1.

Introduction

Prostate cancer is one of the most common types of cancer that affects men. The National Cancer Institute estimates that $\sim 11.2\%$ of men will be diagnosed with prostate cancer at some point during their lifetime and 4.8% of all cancer deaths in 2018 will be due to prostate cancer in the United States.¹ Many men with prostate cancer may be asymptomatic and, without screening, would never know they have the disease. The main purpose of screening for prostate cancer is to identify high-risk localized prostate cancer that can be successfully treated, thereby preventing the morbidity and mortality associated with advanced or metastatic prostate cancer. Nevertheless, the harms of the screening test and subsequent harms from diagnosis and treatment must be weighed. The histopathological analysis has an important role in the management decisions. The current standard reference in prostate cancer diagnosis is the histopathological Gleason grading system,² the accuracy of which depends on the experience and interpretation skill of the pathologist. The Gleason score (GS) system was reviewed in 2014 by the International Society of Urological Pathology (ISUP), aiming to improve the correlation between the histopathological grading system and the prognostic grade groups.³ Therefore, straightforward techniques are continually being developed for helping the assessment of the aggressiveness and the stage of the disease.

Raman spectroscopy is a technique capable of providing chemical information about a biologic tissue sample, and is gaining wide recognition as an objective method for the diagnosis of diseases in tissues.^4–7 It has been successfully tested in a variety of diseases in organs, such as breast,^8–10 brain,^11,12 gastrointestinal,^13–15 and skin cancers.^16–18 Furthermore, Raman spectroscopy has also been used to study prostatic adenocarcinoma, correlating Raman spectrum and the GS system.^19–22 These studies have been performed using near-infrared laser sources, such as 785 and 830 nm, aiming to improve Raman scattering intensity. However, at these wavelengths, biological tissues and common glass sampling exhibit strong fluorescence.

In the context of examining prostate malignancies, Patel and Martin²³ used Raman spectroscopy to characterize the transitional, central, and peripheral zones of normal prostates, revealing larger concentrations of DNA and RNA in the peripheral zone, as well as differences in the relative concentration of lipids and proteins between the three zones. Crow et al.²⁰ and Stone et al.¹⁹ carried out a series of studies on prostatic tissue using Raman spectroscopy. The first publication of normal prostate spectra dates from 2002.²⁴ In 2003, researchers showed the ability to differentiate prostate cancer into three categories ($\text{Gleason}<7$, $\text{Gleason}=7$, and $\text{Gleason}>7$) and from benign prostatic hyperplasia with 89% accuracy from in vitro analysis of frozen biopsies.^19,20

In 2005, the possibility of using fiber optics to perform the examination of urogenital tissues in vitro was demonstrated, opening a perspective for applying the technique in vivo during endoscopic procedures and biopsies.²¹ Statistical regression methods were used to correlate the scattering spectra obtained from the histological samples with the combination of the spectra of several pure biochemical compounds. An increase in the relative concentration of DNA, choline, and cholesterol was shown to be associated with malignancy.^19,25–27

While the molecular specificity of Raman spectroscopy and its ability to make in vivo measurements should significantly reduce the number of unnecessary endoscopic biopsies and prevent biopsy-related costs, it is still to be incorporated for such clinical prostate procedures. A limiting factor of prior research efforts has been the presence of significant background and the associated noise contributions. The use of 1064-nm dispersive Raman spectroscopy provides a viable alternative that reduces intrinsic fluorescence and minimizes tissue’s photochemical changes.^28,29 The signal-to-noise ratio is improved using 1064-nm wavelength instead of 785 nm.³⁰ Thermal or photochemical changes are also reduced when using photons with lower energy, which are not mutagenic like the photons in the ultraviolet range that induces the formation of thymine dimers in DNA. However, the spectroscopic measurement of prostate tissue chemistry using a 1064-nm laser for histopathology is largely unexplored.

This proof-of-concept study aims to propose and investigate the use of 1064-nm laser Raman spectroscopy in the characterization of prostate cancer. We hypothesize that a higher wavelength laser is able to provide significant different spectra between benign and malign sample while keeping the fluorescence effect low.

2.

Methods and Materials

2.3.

Signal Acquisition

To perform the scattering measurements, the specimens were positioned in the central region of the equipment between two glass coverslips (Fig. 1). The use of the 1064-nm laser reduces the intensity of the Raman scattering so that a power of 300 mW was applied on the samples, a value 100 times higher than the power used with the 785-nm laser. Only two measurements were performed on each fragment as a larger number of measurements would increase the temperature and biochemically damage the sample. All scatter data collection was blind, without the knowledge of the diagnosis contained in each sample.

Fig. 1

Experimental setup with specimen positioning.

A total of 74 scans were obtained using a spectral resolution of $0.4{\mathrm{cm}}^{-1}$ and $\sim 6\mu \mathrm{m}$ in diameter at the focus of the laser beam. These parameters were adjusted in agreement with related published works using the 1064-nm laser in biological samples as no published studies using this wavelength for the Raman spectroscopy of prostate samples could be found. A Bruker^® VERTEX 70 FT-Raman spectrometer was used with the 1064-nm Nd:YAG laser as the excitation source. The calibration of the spectrometer followed three steps: the calibration with known lines of neon lamp, with silicon peak at $520{\mathrm{cm}}^{-1}$, and with known peaks of carbon nanotube and single-layer graphene. The Raman signal was collected by a germanium detector cooled by liquid nitrogen. The system uses a ROCKSOLID interferometer with the following acquisition parameters: spreading geometry of 180 deg, one point per the sample, and approximate sampling time of 60 s.

3.

Results

The spectra of the saline solution, formaldehyde, and glass coverslips used in the preparation of the samples are shown in Fig. 2. The Raman spectroscopy of the coverslips produced a broad peak in the region of $1100{\mathrm{cm}}^{-1}$, which could potentially impair the identification of peaks generated by the fragments in this region. Therefore, the same coverslips were used in the experiments so that their influence would be minimized. The spectral analysis of the saline solution produced two broad scattering peaks in the region of interest, the first close to $800{\mathrm{cm}}^{-1}$ and the other at $1650{\mathrm{cm}}^{-1}$. The peak at $800{\mathrm{cm}}^{-1}$ is practically nonexistent in the scattering spectra of the fragments, whereas the $1650\text{-}{\mathrm{cm}}^{-1}$ peak coincides with a peak in the fragments’ spectra. A peak at $1100{\mathrm{cm}}^{-1}$ was observed in the spectrum of the formaldehyde. As it was diluted in water and the peak at $800{\mathrm{cm}}^{-1}$ characteristic of the saline solution was not observed in the fragments, this suggests that the dilution process was sufficient to significantly reduce the presence of the formalin in the fragments.

Fig. 2

Average filtered spectra of the physiological solution (top), formaldehyde (middle), and glass coverslips (bottom). The spectra are offset manually for visual clarity. The fingerprint region from 700 to $1800{\mathrm{cm}}^{-1}$ is shown in light gray.

Figure 3 shows the mean spectra of the benign and malignant groups after noise filtering, background removal, normalization, and baseline correction. As can be seen, the average spectrum of the malignant sample, shown in gray, presents higher values in almost all the fingerprint range. The local minimum at $800{\mathrm{cm}}^{-1}$ is a consequence of the subtraction of the saline solution spectrum, showing that the presence of preparation materials in the fragments was minimized.

Fig. 3

Average spectra of the benign (gray) and malignant (black) samples. The fingerprint region from 700 to $1800{\mathrm{cm}}^{-1}$ is shown in light gray.

Figure 4 shows the average spectra for both samples in the fingerprint region. The Wilcoxon test revealed significantly different signal magnitudes in the range from 980 to $1491{\mathrm{cm}}^{-1}$ (266 measurements), which encompasses eight shift points that have been identified in the literature as characteristic of prostate cancer (A to H). These points are shown in Table 1, with corresponding $p$-values and effect sizes.

Fig. 4

Average spectra of the benign (gray) and malignant (black) samples in the fingerprint region. The most significant peaks are shown in light gray with key points labeled from A to H (see Table 1).

Table 1

Results of the Wilcoxon test showing the significantly different regions in the range of 700 to 1800  cm−1, labeled from A to H (see Fig. 4). References of concordant works are given in the first column. The vibration and molecular description is extracted from Kast et al.36

Machine learning techniques were used to design a classifier able to discriminate between benign and malignant samples. A binary SVM was trained using a Gaussian kernel with $\text{gamma}=0.25$ and $\text{cost}=1$. These values were determined with the aid of a grid searching process over the parameter space, performed by the tune.svm function of the e1071 package. The classifier used the eight measurements described in Table 1, taken at the Raman shifts of 1003, 1062, 1096, 1125, 1269, 1317, 1338, and $1447{\mathrm{cm}}^{-1}$. A leave-one-out cross-validation testing procedure resulted in 96% of accuracy. From the 74 samples, only three benign samples were misclassified as malignant.

A second classifier based on SVMs was designed to predict the GS of the samples. In this case, six binary SVMs using Gaussian kernel, $\text{gamma}=0.25$, and $\text{cost}=1$ were trained to classify between each pair of classes taken from benign, $\mathrm{GS}=6$, $\mathrm{GS}=7$, and $\mathrm{GS}=8$. A voting strategy was used to determine the winner class, i.e., the sample was assigned to the class that was chosen by the largest number of SVMs. The classifier obtained an overall accuracy of 95%. Table 2 shows the confusion matrix for the tests, using leave-one-out cross validation.

Table 2

Confusion matrix and performance metrics achieved by the GS predictor.

4.

Discussion

The use of 1064-nm laser is still underexplored as an excitatory source in the analysis of prostatic tissue. In fact, studies based on Raman spectroscopy for the diagnosis of prostate cancer favor the use of lower wavelengths: 532,³⁹ 633,^37,38 785,^21–23 830,¹⁹ and 832 nm.²⁰ The surface enhanced Raman spectroscopy (SERS) has also been applied to the analysis of blood serum to differentiate between prostate cancer and benign prostatic hyperplasia, using a 633-nm He–Ne laser.^40,41 However, studies based on the coherent anti-Stokes Raman spectroscopy (CARS) technique already adopt the 1064-nm wavelength as the Stokes beam with accurate results.⁴²

The range from 700 to $1800{\mathrm{cm}}^{-1}$ is the region of the spectrum investigated by the majority of the published Raman spectroscopy applications on prostate cancer. The most evident peaks are found in $1000{\mathrm{cm}}^{-1}$ (phenylalanine ring breathing), 1200 to $1350{\mathrm{cm}}^{-1}$ (overlapping of the amide III, $─{\mathrm{CH}}_2─$ “wagging” and $─\mathrm{CN}─$ “stretch” peaks present mainly in lipids and nucleic acids), $1450{\mathrm{cm}}^{-1}$ ($─{\mathrm{CH}}_2$ deformation vibration present in nucleic acids, proteins, and lipids), and in the region of 1600 to $1700{\mathrm{cm}}^{-1}$ ($\mathrm{C}═\mathrm{O}$ stretch, amide I, present in nucleic acids and proteins). These peaks have been detected mostly by studies that used 633- and 785-nm dispersive systems in prostatic parenchyma, as reviewed by Kast et al.³⁶

In the experiments performed with the 1064-nm laser, significant differences were found in the 1000 to $1450{\mathrm{cm}}^{-1}$ range, which are in accordance with previous reported results associated to lipids, nucleic acids, and proteins, among other biomolecules, as described in Table 1. Nevertheless, the experiments were not able to detect differences in the regions of 719 to $729{\mathrm{cm}}^{-1}$ and 781 to $786{\mathrm{cm}}^{-1}$, associated with DNA, RNA, and phospholipids, which were detected by other studies.^21–23,37 In addition, we could not reproduce the findings of Crow et al.,²¹ Taleb et al.,³⁷ and Patel and Martin,²³ which reported differences in the range of 1657 to $1667{\mathrm{cm}}^{-1}$ associated with amide I.

The potential of the Raman spectroscopy to discriminate between benign and malignant samples as well as to predict the stage of the disease has been investigated since 2003 when Crow et al.²⁰ applied linear discriminant analysis to classify a set of spectra from prostate biopsies. The study achieved 89% of accuracy for the Gleason prediction and 92% of accuracy for the classification into benign and malignant classes. Taleb et al.³⁷ examined a set of 30 cells using partial least-squares discriminant analysis and were able to correctly classify the sample into malignant and benign with 100% of accuracy. Discriminant function analysis associated with principal component analysis (PCA) was used by Devpura et al.²² to classify between cancerous and benign tissue samples with 91% of accuracy. PCA and SVM were used by Wang et al.³⁸ to classify the spectra of 50 patients into two groups according to their GS ($<=7$ and $>7$), achieving 88% of sensitivity, specificity, and accuracy.

In the experiments using a 1064-nm excitatory source, we were able to classify the prostate samples into malignant and benign with 96% of accuracy, and to predict their GSs with 95% of accuracy. The choice for a machine-learning method such as the SVM incorporates the ability to discriminate nonlinearly separable classes that are not characteristic of other multivariate analysis such as the LDA. The specificity of the method was consistently high, with average of 98%. The sensitivity varied from 67% to 100%, with an average of 89%. The lowest sensitivity of 67% was obtained for the $\mathrm{GS}=8$ class, due to the fact that this was a reduced subset of six spectra. Nevertheless, the high accuracy of the results is encouraging and should motivate a deeper investigation on alternative wavelengths for the Raman spectroscopy of neoplastic tissues.

5.

Conclusion

In this study, we have proposed and evaluated the application of the 1064-nm laser in Raman spectroscopy as an alternative to the 785-nm modality. In a set of experiments using benign and malignant fragments of the prostate parenchyma, the method was able to reveal differences in the spectra, opening a perspective for its use in cancer diagnosis. Raman spectroscopy using the 1064-nm excitatory laser allows for the evaluation of the scattering from different molecular compounds present in the prostatic tissue, in addition reducing the effect of fluorescence.

Subsequent experiments with a larger sample cohort will be pursued to confirm the method’s ability to discriminate between groups and to provide the accurate GS for a given specimen. A larger sample set and repeated scans should increase the signal-to-noise rate as well as the statistical power of the analysis. In addition to label-free Raman measurements, the use of SERS probes may permit precise visualization of tumor margins in vivo along with the characterization of microscopic tumor invasion. A more advanced perspective of application in prostatic tissue analysis would be the use of the CARS technique that requires a shorter signal integration time and could increase the sensitivity of the method. Ultimately, a 2-D mapping of histopathological sections should be output helping the pathologist to more accurately and objectively determine the prognostic groups of the patients from quantitative information of the biochemical composition of the prostatic tissue.