2.1.1.

Cell culture preparation

The experiment was conducted using EGFP-GL261 tumor cells, a murine glioma cell line with characteristics similar to human GBM and with stable expression of the green fluorescent protein (EGFP). GL261 cells were transfected with EGFP using the electroporation technique. The cells were detached using 0.05% trypsin-EDTA (Gibco), washed with 5 ml of PBS and centrifuged at 1200 rpm for 5 min; then, they were re-suspended in $120\mu \mathrm{l}$ of resuspension buffer (Invitrogen) and $5\mu \mathrm{l}$ of DNA were added. Afterward, the cells were electroporated with two 20-ms pulses at 1200 V using Neon Transfection System $100\mu \mathrm{l}$ pipette tips (ThermoFisher Scientific). The amount of DNA used is around 10 to $20\mu \mathrm{g}$ for each electroporation. The transfected cells were plated and maintained in Dulbecco’s modified Eagle’s medium (DMEM, Gibco) containing 10% fetal bovine serum, 1% penicillin–streptomycin ($100\text{units}/\mathrm{ml}$ of penicillin and $100\mu \mathrm{g}/\mathrm{ml}$ of streptomycin, Gibco), 1 mM sodium pyruvate (Gibco) and 10 mM HEPES (Gibco) in a humidified atmosphere of 5% ${\mathrm{CO}}_2$. The culture medium was replaced with fresh medium 2 to 3 times per week. Once the confluence reaches to about 90%, the cells were harvested using 0.05% trypsin-EDTA. The cells were pelleted by centrifugation at 1200 rpm for 5 min (Sorvall T3 Centrifuge, ThermoScientfic). The pellet was re-suspended with PBS and the resulting solution was injected into the mouse brain using a glass capillary.

2.1.2.

Mice preparation

Adult C57BL/6 mice were maintained at a mice facility with controlled temperature and light, where they received food and water ad libitum. The animals were anesthetized with Avertin ($0.2\mathrm{ml}/\mathrm{g}$ body weight) by intraperitoneal injection. Hair at the top of the head was removed using a hair depilatory cream and fixed on the stereotaxic stage. A single skin cut of $\sim 0.5$ to 1 cm was made using a pair of scissors and the exposed sub-cutaneous tissue was gently removed. A hand drill was used to thin the skull at the injection site, which is the visual cortex. Two microliters of EGFP-GL261 cell suspension containing $\sim \mathrm{40,000}\mathrm{cells}$ in total were aspired into the glass capillary and slowly injected into the mouse cortex. The glass capillary was inserted at $300\mu \mathrm{m}$ at first and then retracted $\sim 50\mu \mathrm{m}$ before injection of cell suspension. A total of 5 min was spent for the injection and additional 2 min before removing the glass capillary. The surface of the injection site was then washed with PBS and skin was closed with simple stitches. Iodine antiseptic cream was applied on the stitches to prevent infection. To relieve pain, 5 drops of paracetamol ($100\mathrm{mg}/\mathrm{ml}$) per 100 ml of drinking water was provided for the animals after surgery. All the instruments used were sterilized.

Another surgery was carried out under anesthesia induced by Avertin ($0.2\mathrm{ml}/\mathrm{g}$ body weight) for placing a cranial window. The mice were placed on a heating pad and fixed on the stereotaxic stage. A longitudinal incision of the skin was performed between the occiput and the forehead, followed by skin and sub-cutaneous tissue removal around the injected site. Using a hand drill, the skull was thinned to open a 5-mm-diameter window. Forceps were used to break and remove the skull gently. The dura mater was continuously rinsed with artificial cerebro-spinal fluid: 125 mM NaCl, 25 mM ${\mathrm{NaHCO}}_3$, 1 mM ${\mathrm{MgCl}}_2$, 2.5 mM KCl, 1.25 ${\mathrm{KH}}_2{\mathrm{PO}}_4$, 2 mM ${\mathrm{CaCl}}_2*{\mathrm{H}}_2\mathrm{O}$, and 10 mM glucose. A coverslip (5-mm diameter and 0.17-mm thickness) was used to cover the window and fixed using strong adhesive glue and self-curing dental cement carefully. We chose a quartz coverslip because, when excited at 785 nm, standard glass emits fluorescence at wavelengths corresponding to the “fingerprint region” of Raman spectroscopy. Again, all the instruments used were sterilized.

Finally, another cranial window was placed at a certain distance from the injection site, in order to have an optical access to healthy brain tissue.

2.1.3.

Optical measurements

Eight animals were included in this study. After two weeks from the injection of EGFP-GL261 cells, each mouse was anesthetized and placed under the stereotaxic device (Fig. 1) for examination through our optical fiber probe (Sec. 2.3.1), whose distal end was placed perpendicularly in contact with each of the two quartz windows. Few animals were inspected also in the following days, in order to record tumor progression; in total, we performed 13 experiments.

Fig. 1

(a)–(c) Simplified scheme of the experiment. (a) A quadri-furcated fiber probe is connected to four excitation sources contained in a $45\times 30{\mathrm{cm}}^2$ box. (c) The excitation light is delivered through an optical window on the mice skull. (b) The generated signal (Raman/flurescence/reflectance spectra) is backward collected and sent to the spectrograph. (d) Picture of an anesthetized laboratory mouse during the experiment with the probe tip pointing toward the skull window.

First, fluorescence spectroscopy was used for real-time ($<0.1\mathrm{s}$) detection of tumor areas by taking advantage of the characteristic EGFP signal. The probe was placed in the point of maximum EGFP intensity; then, fluorescence, Raman, and reflectance spectra were acquired from the same region. Spectral measurements were performed also on the second optical window (i.e., on control brain tissue). Total integration times were selected in order to exploit the dynamic range of the CCD and to reduce noise. For fluorescence acquisitions, they ranged between 0.5 and 5 s—depending on the presence of EGFP and the amount of autofluorescence—obtained by averaging 10 recordings and using 5- and 3-mW excitation power at 378 and 445 nm, respectively. Each Raman measurement required averaging 10 recordings, resulting in 25- to 50-s integration times using 75-mW power. Reflectance spectra were acquired by averaging 100 recordings in 0.1 to 5 s, using $\sim 5\text{-}\mathrm{mW}$ typical output power (from a nominal bulb power of $\sim 5\mathrm{W}$). The inspection typically required a total of less than 4 min for each animal, resulting in 1 spectrum for each spectroscopic technique and for each region. Tumor areas were later examined through a multiphoton microscope (Sec. 2.3.2) in order to observe EGFP-labelled cells and to estimate tumor size.

2.3.1.

Fiber-probe system

Spectral measurements were performed through the experimental setup described in Ref. 6. As shown in Fig. 1, this system consists of four excitation sources, a quadri-furcated fiber probe (EMVision LCC, Loxahatchee, Florida), and a spectrograph (MicroHR, HORIBA Scientific, Edison, New Jersey) equipped with a 600-lines/mm grating and with an open-electrode CCD camera (Syncerity, HORIBA Scientific, Edison, New Jersey) cooled at $-60°\mathrm{C}$.

In particular, we used two laser diodes (TEC 042, Sacher Lasertechnik, Maburg, Germany) for exciting fluorescence at 378 and 445 nm, respectively; a narrow-band laser diode at 785 nm (FC-785-350-MM2-PC-1-0-RM, RGBLase, Fremont, California) for near-infrared (NIR) Raman spectroscopy; and a halogen lamp (HL-2000-LL, Ocean Optics, Dunedin, Florida) for reflectance spectroscopy.

The probe, which is connected to all light sources, consists of a bundle of 11 optical fibers: a central fiber ($300\text{-}\mu \mathrm{m}$ core diameter) for delivering 785 nm; 7 surrounding fibers ($300\mu \mathrm{m}$ each) for Raman collection; 2 fibers ($200\mu \mathrm{m}$ each) for delivering 378/445 nm and white-light from the lamp, respectively, and 1 fiber ($200\mu \mathrm{m}$) for collecting both fluorescence/reflectance signals, all of them placed near to the central fiber. A narrow bandpass filter centered at 785 nm was placed in front of the Raman delivering fiber in order to clean the laser line, blocking unwanted fluorescence and Raman signals generated inside the optical fiber itself. A ring-shaped long-pass optical filter was placed at the distal end of the probe, preventing 785-nm laser light from entering the seven collection fibers. A motorized filter wheel allows changing emission filter according to the spectral technique used: a long pass filter at 400 nm (Longpass 400 nm, Melles Griot, Albuquerque, New Mexico) for fluorescence spectroscopy excited at 378 nm; a long pass at 458 nm (LP02-458RS-25, Semrock, Rochester, New York) for fluorescence spectroscopy excited at 445 nm; a long pass at 785 nm (LP02-785RE-25, Semrock) for Raman spectroscopy; and a neutral density filter with $\mathrm{OD}=0.3$ for diffuse reflectance. These filters prevent laser light from entering the monochromator and lamp light from saturating the detector.

The device was designed to be easily handled (as a pen) for fast examination of suspicious tissue areas.

2.3.2.

Two-photon excited fluorescence microscope

High-resolution two-photon excited fluorescence (TPEF) in vivo imaging was performed on a Prairie Ultima Multiphoton microscope equipped with a mode-locked Ti:sapphire laser (Chameleon Ultra II, Coherent, Santa Clara, California). Laser excitation was set at 910 nm for EGFP detection and a maximum power of $\sim 35\mathrm{mW}$ was used for data acquisition. Images were acquired with a water immersion lens (XLUMPLFLN20XW, Olympus, Tokyo, Japan; $20\times$ magnification, $\mathrm{NA}=1$) at $1024\times 1024\text{pixel}$ resolution and zoom = 1, leading to $600\mu \mathrm{m}$ field of view with linear resolution of $0.594\mu \mathrm{m}/\text{pixel}$.

2.4.1.

Pre-processing of spectra

About 26 spectra—13 from healthy mice brains and 13 from GBM tissues—were recorded for each optical technique (fluorescence excited at 378 and 445 nm, Raman, reflectance). The transmission spectrum of the fiber-probe setup was used for correcting all the recorded spectra in order to remove artifacts. Also, further processing was performed on Raman and reflectance datasets. In particular, we used a sixth-order polynomial fitting through an automated routine (Vancouver Raman algorithm¹⁷) for removing NIR fluorescence and background signals from the Raman spectra, which were later normalized to the ${1445\text{-}\mathrm{cm}}^{-1}$ Raman band (corresponding to ${\mathrm{CH}}_2$ and ${\mathrm{CH}}_3$ deformations¹⁸) intensity. Reflectance spectra were corrected for the lamp emission spectrum and normalized to the intensity recorded at 570 nm, which is one of the oxy/deoxy-hemoglobin isosbestic points.

2.4.2.

Spectral classification analysis

This study aims to discriminate healthy brain and glioma tissues in a label-free modality, e.g., based on the information recorded through Raman and reflectance spectroscopies. In this framework, PCA and LDA are well-known methodologies^19–21 used for the classification of spectral data;^6,22,23 we implemented each method in a separate custom-made routine on LabVIEW (National Instruments, Austin, Texas). PCA was used for reducing noise and redundant information by selecting four principal components (PCs) of each dataset and discarding all the others. Also, a decreased dimensionality reduces the risk of overfitting the recorded data and, thus, overestimating classification accuracy. LDA, then, was employed to build a linear combination of the selected PCs in order to maximize discrimination between different tissue classes. We specifically designed an LDA routine for generating a classifier and maximizing the area under its receiver operating characteristic curve (auROCc), a parameter useful for estimating the classification accuracy.^24,25 For each spectrum, the algorithm creates a linear combination by multiplying each selected PC loading by a random number between 0 and 100, then summing all the four products into a score. The ROC curve of normal and tumor scores and the corresponding auROCc values are then calculated and recorded. The algorithm iterates such routine, each time comparing the new auROCc value with the maximum value among those ones already calculated, and stops after $N$ consecutive iterations (e.g., $N=500$) where no further maximum is found. Finally, the linear combination providing the best accuracy is saved and can be used to classify a group of unknown spectra. A cross-validation function can be selected for dividing the dataset in ($k$)-folds, using ($k-1$)-folds for training the classification algorithm and 1 fold for testing it, and repeating the process ($k$)-times by rotating the folds in the two groups.

Two additional analytical methods were applied to the recorded spectra: ratiometric scoring (i.e., calculating the ratio between photon counts recorded at two different wavelengths/wavenumbers) and slope analysis (i.e., applying a linear fit to a region of the spectrum in order to derive its slope). Both analyses were implemented in LabVIEW.

2.4.3.

Image analysis

TPEF images were reconstructed and analyzed through ImageJ (National Institutes of Health, Bethesda, Maryland). The radial (i.e., parallel to brain surface) extension of each tumor was estimated by measuring the average distance between opposite margins. Instead, tumor axial (i.e., perpendicular to brain surface) extension could not be estimated due to the limits in NIR light penetration (and visible fluorescence back-propagation) in highly scattering media such as brain tissue; roughly, the microscope was not able to image deeper than $400\mu \mathrm{m}$.

3.2.1.

Tumor localization

Tumor areas in mice brain were detected through their EGFP emission peak excited at 378 and 445 nm [Figs. 3(a) and 3(b), respectively]. In fact, after placing each animal on the stereotaxic stage, we continuously acquired fluorescence spectra while moving the fiber probe along the tumor window; then, we positioned the probe where the fluorescent label signal was maximized. We found that EGFP increased day by day, as tumor size increased over time; also, we recorded the autofluorescence spectrum of healthy brain tissues from the center of the other quartz window for a qualitative comparison against spectra of EGFP-labelled tumor. However, only a small number (maximum 3) of time points was recorded for each animal, due to the limited time window—few days—available for measurements before its death; moreover, in order to reduce the stress for the animal, a 1-day pause was interposed between every two sessions.

Tumor radial extension was correlated with fluorescence intensity excited at 445 nm and emitted at 510 nm, corresponding to the EGFP emission peak [Fig. 4(c)]. These variables seem to be related ($R^2=0.828$, Pearson’s $r=0.908$ for a linear fit): normal dots are clustered in one small area, while tumor dots follow an approximately linear trend. However, there are few considerations to be made. In the first place, peak intensity is calculated relatively to tissue autofluorescence emitted at 660 nm; thus, biological variation may introduce a certain degree of “noise” in the reported values. Second, the intensity of EGFP emission depends on the spatial convolution between tumor volume (i.e., its radial and axial extensions) and the excitation volume of the probe. Since the radial laser spot is $\sim 1{\mathrm{mm}}^2$, fluorescence intensity for tumors larger than that size depends only on their axial extension; it can be expected that, in such cases, EGFP intensity increases with tumor depth. Then, if radial and axial extensions grow proportionally together, this is a possible explanation for seeing a linear trend in Fig. 4(c).

Fig. 4

(a) Results of the scoring algorithm based on the linear combination of Raman and reflectance PCA-LDA scores, presented as normal/tumor (black/red) scores generated by the classifier. (b) ROC curves of the multimodal classifier; the red dot corresponds to a suggested threshold for discriminanting normal and tumor scores. (c) and (d) Correlation of the tumor radial size with, respectively, the EGFP peak intensity and with the scores of the PCA-LDA classifier applied to reflectance data. (e) Correlation of the EGFP peak intensity with the reflectance scoring. A linear fit (dash-dotted red line) of the tumor datapoints was superimposed for each dataset presented; in graph (c), the intercept was fixed to the mean value of normal datapoints. (f) Slope values of normal (black) and tumor (red) refletance spectra calculated in the 625 to 730 nm range.

3.2.2.

Raman results

We recorded two Raman spectra from each animal inspection, i.e., from both healthy and tumor brain regions. Figure 3(d) shows the average spectra of these tissue types, which are characterized by the typical Raman peaks of amino acids (such as proline and valine bands¹⁸ at $950{\mathrm{cm}}^{-1}$), nucleic acids²⁸ (1060 and $1575{\mathrm{cm}}^{-1}$), lipids ($1445{\mathrm{cm}}^{-1}$), amide III and tryptophan²⁹ ($1210{\mathrm{cm}}^{-1}$), and amide I²⁷ ($1650{\mathrm{cm}}^{-1}$). By normalizing spectral intensity to the ${1445\text{-}\mathrm{cm}}^{-1}$ band, the major differences seem to reflect higher amino acids, tryptophan, and amide III relative (i.e., with respect to the lipid content) concentrations in healthy brain, and higher amide I relative concentration in tumor tissues. These findings could be partially explained with the altered cell metabolism in gliomas, which causes an increased rate of lipid synthesis during their growth³⁰ and lower levels of amino acids.³¹ A classifier based on ratiometric scoring resulted in 77% and 79% auROCc (data not shown) by calculating the ratio between Raman photon counts registered at the ${1450\text{-}\mathrm{cm}}^{-1}$ and, respectively, 985- and ${2030\text{-}\mathrm{cm}}^{-1}$ bands.

The first 15 PCs of the Raman dataset were calculated through standard PCA. We selected four of them, based both on the discrimination capability of each one (evaluated through its auROCc) and the improvement provided to the PCA-LDA analysis (again, measured as increased/decreased auROCc value). A classification model was generated from the linear combination of PC3, PC7, PC10, and PC13; the resulting classifier was denoised and reported in Fig. 3(d) in order to show the correlation between its major peaks and the observable differences between normal and tumor spectra. Finally, each Raman spectrum was assigned a “score” [Fig. 3(e)] based on the classifier, whose auROCc was found to be 92% [Fig. 3(f)] when training the algorithm on all dataset spectra. Then, in order to validate this result, we performed an automated 4-fold cross-validation (4f CV) analysis—10 normal and 10 tumor samples for training, 3 normal and 3 tumor samples for testing—resulting in 86% average auROCc (see Table 1).

Table 1

Specificity/sensitivity and auROCc values obtained using the classification models based on PCA-LDA. First and second rows show the results from the separate classification of Raman and reflectance datasets, respectively, while the third row reports their combined analysis. The last column on the right reports auROCc values obtained through 4f CV.

3.2.3.

Reflectance results

Figure 3(g) shows the average reflectance spectra recorded from normal murine brain and GL261 tumor areas. Each spectrum was normalized to an isosbestic point of oxy- and deoxy-hemoglobin (570 nm) in order to take advantage of their different molar extinction coefficients, to qualitatively compare their relative abundance in the two examined tissue types and, hence, to derive information on the corresponding oxygen levels. In fact, one of the most characteristic features of these spectra regards oxy-hemoglobin absorption peaks³² (540 and 575 nm), indicating that such substance is one of the major absorbers. Normal and tumor spectra intersect each other also in the others’ isosbestic points³³ (452, 529, 545, and 584 nm), with the only exception of 500 nm. It is unclear whether the presence of EGFP may be partially responsible for missing the latter intersection, because a similar behavior can be observed in Ref. 14 between label-free normal and cancer brain tissue spectra.

The major differences in Fig. 3(g) take place between 600 and 730 nm: murine gliomas show higher relative absorption in this region, suggesting a higher concentration of deoxy-hemoglobin with respect to oxy-hemoglobin, thus lower oxygenation with respect to normal brain. Indeed, this is consistent with the typical hypoxic environment³⁴ of tumors. These patterns are so prominent that a classifier based on ratiometric scoring resulted in 82% auROCc (data not shown) by calculating the intensity ratio between 570 and 690 nm.

Another feature can be observed at longer wavelengths ($>625\mathrm{nm}$), where healthy spectra have slightly higher slope than tumors. It is known³⁵ that cell nuclei dimension affects reflectance spectra in this region, specifically with respect to the reduced scattering coefficient (which is inversely proportional³⁶ to the square root of scatterers diameter). Thus, as reflectance intensity is an approximately linear³⁷ function of such coefficient, the ratio between normal and tumor spectral slopes could reflect their different nuclear size. We implemented a linear fitting algorithm in LabVIEW for deriving normal and tumor slopes between 625 and 730 nm [Fig. 4(f)], finding a normal-to-tumor slope ratio of $2.6\pm 1.5$. Formally, this value is consistent with typical tumor-to-normal root square nuclear size ratio (for example, we calculated a $1.25\pm 0.04$ ratio from data reported in Ref. 38), although the reported standard error is relatively high. A classifier based on slope values resulted in 73% auROCc.

Finally, the same PCA-LDA methodology used for the analysis Raman spectra was applied to the reflectance dataset. The first 15 PCs were calculated through standard PCA, then 4 of them (PC1 to PC3 and PC5) were used to generate a classification model. The PCA-LDA classifier [reported, after denoising, in Fig. 3(g)] reached 93% auROCc [Table 1, Figs. 3(h) and 3(i)] when training the algorithm on all dataset spectra; a 4f CV resulted in almost the same accuracy (92%). Moreover, we evaluated the correlation between our model and tumor diameter [Fig. 4(d)] or EGFP intensity [Fig. 4(e)]. Using a linear fit, a positive correlation was observed in both cases (Pearson’s $r=0.459$ and 0.552, respectively); however, the regression line did not result in good fitting of the data ($R^2=0.211$ and 0.305). Thus, the classifier based on reflectance spectroscopy seems capable of detecting the tumor and potentially sensitive to its diameter, although more experiments and training of the model would be required to quantitatively evaluate its size.

3.2.4.

Multimodal results

We reached high sensitivity (92%) in classifying Raman spectra and high specificity (100%) for reflectance ones. Since Raman and reflectance spectroscopies provide complementary information, combining their results could improve tissue characterization. Thus, the scores resulting from their classification models were linearly combined through the same LDA routine in order to maximize discrimination between normal and tumor areas. We found a sensible increase in auROCc value (97% from the whole dataset and 95% from the cross-validation), while the suggested threshold corresponded to 92% sensitivity and specificity.

Tumor multimodal scores were also correlated to GBM size and EGFP signal, but the results from both linear fits were worse than the corresponding analyses on reflectance spectra. Such outcome reflected the fact that Raman scores were totally uncorrelated to tumor dimensions, hence their contribution to multimodal scoring reduced the correlation observed from reflectance data.