Hyperspectral Raman microscopy can accurately differentiate single cells of different human thyroid nodules

: We report on the use of line-scan hyperspectral Raman microscopy in combination with multivariate statistical analyses for identifying and classifying single cells isolated from clinical samples of human thyroid nodules based on their intrinsic Raman spectral signatures. A total of 248 hyperspectral Raman images of single cells from benign thyroid (n = 127) and classic variant of papillary carcinoma (n = 121) nodules were collected. Spectral differences attributed to phenylalanine, tryptophan, proteins, lipids, and nucleic acids were identified for benign and papillary carcinoma cells. Using principal component analysis and linear discriminant analysis, cells were identified with 97% diagnostic accuracy. In addition, preliminary data of cells from follicular adenoma (n = 20), follicular carcinoma (n = 25), and follicular variant of papillary carcinoma (n = 18) nodules suggest the feasibility of further discrimination of subtypes. Our findings indicate that hyperspectral Raman microscopy can potentially be developed into an objective approach for analyzing single cells from fine needle aspiration (FNA) biopsies to enable the minimally invasive diagnosis of “indeterminate” thyroid nodules and other challenging cases.


Introduction
Thyroid cancer is the most common endocrine malignancy and ninth most common overall cancer with an estimated 53,990 new cases in the United States in 2018 [1]. It is more prevalent in females, accounting for more than 75% of the cases. Thyroid cancer can occur in any age, but it is most common after age 30, with increasing aggressiveness in older patients [2]. The cornerstone for evaluating most thyroid nodules is a neck ultrasound followed by fine-needle aspiration (FNA) in sonographically suspicious nodules. Approximately 10-30% of thyroid nodules have "indeterminate" cytology according to the criteria set forth by the Bethesda System for Reporting Thyroid Cytopathology [3]. In these cases the cytopathologist cannot determine if the nodule is benign or malignant and the patient is faced with the uncertainty of whether the thyroid should be surgically removed. Recently, various genetic based molecular studies have been developed to aid clinicians in the management of patients with indeterminate thyroid nodules; but, the positive predictive value has been suboptimal [4]. As such, thyroidectomy remains the treatment of choice, although majority of the excised nodules are ultimately benign. Hence, a novel approach that can more accurately diagnose and differentiate thyroid nodules would avoid unnecessary surgeries and have a major impact in patient care and management.
Raman spectroscopy is a label-free spectroscopic technique based on inelastic scattering of light by vibrational modes of chemical bonds that allows for the identification of intrinsic molecules (e.g. protein, lipids, amino acids, nucleic acids) in cells and tissues. Subtle differences in chemical composition and structure can lead to changes in peak intensities or positions in a Raman spectrum. Raman spectroscopy provides several advantages for cytopathology applications [5,6]. It can provide intrinsic chemical information of the sample without requiring exogenous labels or stains, has subcellular spatial resolution if implemented into a confocal microscope, and is nondestructive and noninvasive. Previous studies have demonstrated the use of Raman spectroscopy to improve the diagnosis of thyroid tissues [7][8][9][10][11]. Here, we extend this technology for diagnosing human thyroid cancers at the single cell level, with the goal of developing Raman spectroscopy as an ancillary spectral cytopathology tool to improve the accuracy of diagnosing thyroid nodules. In this study, we performed linescan hyperspectral Raman microscopy on single cells isolated from benign and neoplastic human thyroid nodules from clinical samples and applied multivariate statistical methods, principal component analysis (PCA) and linear discriminant analysis (LDA), to analyze the multidimensional spectral data for the purposes of optimizing group separation and determining the diagnostic accuracy of the Raman spectral signatures in various thyroid nodules.

Sample collection
This study is approved by our Institutional Review Board (UC Davis, Sacramento, CA). All patients were consented prior to study enrollment. Representative samples of the fresh nodules were collected for the study. Nodules that had insufficient residual tissue after diagnostic sampling were excluded from the study. The diagnostic materials were processed according to routine diagnostic surgical pathology with hematoxylin and eosin (H&E) stain, and the final diagnosis rendered is confirmed by a second pathologist for the study.

Sample preparation
Tissue samples were dissociated into single cells using established methods [12]. Briefly, the samples were incubated at 37°C in a collagenase (Worthington Type 2) solution 300U/ml in Hank's balanced salt solution (HBSS) for a few hours to digest the tissue. After digestion, single cells were isolated from larger pieces of tissue fragments by using a nylon mesh with 70μm pore size (Corning cell strainer). The isolated cells were washed a few times by centrifugation in HBSS, after which the supernatant was discarded and the packed cells resuspended for a few minutes into a 4% paraformaldehyde in phosphate buffered saline (PBS) solution for fixation. The fixed cells were then washed by centrifugation and the supernatant was re-suspended in PBS solution. The cell solution was pipetted onto a #1 thickness quartz coverslip that was mounted in a cell chamber holder (Thermo Fisher Scientific). Cells remained immersed in PBS solution for the duration of the Raman spectroscopy measurements.

Hyperspectral Raman microscopy
Hyperspectral Raman images of individual cells were acquired using a previously published method [13]. Briefly, a master oscillator power amplifier fiber laser system (Sacher-Laser) with a wavelength of 785 nm and a maximum power of 2 W is used as the excitation source. The laser beam passes a narrow 785 nm maxline laser-line clean-up filter (Semrock, LL01-785) to ensure monochromatic excitation and an achromatic cylindrical lens (Thorlabs, f = 100 mm) that focuses the Gaussian beam into a line profile. The cylindrical lens sits on a rotational mount for adjusting the orientation of the line to ensure that it is properly imaged onto the entrance slit of the spectrometer (PI Acton, SpectraPro SP2300i). After the cylindrical lens, the line-profile is imaged by an achromatic lens (Thorlabs, f = 500 mm) onto the back aperture of a 60x, 1.2 N.A. water immersion objective lens (Olympus, UPlanSApo) and focused into the sample plane. The length of the line at the sample plane is 50 µm with a diffraction-limited width. The cell sample sits on a motorized flat top translational stage (ProScan Prior II) of an inverted microscope (Leica, DM IRM), allowing for scan. The Raman signals generated from the line shaped focal region are collected by the same objective lens and separated from the excitation source by a 785 nm dichroic long pass filter (Semrock, LP02-785RU). The Raman signals pass through another razoredge long pass filter (Semrock, LP02-785RE) and are imaged by an achromatic lens (Thorlabs, f = 125 mm), onto the entrance-slit of the spectrometer. The slit is adjusted to a width of 20 µm. A 600 grooves per mm grating is used to disperse the Raman signals from the line pattern, which is imaged onto a back-illuminated deep-depletion CCD detector (PI Acton, Pixis100). The image of the 50 µm long line is projected onto 100 pixels on the CCD chip, resulting in 0.5 µm per pixel. Typical Raman acquisition times per line is 50 seconds leading to a full hyperspectral Raman image of a single cell within minutes by scanning the cell with 1µm step in the direction perpendicularly to the excitation laser line.

Data analysis
Background removal was first performed on the Raman spectra using a fully automated method for subtraction of fluorescence from biological Raman spectra [14]. Raman spectra were normalized with respect to the area under the curve. Multivariate statistical analysis was then performed on the multidimensional Raman spectral data for objective identification and classification of single thyroid cells. PCA is an unsupervised method that is used to identify the combination of Raman spectral features that maximize the data variance. These features are captured in a new set of variables called principal components (PCs) in a reduced dimension. The first few PCs typically account for the majority of the data variance. However, as an unsupervised method, PCA has no prior knowledge about the groupings of the spectral data, which means it is not suitable for the purposes of group separation. LDA is a supervised technique and is useful for discriminating between groups. So, for the purposes of optimizing group classification, a PCA-LDA model was developed in which PCs were used as the input variables for LDA. A 'leave-N-out' cross-validation technique was used to test the classification sensitivity and specificity of the PCA-LDA model. This procedure involves taking all K-N cells as a training set to build the LDA model, which is then used to classify the N 'blind' cells that were left out. This is done repeatedly for every possible group of N in the set of K cells. The accuracy of a prediction cross validation method of Raman spectra was presented by using the confusion matrix, where cells were classified as true negative (TN), false positive (FP), true positive (TP) and false negative (FN). Diagnostic accuracy, sensitivity and specificity were calculated. All homemade algorithms were written in Matlab (Mathworks, USA). Table 1 summarizes the patient sample characteristics used in this study. 228 cells from 10 different thyroid nodules (5 benign, 5 papillary carcinomas) underwent Raman spectroscopic analysis and hyperspectral Raman images were acquired from each individual cell.  shows bright-f d (g) benign sin n (h-l) cell are 003, 1080, 129 sizes of the Ram at there are se . Brightfield and R ular cell for select rt the hyperspe presents the to all pixels belon all subsequent erage Raman s y shadows rep differences be Fig. 2(c), are re m −1 , 1362 cm − r assignment of  [17]. Previou ancer [18,19]. scopy, suggest of the major al. [21], by usi cant increases i arker to discrim n spectra of (a) 12 ray shadows repre ce spectrum (beni shade) and CVPTC r peak intensit o malignant ce phan, and prot igned to phosp which are assi ious literature issues [10,11] Figure 3(a first three pri Raman spect highlight the maximizing t difference spe benign and CV   combination w s from differen roups have use lines [15,22], udy on clinica tral image into sition of the e ll by sampling d diagnostic ac C n alues for with PCAnt types of ed microour work al patient o a single entire cell g only an curacy of our data may be attributed to our method's ability to adequately sample the entire cell. Our preliminary results even show excellent discrimination of cells that cannot be distinguished by current cytopathologic FNA analysis. Future studies will focus on FNA biopsy samples and will analyze other "indeterminate" thyroid nodules. We believe this novel approach can be developed into an objective and accurate ancillary tool for analyzing FNA samples to improve diagnostic cytopathology and avoid unwarranted surgeries.

Funding
Collaborative for Diagnostic Innovation Grant from the UC Davis Department of Pathology and Laboratory Medicine; UC Davis Academic Senate Grant.