. Before the Raman experiment, 20μm tissue sections were sliced from frozen tissue blocks at −22C and transferred onto low Raman background glass slides (MgF2 substrates). Serial sections were transferred onto microscope slides for hematoxylin and eosin (H&E) staining. The H&E technique stains cell nuclei in purple (hematoxylin), and intracellular or extracellular protein in red (eosin). Tissue structure identification and histopathological diagnosis were provided by a board-certified dermatologist.

Raman imaging experiment and data preprocessing
Raman imaging was performed using a custom-built confocal Raman microscope. We used a longer wavelength laser (830nm) to minimize tissue autofluorescence. Reflectance confocal and bright-field images are also collected simultaneously. A detailed system description can be found elsewhere [19]. The lateral, axial and spectral resolution of the system was approximately 1 μm, 8 μm and 8 cm −1 , respectively. The power delivered to the sample was approximately 45mW. Raman images were collected from "tissue-level" regions varying from 60 × 60 μm 2 to 100 × 100 μm 2 (2μm steps, 2s per step). The "tissue-level" regions consist of ~10 -100 cells, approximating the resolution of a dermatologist reading an H&E slide.
Raw Raman spectra underwent wavenumber calibration, dark noise removal, cosmic ray removal, smoothing, and a fifth-order polynomial fitting [26] to remove tissue fluorescence background. Spectral response calibration was conducted using a tungsten halogen lamp (LS-1-CAL, Ocean Optics, FL, USA). The effective spectral range was 800 to 1790 cm −1 . Data were normalized to the area under curve.

Clustering analysis
Raman pseudo-color images were generated by k-means clustering. K-means is an unsupervised algorithm for cluster analysis and can easily handle large amounts of Raman spectroscopy data for cell [27] and tissue [14] imaging. The first 100 principal components accounting for 95% -99% of the variation in the data set served as input for K-means. The number of clusters was determined by visual comparison of the pseudo-color image and histopathology. Each cluster was represented by a centroid Raman spectrum and assigned a different color. To eliminate spectral outliers, any spectrum that belonged to a cluster that was more than three times the standard deviation from the mean of that cluster was omitted [14]. We then annotated the centroid Raman spectrum of each cluster as either BCC or normal skin structures: inflamed dermis (Inf), epidermis (Epi), dermis (Der), hair follicle (HF), hair shaft (HS), sebaceous gland (SG), and fat. HS is a long filament in the center of the follicle extended above the surface of epidermis (also called hair). HF is the sheath of cells and connective tissue that surrounds the root of a hair. HF and HS were separated because they had heterogenous biochemical composition [17]. While muscle tissue can be present in Mohs sections and has been studied in other Raman studies [6,20], it generally appears in a small minority of cases; thus, we have excluded it from this study.

Model fitting and statistical analysis
Each centroid Raman spectrum was described as a linear combination of the model components according to a non-negative linear least-squares fitting criteria. The fit coefficients were then visualized using scatter plots. The variation of the fit coefficient is represented by the error bar generated by the 95% confidence interval. The fit coefficients of BCC and individual normal tissue structures were statistically compared. To account for dependencies in the data inherent from measuring multiple skin structures per patient, linear fixed-effects models were employed with the skin structures (epidermis, dermis, etc.) treated as a fixed effect and the patient treated as a random effect. The models were fitted using restricted maximum likelihood and p-values were derived from t-tests using Satterthwaite approximations [29,30].

Diagnostic algorithm
A logistic regression classifier was built to discriminate BCC from normal tissue structures based on their fit coefficients. Leave-one-patient-out receiver operator characteristic (ROC) analysis was used to determine the optimum number of input model components., i.e., the models were trained using a subset of 29 patients and tested on the remaining one patient. To avoid overfitting, the number of input model components was in all cases no more than 4.
Leave-one-spectrum-out ROC analysis was also performed for inspection of the misclassified spectra. The area under the ROC curve (AUC) was calculated to measure the discriminatory power of the classification model. The combination of sensitivity and specificity with of greatest clinical value was obtained from the ROC curve. Sensitivity determines the ability of the model to correctly identify the positive group, whereas specificity is the ability of the model to correctly identify the negative group. Positive predictive value (PPV) and negative predictive value (NPV) were derived from Eq. (1) and (2), respectively. PPV is the probability that the positive group (BCC) identified by the model is truly positive. NPV is the probability that the negative group (normal tissue) identified by the model is truly negative.

Annotated tissue spectra database
In total, we obtained 223 centroid Raman spectra from 30 patients, including 50 spectra from BCC, and 173 spectra from normal structures (including inflamed dermis (N = 19), epidermis (N = 26), dermis (N = 47), hair follicle (N = 31), hair shaft (N = 18), sebaceous gland (N = 22), and fat (N = 10)). Figure 2 shows a typical example of Raman experiment. Good visual correspondence was observed between Raman pseudo-color images and H&E images of the serial section. Figure  3 shows the mean Raman spectra of BCC and normal structures. The main differences between BCC and epidermis/hair follicle/inflamed dermis can be found at 1093, 1577, 1663 cm −1 (assigned to nucleus), while the main differences between dermis and BCC can be found at 856, 940, 1248 cm −1 (assigned to collagen).    Table 2.    Figure 7 represents separating BCC from different categories of normal structures using the fit coefficients of two primary model components. Figure 7(a) shows that epidermis and hair follicle have the largest overlap with BCC compared to other normal structures. The overlap occurs when epidermis and hair follicle have comparable level of keratin and nucleus content as BCC. Figure 7(b) shows that fat and sebaceous gland can be easily separated from BCC because they have distinct nucleus and lipid content. Figure 7(c) (d) demonstrate inflamed dermis/dermis can also be separated from BCC using their distinct differences in nucleus and collagen.

Discussio
In this study, skin structur discriminatory components e Our result BCC and norm has statistical tissue structur 3(b), mean BC to nucleus co with the H&E density of cel nucleic contri [6,20].
Another im compared to i HF. Keratin normal epide between epid maturation an 7 [25]. HF is also rich in keratin. The epithelium of HF forms a cylinder with different concentric layers, including the inner and outer root sheath, with each one expressing a distinct pattern of keratin [32]. Because keratin expression is closely related to differentiation of tumors, it plays a significant role in identifying the origin of BCC. Several studies have discovered that BCC may arise from germinative cells within the basal layer of epidermis or follicular structures [33,34].
Furthermore, BCC has significantly smaller amounts of collagen and elastin compared to dermis. The difference spectra in Fig. 3(c) shows that BCC spectra have weaker contribution from 856, 940, 1248 cm −1 assigned to the vibrational modes of collagen type I and elastin. This is likely the result of the epidermal origin of the BCC, high in cellular content and low in connective tissue. We also found collagen in inflamed dermis is higher than BCC, but lower than normal dermis. This confirms the observation of an early study that a dense inflammatory infiltrate appears to have less collagen than normal dermis [14].
Finally, BCC has significantly larger amount of triolein and ceramide compared to dermis, and larger amount of triolein compared to epidermis. The difference spectrum of BCC and dermis also shows higher contribution at 1080, 1128, 1440 cm −1 associated with lipids. A previous biophysical model built upon nonmelanoma skin cancer found an increased contribution of triolein to BCC spectra [35]. Another biophysical model also found an increase in triolein in urological carcinoma lesions [36]. Those studies suggested that maintaining sufficient lipid levels may be necessary to sustain fast tumor growth.
Some normal structures have high biochemical similarity. For instance, the biochemical compositions of epidermis and HF are highly similar, which agrees with the fact that HF is an invagination of normal dermis [17]. As a result, a previous study grouped epidermis and HF together for clustering analysis [6]. The biochemical composition of HS may resemble either HF or SG. This is because HS consists of terminally differentiated keratinocytes that are produced by HF [32], but it is sometimes coated by the sebum secreted by SG [37]. This also explains why the fit coefficients of HS have larger variation compared to the other components. Fat and SG both have low nucleus and high triolein content, so they can be easily discriminated from BCC ( Fig. 7(b)).
We evaluated the diagnostic performance by prioritizing either high sensitivity or high specificity (Table 4). Achieving high specificity is clinically significant in tissue-conserving surgeries such as Mohs, when preserving normal tissue is a critical concern. Appropriate use criteria of Mohs include tumor location (such as "mask areas" of face), size, and patient type [38]. On the other hand, achieving high sensitivity is more clinically significant when the primary goal is to capture the entire margin based on aggressive growth histology rather than preserving normal tissue.
By prioritizing high sensitivity, we achieved 100% sensitivity and 84% specificity in discriminating BCC from all normal tissue structures. Our results show that that nucleus accounts for most of the discriminant ability. By using the fit coefficient of nucleus alone, 100% of the spectra annotated as dermis (40/40) and fat (10/10) were correctly classified. Most of the spectra annotated as SG (95%, 21/22) and HS (89%, 16/18) were also correctly classified. The misclassification in these latter two categories may be due to unknown tissue structures grouped as the same cluster as SG or HS, leading to high fitting error. On the contrary, HF, epidermis and inflamed dermis were the normal structures that were more easily misclassified as BCC. 16 out of 31 spectra annotated as HF were correctly classified (52%), 20 out of 26 spectra annotated as epidermis were correctly classified (77%), and 16 out of 19 spectra annotated as inflamed dermis were correctly classified (84%).
HF and epidermis were most easily misclassified as BCC. Figure 7(a) demonstrates that those misclassifications occur due to high nucleus content in some of HF and epidermis. The main reason is that HF may have abundant basal cells in inner and outer root sheath layer, whereas epidermis is rich in basal keratinocyte stem cells in stratum basale layer and polyhedral keratinocytes in stratum spinosum layer. About 16% of inflamed dermis was also misclassified as BCC. Corresponding with the H&E images indicated that inflamed dermis regions have higher number of nucleus than normal dermis, so their spectra were more similar to BCC than normal dermis.
By prioritizing high specificity, we achieved 99% specificity, 52% sensitivity and 93% PPV. High specificity indicates a region has high risk, so a dermatologist could remove more tissue from the corresponding region with high confidence of it being tumor [39]. In a Mohs guidance setting, this approach could reduce the number of skin samples processed for histopathology. The dermatologist would still make a histopathological diagnosis on the final stage to ensure all the tumors are entirely removed; thus, one may be able to tolerate lower specificity. Ultimately, this system would need to be tested in an intraoperative setting to determine its impact on reducing the number of tissue samples needing histology processing. Our results showed 24 out of 50 spectra annotated as BCC were misclassified as normal structures. All the spectra annotated as dermis, HF, HS, SG and fat were classified correctly. Most of the spectra annotated as inflamed dermis (95%, 18/19) and epidermis (96%, 25/26) were correctly classified.
Although we have demonstrated Raman microscopy is highly accurate in evaluating skin tumor surgical margin, one major limitation is lengthy acquisition time. To raster scan a tissue 1mm 2 in size, it would take around 10 -20 hours, making it unpractical for intraoperative use. To overcome this limitation, several wide-field imaging techniques could be employed. Kong et al. integrated Raman microscopy with tissue autofluorescence imaging and achieved one or two orders of magnitude faster speed [6]. Karen et al. developed fluorescence confocal mosaicking microscopy and proved its potential for rapid assessment of BCC margins during Mohs [40]. Flores et al used fluorescence confocal mosaicking microscopy to enable rapid detection of residual tumor directly in the surgical wounds on patients [41]. Further directions for this work include speeding up Raman acquisition without losing molecular specificity, and combining Raman spectroscopy with wide-field imaging technique for fast intraoperative surgical guidance.

Funding
Cancer Prevention & Research Institute of Texas (RP130702).