2.1.

Groups under Study

The research was carried out according to the principles of good clinical practice. Protocol of the research was approved by the Ethics Committee of the Siberian State Medical University (Tomsk, Russia), Ref. Number 4532 on July 07, 2015. All participants were informed about the details of the research and signed the Informed Agreement. The interaction with the patients was limited to the sampling by a saliva probe placed into a disposable container and the scraping of cell material by a sterile dental metal spatula from the mucous membrane of the cheek for cytological analysis.

The study excluded patients with lichenoid reactions, drug reactions, and systemic lupus erythematosus; patients with restorations of amalgam and acute inflammatory diseases of oral cavity (abscesses, pericoronitis, periodontitis); and patients who were receiving drugs that affect immune response (including oral contraceptives, corticosteroids, retinoids, tacrolimus, cyclosporine, and azathioprine) at the time of the study and for 6 months prior, as well as pregnant or nursing women.

The study participants involved 30 patients (2 males and 28 females) with clinically and histologically proven OLP. This group included two subgroups with the erosive form of OLP ($n=15$) and with the reticular and papular forms of OLP ($n=15$). In all cases, the diagnosis of OLP was confirmed clinically and by histopathological examination.

The control group consisted of six healthy volunteers (one male and five females) without inflammation in the mucous membrane in the oral cavity and without periodontitis.

2.2.

Sampling Protocol

For 12 h before the sampling, the participants ingested no meals or medications; alcohol and tobacco were also excluded. On the morning of the saliva sampling, toothbrushing, ingestion of food, liquids (including water), or medications, smoking, and chewing gum were excluded. An unstimulated saliva probe was taken on an empty stomach between 8 am and 9:30 am. For the free flow of saliva, the participant tilted his or her chin down to the chest and opened his or her mouth; saliva dripped into the lower lip of a cuvette over 5 min. The cuvette was tightly closed with a cap. The samples were analyzed within 1 h after collection.

Additionally, for every participant, we produced swabs by scraping the mucous membrane of the cheeks with a sterile dental metal spatula; the scraping was then placed on a glass slide and dried for 1 h, fixed in 96% ethanol, and washed in water for 10 min. Smears were stained with Romanovsky–Giemsa stain.

Multiple sampling was used to provide statistical analysis.

2.3.

Experimental Base

The stained slides were analyzed by the optical microscope Carl Zeiss Axio Scope (Germany). The scrapings were assessed by the degree of destruction of the epithelial cells. The epithelial cells were divided into the following types: the zero type of cells had a normal nucleus and cytoplasm structure; the first type of cells had no more than 50% damage to the cytoplasm and normal structure of the nucleus; the second type of cells had more than 50% damage to the cytoplasm and partial damage to the nucleus; the third type of cells had complete damage to the cytoplasm but not complete damage to the nucleus; and the fourth type of cells had complete damage and disintegration of both the cytoplasm and the nucleus.

For every slide, 100 epithelial cells were selected and the nuclear/cytoplasmic ratio (N/C) was calculated and each stage of differentiation of epithelial cells was defined according to N/C value. When the N/C was 0.50 to 0.49, a cell was considered at the first stage of differentiation; when N/C was 0.40 to 0.49, a cell was considered in the second stage; when N/C was 0.30 to 0.39, a cell was in the third stage; when N/C was 0.20 to 0.29, a cell was in the fourth stage; and when N/C was 0.10 to 0.19, a cell was in the fifth stage of differentiation. In the absence of the nucleus ($\mathrm{N}/\mathrm{C}=0$), a cell was attributed to the sixth stage of differentiation (Fig. 1). Then, the index of cellular differentiation (ICD) was calculated according to the formula:

Fig. 1

Various stages of differentiation of epithelial cells: (a) a cell considered to be at the first stage of differentiation; (b) a cell in the second stage; (c) a cell in the third stage; (d) a cell in the fourth stage; (e) a cell in the fifth stage; and (f) a cell in the sixth stage.

Here, the numbers 1 to 6 denote the differentiation stages and also are weighting factors of the stages. The latin letters “a,” “b,” “c, “d,” “e,” “f” denote the number of cells corresponding to the specific stage of differentiation.

Callimeri and Smith⁴² found that increase in the N/C was an indicator of the transformation from a benign cell to a malignant one. Cowpe et al.²⁵ showed that decrease in cytoplasmic diameter, and hence cytoplasmic perimeter and CA, was an early indicator of malignant cell transformation. They found that decrease in the cell diameter occurred in cell dysplasia.

The spectral analysis of saliva samples was produced by the time-domain THz spectrometer T-SPEC (EKSPLA, Estonia). The main characteristics of this spectrometer are a spectral range of 0.3 to 3.5 THz, a dynamic range of up to 90 dB, and spectral resolution of at least 2.3 GHz. To increase the signal-to-clutter ratio, the average of 1024 scans was used.

Special cuvettes were designed for saliva sampling (Fig. 2). The cuvette had a part for measurements (“b” in Fig. 2), which provided a saliva layer thickness of 1 mm, because a thicker layer of saliva almost totally absorbs the THz radiation. The bottom part “c” of the cuvette was designed to produce the reference signal. Tests have shown that the cuvettes are transparent enough in the THz range (Fig. 3). Figure 3(a) shows the intensity of output THz signal when the radiation passed through room air without the cuvette (Ia) and when the radiation passed through the cuvette (Ip) in the point “X” (Fig. 2).

Fig. 2

The cuvette designed for saliva sampling.

Fig. 3

The intensity of output signals for when the THz radiation passed through room air without the cuvette (Ia) and when THz radiation passed through both (a) room air and the cuvette at the point “X” (Ip) and (b) the ratio of these signals.

The measurements of output signal at five points on the cuvette surface were performed. This procedure was repeated for 20 cuvette copies.

2.4.

Data Preprocessing and Analysis

The THz absorption spectra of saliva samples were considered as feature vectors of the participant’s state. Principal component analysis (PCA) was used to select informative features in the initial feature vectors. The basic idea of PCA is to find a reduced number of new variables that are adequate for recovery of the initial variables, possibly with insignificant errors. The basic idea of PCA consists in projection of correlate variables into a lower number of uncorrelated variables called principal components (PCs).⁴³

The supervised support vector machine (SVM) binary classification method with Gaussian radial basis function kernel $K(x,y)=\exp (-\frac{⟨(x-y),(x-y)⟩}{2{\sigma }^2})$ or polynomial kernel $K(x,y)={(x^{\mathrm{T}}y+1)}^{\rho }$ was used for dichotomous classification of the groups under study.⁴⁴ Here, $x,y$ are the feature vectors, σ is the optimization parameter, T is the operation of matrix transposition, and $\rho$ is the degree of the polynomial.

There are several approaches to constructing a multiclass classifier using binary classifiers.⁴⁵ According to the one-versus-all method, we had to construct $N$ independent binary classifiers, so that every classifier would separate a specific class’s feature vectors from all other classes’ feature vectors.⁴⁶ According to the one-versus-one (also known as all-versus-all) method, we had to construct $N(N-1)$ independent binary classifiers, each of which would separate $i$’th class feature vectors from $j$’th class feature vectors.⁴⁷ The latter method was shown to provide the better results.⁴⁸

Nonparametric Mann–Whitney $U$ test was also used, and $p<0.05$ was considered statistically significant.

3.1.

Cytological Analysis

The cytological analysis was carried out as follows. The number of participants in the control group, the subgroup of patients with the erosive form of OLP, and the subgroup of patients with reticular and papular forms of OLP were denoted as $N_1,N_2,N_3$, respectively. Then, the indexes $j_1=\mathrm{1,2},\dots ,N_1$, $j_2=\mathrm{1,2},\dots ,N_2$, and $j_3=\mathrm{1,2},\dots ,N_3$ denoted participants in corresponding group.

A swab was taken from every participant, then the stained slide was prepared, from which the random statistical sampling of 100 cells was analyzed. The cells in a sample were marked by $\alpha =\mathrm{0,1},\dots ,4$ depending on type. Then, the number of cells of type $\alpha$ in a swab from patient $j_1$ from the first group was denoted as $n_{j_1\alpha }^{(1)}$, the same for the second group was $n_{j_2\alpha }^{(2)}$, etc.

The results of cytological analysis are presented in Table 2 in terms of the total number of cells of a specific type in the specific group. For example, $A_{1\alpha }=\sum _{j_1=1}^{N_1}{n}_{j_1\alpha }^{(1)}$ is the total number of cells of type $\alpha$ in the first group. Its relative value (in %) is $\frac{A_{1\alpha }}{N_1\times 100}\times 100\%$. In Table 2, $p_{ij}$ is the significant level of difference between average characteristics of $i$’th and $j$’th groups according to ${\chi }^2$ Pearson criterion.

According to cytological studies, the number of epithelial cells of the 0 type of destruction in the control group was 61.2%, whereas for patients with erosive, reticular, and papular forms of OLP, this parameter was decreased significantly to 34.7% and 54.0%, respectively ($p<0.05$) (Table 1). For the erosive form of OLP, zero type and the fourth type of epithelial cells were observed almost equally (34.7% and 31.5%, respectively). Despite the fact that the zero type and the first type cells predominated for the group with reticular and papular forms of OLP, content of these cells was significantly lower in comparison with the control ($p<0.05$).

Table 1

Degree of destruction of buccal epithelial cells in the groups under study.

The mean value of the ICD was $448.4\pm 5.7$ in the control group and was $397.4\pm 7.7$ and $438.6\pm 8.3$ in the group with the erosive form of OLP and in the group with the reticular and papular forms of OLP, respectively. Thus, the mean value of the ICD for the reticular and papular forms group showed no statistically significant distinction compared to the control group.

3.2.

Spectral Analysis of Saliva Samples

The shapes of the averaged absorption spectra of the saliva samples for patients with erosive OLP, reticular and papular OLP, and healthy volunteers processed by the median filter are shown in Fig. 4. The log-ratio of the output intensities $\log \frac{I_{\mathrm{c}}}{I_{\mathrm{b}}}$ was used to exclude the influence of the cuvette. Here, $I_{\mathrm{b}}$ and $I_{\mathrm{c}}$ are the output intensities of THz radiation that passed through parts “b” and “c” of the cuvette, respectively.

Fig. 4

The dependence of the absorption of THz radiation in terms of $\log \frac{{\mathrm{I}}_c}{{\mathrm{I}}_b}$ on the frequency for (c) the control group, (a) patients with erosive OLP, and (b) reticular and papular OLP. The confidence intervals are shown. Here, $I_{\mathrm{b}}$ and $I_{\mathrm{c}}$ are the output intensities of THz radiation that passed through parts “b” and “c” of the cuvette, respectively (see Fig. 2).

The quality of spatial separation of the groups under study in the PCs’ space strongly depends on the quantity of the PCs used. We took into account the first eight PCs. The example of the projection of the objects under study on the plane of the second and third PCs is presented in Fig. 5. Evidently, the entire healthy control group is spatially separated from the groups of OLP patients.

Fig. 5

The projection of the objects under study on the plane of the second and third PCs.

To construct a multiclass classifier, we used a set of one-versus-one binary SVM classifiers. The selection of the training subset from all groups under study was repeated randomly many times. The remainder of the initial data was used to test the classifiers. The ratio of separation of initial data on training and testing subsets varied from 0.5/0.5 to 0.85/0.15. A value of the training subset close to 25% from initial data was shown to be optimal.

At first, we constructed the following binary SVM classifiers: patients with the erosive form of OLP versus patients with reticular and papular forms of OLP; patients with the erosive form of OLP versus healthy volunteers; and patients with reticular and papular forms of OLP versus healthy volunteers. The results of testing of the classifiers are presented in Table 2 in the terms of sensitivity and specificity:

Here, TP is the quantity of true-positive classification results and FN is the quantity of false-negative classification results. The results of dichotomous SVM classification with Gaussian radial basis function kernel of the saliva absorption spectra samples for the groups under study are presented in Table 2.

Table 2

SVM classification of the testing set of saliva absorption spectra samples for the groups under study.

The results of differential (multiclass) diagnosis based on saliva absorption spectra samples analysis and three one-versus-one classifiers from Table 2 are presented in Table 3. The evaluations were carried out using a merged testing set that included patients with erosive, reticular, and papular forms of OLP and healthy volunteers, as shown in Table 3. The feature vector of a representative from the testing set was analyzed by every classifier from Table 2. The differential diagnosis rule was based on the result that was selected in more than 50% of outcomes. The absorption spectrum of every sample was measured four times. Thus, differential classification, based on using a single scan for a specific sample or a single sample (using four scans for a specific sample), was carried out.

Table 3

Differential diagnosis based on the set of SVM classifiers from Table 2.

In the case of using a single sample, the diagnosis considered to be determined if it coincided for more than two spectra scans for the specific sample under study. This approach in fact represents two-stage classification. An analog of this approach is classification by the “majority vote,” when a solution is based on using several classifiers, because using multiple feature vectors for classification of the same object supposes using multiple classifiers. The classification by the majority vote allows increasing the robustness of the decision rules created because of using multiple descriptors for the same object.^49,50

The results of differential classification based on using saliva samples’ absorption spectra scans were shown to be high. The two-stage classification approach based on using several scans for a specific sample provided 100% accuracy in differential classification of the collected data.