1.1.

Optical Diagnosis

Optical methods present a viable approach for providing automated, noninvasive, real-time diagnosis of skin lesions, as well as guidance of therapy. There has been much interest in recent years to develop “optical biopsy” methods for tissue diagnosis. Several optical techniques have been applied in a variety of organ systems, with varying degrees of success.

A number of optical imaging modalities have been employed to study skin structures and malignancy, including optical coherence tomography (OCT), laser scanning microscopy, and polarized light imaging. Welzel obtained OCT images from various skin lesions and were able to differentiate skin layers and morphological changes.³ Knüttel utilized OCT to measure refractive index and scattering differences in skin structures.⁴ However, OCT typically provides axial sections that are based on morphological structures alone and requires expert interpretation. On the other hand, confocal reflectance imaging typically provides transverse sections of tissues by depth resolution that is again based on tissue morphology. This method has been used to study a number of skin conditions and diseases. ^{5, 6, 7, 8} In particular, Rajadhyaksha constructed a portable confocal microscope for in vivo skin imaging with positional accuracy within $\pm 25\mu \mathrm{m}$ .⁹ Thus, both confocal imaging and OCT are two methods that provide in vivo morphological images that still require an expert to interpret. Both these imaging methods provide little quantitative information and no biochemical information with respect to the disease state.

Pellacani and Seidenari compared polarization microscopy with epi-illuminescence microscopy of pigmented skin lesions and found that the polarization revealed excellent contrast between morphological structures while obscuring the observed pigment colors in the lesions.¹⁰ Yaroslavsky used polarization imaging to accurately detect locations and shapes of dye-enhanced samples with BCC and SCC in vitro.¹¹ Olivier found polarization imaging to be a superior alternative to clinical examination for determining necrotic regions of skin flaps.¹² While these results show that polarization microscopy provides semiquantitative information about tissue state, it provides little biochemical information and lacks the capability for automated, graded tissue diagnosis.

Spectroscopic techniques, in contrast, are based on quantitative measurement of specific native tissue chromophores. Clinical application of these techniques can be performed by unskilled personnel, with automated, statistical algorithms providing objective correlation of measured data with pathological state. In the case of skin cancer diagnosis, perhaps the most often applied spectroscopic techniques have been fluorescence spectroscopy and Raman spectroscopy.

Fluorescence spectroscopy has been used to characterize melanomas¹³ as well as nonmelanoma skin cancers (BCC and SCC) from normal skin spectra.¹⁴ Fluorescence spectra could be correlated to the disease state, and some of these differences were attributed to tryptophan moieties related to the carcinomas. Doukas found a linear relationship between the fluorescence intensity at $295\mathrm{nm}$ and epidermal keratinocyte proliferation.¹⁵ In another study, Gillies found this same relationship in psoriatic lesions and observed a clear distinction between the fluorescence signals of epidermis and dermis.¹⁶ These studies indicate that the biochemical constituents detected by fluorescence spectroscopy are capable of predicting morphological and pathologic changes. However, because there are relatively few autofluorescent biological markers, the fluorescence spectra provide a limited capability to perform graded diagnosis of skin lesions. Moreover, Lauridsen found pigmentation to be the greatest factor in the variability of fluorescence spectra from normal skin measured in vivo.¹⁷ Sandby-Møller also found skin redness due to sun exposure affected the fluorescence spectra, while epidermal thickness had no effect.¹⁸ Thus, intrinsic tissue parameters such as skin color can modulate the fluorescence spectra measured, thus affecting the performance of fluorescence spectroscopy for skin cancer diagnosis.

Raman spectroscopy is an optical technique that probes the vibrational energy levels of molecules. In the Raman spectrum, specific peaks correspond to particular chemical bonds or bond groups. Because of Raman’s chemical specificity, it has the ability to discern the slight biochemical changes associated with malignant transformation, thereby aiding in differential diagnosis. Furthermore, when utilized in a confocal configuration, Raman spectra can be acquired from various depths and locations within the tissue. Several groups have used Raman spectroscopy to study the molecular composition and biochemistry of normal skin. One study on the normal skin measured in vivo shows Fourier-transform (FT) Raman spectra to vary as a function of hydration state and related collagen structure, while pigmentation was found to have very little contribution to spectral differences.¹⁹ Caspers correlated confocal Raman spectra measured in vitro and in vivo with molecular composition and hydration of skin layers. ^{20, 21, 22, 23} FT-Raman spectra have also been measured from normal and psoriatic lesions in vitro, and distinct differences due to lipid content and related keratin content were found.²⁴ Hata found a relationship between Raman spectral differences and carotenoid concentrations and therefore, cancerous and precancerous skin lesions.²⁵ For skin cancer detection, Gniadecka used FT-Raman spectroscopy to characterize skin structures²⁶ and to differentiate BCC from normal skin in vitro.²⁷ Sigurdsson utilized surface-level FT-Raman spectra to discriminate BCC and melanoma from actinic keratosis, pigmented nevi, and normal skin tissue with $>94\%$ sensitivity and $>98\%$ specificity.²⁸ Nuclear and tissue environmental changes between BCC and normal epidermal cells were found using confocal Raman microscopy by Short ²⁹ Nijssen used a confocal system to generate Raman maps of 15 skin sections and were able to diagnose BCCs with 100% sensitivity and 93% specificity.³⁰ Thus, these different studies demonstrate the ability of Raman spectroscopy to detect subtle molecular transformations associated with skin and skin malignancy. Furthermore, a confocal approach has been shown to be essential in order to maximize Raman signal collection from relevant structures while avoiding interference from irrelevant sources. However, none of the published reports explore the capability of Raman spectroscopy to differentiate both melanomas and nonmelanomas (BCC and SCC) from normal skin toward clinical detection, nor compare the diagnostic performance of depth-resolved Raman measurements with non-depth-resolved measurements such as those obtained from a simple fiber probe.

While skin is easily accessible for optical diagnostic techniques, its stratified composition, inherent variability between and within patients, and presentation of malignancy at various depths within the skin provide a unique challenge. Any optical diagnostic technique, then, must not only be able to discern normal skin from various pathological conditions, but also be spatially resolved in order to minimize the effect of extraneous contributors in the superficial layers. Due to its molecular specificity, clinically viable acquisition times ( $<30$ seconds), and spatial resolution, confocal Raman spectroscopy was selected for differential skin cancer diagnosis. The goal of this study was to characterize the Raman signatures of normal and malignant skin and to develop a strategy for clinical implementation. Using an in-house built confocal Raman microspectrometer, spectra were acquired from various normal and malignant skin tissues in vitro at various depths, and the spectra were characterized using simple and multivariate statistical techniques. Logistic discrimination was performed to predict pathological classification of the samples, and preliminary analyses indicate the ability of confocal Raman spectroscopy to successfully differentiate the various types of skin lesions and suggest that further clinical studies are warranted.

2.1.

Skin Samples

All skin samples were obtained under a protocol approved by the Vanderbilt University Institutional Review Board (IRB). Each sample was obtained fresh-frozen from the Vanderbilt University Medical Center immediately after resection. Thirty-nine samples were obtained for this study, of which 17 were normal, 8 basal cell carcinoma (BCC), 7 squamous cell carcinoma (SCC), and 7 melanoma. Normal samples were obtained from either breast reduction or amputation procedures. Malignant samples were partial sections of lesions that had been surgically removed. All samples were maintained at $-80°\mathrm{C}$ until time of spectral study, at which point they were thawed in buffered saline. Because of their small size and harvest from within the resected tumor boundaries, gross tissue pathology from the parent lesion was used as the gold-standard for classification.

2.2.

Confocal Raman Microscope

The confocal Raman microscope (CRM) used in this study is an in-house built system designed specifically for tissue interrogation and is shown in Fig. 1 . The excitation source of the microscope is an external cavity diode laser (ECDL) built in-house, which includes a 150-mW diode at $825\mathrm{nm}$ (DL8032, Sanyo Electric, Japan) and an 1800-line/mm, gold-coated, high-modulation holographic grating (ThermoRGL, Rochester, New York). The ECDL output beam is shaped by a weak cylindrical lens to minimize astigmatism, and an anamorphic prism pair transforms the elliptical cross section into a circular beam with Gaussian energy distribution. A longpass dichroic beamsplitter at $840\mathrm{nm}$ (Omega Optical, Brattleboro, Vermont) separates the excitation beam path from the emitted Raman scatter. A second dichroic (hot) mirror (Edmund Industrial Optics, Barrington, New Jersey) separates both the excitation beam and the Raman scatter from the shortpass-filtered white light illumination source to allow concurrent Raman acquisition and white light viewing of the sample via a color video camera. An achromatic near-infrared-optimized microscope objective ( $0.35\mathrm{NA}$ , Nachet, France) serves both to focus the excitation light on the sample and to collect the Raman scatter. A longpass edge filter at $840\mathrm{nm}$ (Omega Optical, Brattleboro, Vermont) eliminates residual Rayleigh scatter. The Raman signal from the sample is focused into a $100\text{-}\mu \mathrm{m}$ core diameter, multimode optical fiber, which serves as the confocal aperture. The fiber is connected to a holographic imaging spectrograph (HoloSpec f/1.8i, Kaiser Optical Systems, Ann Arbor, Michigan) and liquid-nitrogen-cooled, back-illuminated, deep-depletion CCD (EEV1024EB, Roper Scientific, Trenton, New Jersey) for detection. In addition, a motorized microscope stage (4400RP, Conix Research, Springfield, Oregon) allows automatic positioning of the sample with mechanical resolution better than $0.7\mu \mathrm{m}$ .

Fig. 1

Schematic of confocal Raman microscope used in this study: ECDL=laser, AP=anamorphic prism pair, $\mathrm{C}+=\mathrm{positive}$ cylindrical lens, BP=bandpass filter, DM=dichroic mirror, M=mirror, $\mathrm{BS}=50∕50$ beamsplitter, SP=shortpass filter, LP=longpass filter, and Obj=microscope objective.

Axial and lateral optical resolution of the CRM was determined by the average full-width-at-half-maximum (FWHM) of the intensity profile of a small $\left(1\mu \mathrm{m}\right)$ polystyrene bead as it was repeatedly scanned through the beam focus in the three respective dimensions. This produced an axial resolution of $22\mu \mathrm{m}$ and lateral resolutions of 2.4 and $6.6\mu \mathrm{m}$ in the $x$ and $y$ directions, respectively. The axial resolution was deliberately increased over true diffraction-limited resolution (theoretically $\sim 10$ to $14\mu \mathrm{m}$ with $0.35\mathrm{NA}$ objective) in order to guarantee tissue-level measurement volumes ( $\sim 2$ to 3 cells) while maintaining volume resolution and increasing overall collection efficiency. The discrepancy between the lateral resolutions can be attributed to residual astigmatism in the excitation beam. The spectral resolution of the detection system was calculated to be $<7{\mathrm{cm}}^{-1}$ , based on the holographic grating dispersion, slit width, and CCD pixel size.

2.3.

Data Acquisition and Processing

Raman spectra were recorded from the surface of each tissue specimen (determined by focusing the microscope on the tissue surface using a video image) and at $20\text{-}\mu \mathrm{m}$ increments below the surface, down to a depth of at least $100\mu \mathrm{m}$ (determined by translation of the motorized stage). These depth measurements were made at various locations (2 to 5, depending on sample size) within each tissue specimen to ensure repeatability of the spectral data. All spectra were recorded using a 30-s integration time, with an excitation power of $40\mathrm{mW}$ at the sample. The spectral dispersion of the system was characterized using the atomic emission lines of a $\mathrm{Ne}:\mathrm{Ar}$ lamp; wave number calibration was performed using naphthalene and acetamidophenol as standards. Spectra were binned to half the spectral resolution of the system for direct comparison. High-frequency readout noise and shot noise were removed from the spectra by a second-order Savitzky-Golay filter.³¹ Tissue autofluorescence was subtracted using an automated modified polynomial fitting method.³² To account for inherent variation in intra- and intersample absolute signal intensity, the spectra were normalized to their respective mean intensity across all wave numbers.

2.4.

Data Analysis

For each sample, spectra from all measurement locations were averaged, and a mean spectrum was calculated for each depth. Outlier spectra were determined as those residing outside three standard deviations $\left(\sigma \right)$ of the respective mean spectrum (i.e., from same sample, same depth, different locations), and mean spectra were recalculated with outliers removed. All further data analysis used only the mean spectral values from each sample, at each measured depth.

Integrated Raman spectral surrogates were created from the depth-resolved spectra to approximate spectra that would be obtained from a traditional, nonconfocal Raman probe. These integrated spectra were created by summing the intensity values of the depth-resolved spectra (processed as earlier but without any normalization) at each respective wave number across all measured depths and locations within each sample. While depth-related optical absorption was not mathematically factored into this integration, the depth resolution of the Raman microscope and the associated decrease in Raman intensity with increasing measurement depth allows a general approximation of non-depth-resolved spectra. Outliers, identified earlier, were excluded from integration. Each resultant spectrum (one for each tissue specimen) was then normalized to its mean intensity across all wave numbers to eliminate inherent intersample absolute intensity variation.

Each measurement depth (and the integrated spectra) was evaluated independently to assess the effect of measurement depth on the pathological predictive capabilities of the Raman spectra, using leave-one-out cross-validation. The analysis technique employed has been described previously^{33, 34, 35} and consists of two steps: (1) extraction of diagnostic features from the spectra using the nonlinear maximum representation and discrimination feature (MRDF), and (2) development of a probabilistic scheme of classification based on linear sparse multinomial logistic regression (SMLR) for classifying the nonlinear features into corresponding tissue categories.

Given a set of input data comprising samples from different classes with a given dimensionality, nonlinear MRDF aims to find a set of nonlinear transformations on the input data that optimally discriminate between the different classes in a reduced dimensionality space. It uses nonlinear transforms that are polynomial mappings of the input data. In the case of spectral data, the aim of nonlinear MRDF is to compute $K\ll N$ nonlinear transformation vectors, ${\Phi }_K$ , from $N$ -dimensional (where $N$ is the number of wave numbers over which spectra were recorded) spectra of skin tissues, such that the projections of the input data on ${\Phi }_K$ from the different tissue categories are statistically well separated from each other. Since the dimensionality $N$ is much larger than the size $S$ of the data, a two-stage MRDF with restricted polynomial transform at each stage was used. In the first stage, the input data $x$ (intensities corresponding to wave numbers of the spectra) from each tissue type were raised to the power $p^{\prime }$ and subject to a transform to produce the first stage output features $y_M^{\prime }$ in the feature space of reduced dimension $M⩽S\ll N$ . In the second stage, the reduced $M$ -dimensional output features $y_M^{\prime }$ for each tissue type were further raised to the power $p^{\prime }$ and subject to a second transform to yield the final output features $y_K$ in the nonlinear feature space of dimension $K$ $(K⩽M)$ . Since the nonlinearities introduced in the two stages were different ( $p^{\prime }$ in the first stage, and $p$ in the second stage), this is expected to produce more general nonlinear transforms on the input spectral data leading to improved separation of the final nonlinear features for the tissue types in the new feature space.

Classification with SMLR is a probabilistic multiclass model based on a sparse Bayesian machine-learning framework of statistical pattern recognition. The central idea of SMLR is to separate a set of labeled input data into its constituent classes by predicting the posterior probabilities of their class membership. It computes the posterior probabilities using a multinomial logistic regression model and constructs a decision boundary that separates the data into its constituent classes based on the computed posterior probabilities following Bayes’s rule. Classification of a given set of input data $\mathbf{x}$ is based on the vector of posterior probability estimates yielded by the SMLR algorithm, and a class is assigned to each dataset (transformation of the original spectrum) for which its posterior probability is the highest.