2.1.

Sample Collection and Handling

Biopsied normal, premalignant, and malignant oral tissue samples were obtained from the college of Dental Surgery and Sai Baba Cancer Hospital, KMC Manipal Academy of Higher Education (MAHE), India. Tissues from uninvolved areas from the same subjects were used as healthy controls. A mirror image of each sample was fixed in 10% neutral buffered formalin and was sent for histopathological certification. The samples were kept moist with saline $(\mathrm{pH}=7.4)$ , and spectra were recorded within half an hour of tissue removal. Samples were mounted on a quartz plate and fluorescence measurements were carried out with 325-nm laser excitation. Whenever necessary, the tissue samples were snap frozen in liquid nitrogen and stored at $-80°\mathrm{C}$ for later analysis. Trial runs showed that the spectra remain unchanged at least for $2\mathrm{h}$ after biopsy, if kept in saline.¹⁸ For each sample, spectra were recorded from several points, separated by few hundred micrometer to few millimeters, and were treated as independent data because of the inhomogeneous nature of the tissue, and in addition, to explore the feasibility in surgical boundary demarcation. Details of the sample preparation and spectral recording are mentioned elsewhere. ^{18, 19, 20, 21} No particular care was taken regarding the orientation of the tissue under study. Biopsied oral tissue subjects used in this study were of different sizes, having their surface area (approximately) in the range of 15 to $25{\mathrm{mm}}^2$ and of thickness in the range 2 to $5\mathrm{mm}$ . The sample details are given in Table 1 .

Table 1

Sample details.

2.2.

Experimental Setup

A block diagram of the experimental setup for recording the fluorescence spectra is shown in the Fig. 1 . The excitation used in this study is $325\mathrm{nm}$ at a 10-Hz repetition rate with energy per pulse of 100 to $200\mathrm{\mu }\mathrm{J}$ from an $\mathrm{Nd}\text{-}\mathrm{YAG}∕\mathrm{MOPO}∕\mathrm{FDO}$ source (Spectra Physics, Quanta Ray, Model: PRO 230 10, MOPO SL). The fluorescence emission from the tissue samples was collected by fiber optic probe, details of which are given elsewhere.¹⁸ In brief, we used a typical homemade seven-fiber probe with a central fiber for excitation and the surrounding six fibers for collection of fluorescence. The individual fibers (Thorlabs FG 200-UEP Fiber type) were of $240\mathrm{\mu }\mathrm{m}$ diameter with a numerical aperture of 0.25. The excitation end of the fiber was held perpendicular to the tissue surface under study and the distance between the tip of the fiber and the tissue surface was maintained to 2 to $3\mathrm{mm}$ , as optimized for better fluorescence intensity. To adjust the distance between the tissue surface and the fiber tip and to position the excitation at different sites of a tissue sample, the excitation end of the fiber was mounted on an $XYZ$ micrometer translational stage. The output of the fiber was coupled into an imaging spectrograph (Spectra Pro 150 spectrograph $300\mathrm{g}∕\mathrm{mm}$ , 300-nm blazed grating) equipped with an Andor intensified CCD (ICCD) system (Andor Technology, Northern Ireland, ICCDBH-501-25F-01) for spectral recording. A time delay of $500\mathrm{ns}$ provided by a DG 535 delay and gate generator (Stanford Research Systems) was used to compensate for the cable and other delay times. The total fluorescence after the delay was recorded with an optimized 75-ns gate.¹⁸ All spectra were recorded with a slit width of $250\mathrm{\mu }\mathrm{m}$ (5-nm bandpass) and with an average of 100 laser pulses $\left(10\mathrm{Hz}\right)$ .

Fig. 1

Schematic of the experimental setup for LIF measurement.

2.3.

Data Analysis

In this study we used 143 fluorescence spectra (60 normal, 26 premalignant, and 57 malignant) from 19 normal, 5 premalignant, and 19 malignant oral tissue samples. These spectral samples are the same samples, as are used in Ref. 27 for PCA and ANN analysis. In this study, these 143 spectra are treated as independent data for the following reasons. First, in epithelial cancers (before it has penetrated the underlying tissue), often a tissue specimen can have normal and neoplastic regions adjacent to each other. ^{30, 31, 37, 38} The pathologist, therefore, examines several sites on a biopsy sample and even if only one site shows as abnormal, the sample is considered as abnormal. Second, for in vivo applications in screening, surgical boundary demarcation, and optically guided biopsy it is necessary to identify the exact site where malignancy is suspected. The variations in the spectra from site to site will be the deciding factor in this case and they must be treated independently. Third, in epithelial cancers, the abnormal proliferation begins near the basal lamina (dysplasia), which then advances into the epithelial region (carcinoma in situ), and later, in the invasive stage, penetrates into the connective tissue region. We have shown³⁸ by Raman spectroscopy that in biopsied tissue samples of oral cancer, the subepithelial region showed very similar spectra for normal and malignant samples, and only the epithelial spectra discriminated between the two very well for many samples. This means that unless one is sure of the region (epithelial, subepithelial, or connective tissue) one is looking at in a biopsy sample, there can be both normal and malignant types of spectra from adjacent sections of the same sample. This possibility is further supported by the result that in none of the spectra of normal samples (oral, breast, cervix, and ovary systems that we have studied) recorded from different sites of the same specimen, did we observe any anomaly, since normal samples are normal at all sites. Only malignant samples have shown this anomalous behavior,³⁹ with some sites giving typical normal spectra and other sites of the same sample showing typical malignant spectra, since these samples have the possibility for both types of sites existing together. In view of these points, when there is no a priori information available for a sample it is necessary to examine many sites on the sample and treat each site as independent. This is necessary even when one is guided by a visual examination by the clinician for biopsy because of possible errors such as past pointing. ^{37, 38, 39, 40, 41, 42, 43, 44} In this study, we used MATLAB@R6-based PCA and ( $k$ -NN) analysis to discriminate among normal, premalignant, and malignant samples.

The PCA method is a linear feature transformation technique that preserves the variance of the feature space. This is a classical method, usually employed for the dimensionality reduction of a feature space.⁴³ When many of the variables in a multidimensional data set have a significant correlation, PCA successfully captures nearly all the information of the original data set in the first few principal components (PCs). PCA mainly has three effects. First, it transforms the data by defining new variables, termed PCs, which are statistically uncorrelated. Second, it orders the resulting PCs in such a way that those possessing larger variances are arranged first. And third, in many problems, the first few components capture most of the discrimination capability. In such cases, only the first few components must be considered.

PCA is applicable for data analysis problems in several domains where the data elements are significantly correlated. In PCA, the PCs are arranged in the order of their contribution to the variance of the entire spectral data set. The first PC accounts for as much of the variability in the data set as possible, and each succeeding component accounts for the next largest amount of variation and is independent of the previous PC. Each new PC is orthogonal to the previous PC and contains the maximum information unexplained by the previous one. Thus, PCA processing can dimensionally reduce the variables of the original spectral data into a small number of informative PCs that fully describe the variations in the spectral data within the limitations of noise. For doing PCA, the six features mean, median, standard deviation, energy, maximum intensity, and the spectral residuals are extracted from each spectrum and MATLAB@R6 (Refs. 45, 46) based PCA is performed on the reference database. Applying PCA to the feature space matrix, the original data are transformed into a set of PC scores. Since most of PCs account mainly for noise and do not provide diagnostic information, it is necessary to identify a small group of informative PCs to build the algorithms for classification of oral spectra of different conditions (normal, malignant, and premalignant). Including a minimal number of PCs reduces the complexity of a diagnostic algorithm. The contribution of each PC to the total variance of spectral data is proportional to its eigenvalue. High-order PCs often account for less than 1% of the total variation and represent mostly noise⁴⁴ only. In this paper, PCs that have a variation of more than 1% variance in the spectral data are considered to be informative PCs.

In PCA, first the calibration standards for each class of the samples (normal, premalignant, and malignant) are prepared using pathologically certified normal, premalignant, and malignant, samples. In this case, we randomly chose 20 spectra each from the normal, premalignant, and malignant groups of spectra and the calibration sets were prepared. Once the calibration standards were established, any new sample could be classified as normal, premalignant, or malignant by comparing it with the standard models for normal, premalignant, and malignant obtained from pathologically certified samples.

The standard calibration sets were optimized by removing outliers from the model set using cluster analysis and selecting an optimum number of factors (PCs) required for the dimensional data analysis.^{47, 48} Including outlier samples in the training set introduces a bias to the final model.^{47, 48} The cluster plot for detecting outliers in the calibration model is generally plotted between the PC scores of the first two PCs because the first two PCs represent the plane of best fit through the data.

After the PCs with diagnostic information were identified, a $k$ -NN method was used to build the algorithm for classification of the oral tissue spectra of different pathological conditions (normal, malignant, and premalignant). To construct algorithms for $k$ -NN classification, only those PCs that have significantly different projection scores for normal, malignant, and premalignant tissues were selected. The spectral samples used in $k$ -NN analysis are the same set of normal, pre-malignant and malignant samples as used in PCA.

Nearest neighbor methods provide an important data classification tool for recognizing object classes in pattern recognition domains. ^{34, 35, 36, 49, 50} This method classifies an unknown sample to that class having most “similar” or “nearest” sample point in the training set of data. The nearest sample was found by using concept of distances or metrics known as Euclidean distance. Euclidean distance is the square root of sum of the square of the $x$ distance plus the square of the $y$ distance for a given pair of cases $x$ and $y$ . Once the Euclidean distance is identified, the $k$ -NN classifies an unknown sample to that class having most similar or nearest sample point in the training set of data. Nearest neighbor methods can detect a single or multiple numbers of nearest neighbors. ^{36, 49, 50, 51} In this study, the single nearest neighbor method was used to classify oral lesions including different pathological conditions (normal, malignant, and premalignant).

3.1.

PCA

In this classification problem, to reduce the dimensionality, we used six features, each from 143 sample spectra (60 calibration + 83 test) on which PCA was performed. Twenty spectra each, randomly chosen from certified normal, premalignant, and malignant tissue samples were combined to discern the best approach to prepare calibration sets in the three classes. Six features of 60 training samples belonging to the normal, premalignant, and malignant groups were first pooled together and PCA was performed on this reference database. Since only a few PCs are required to express the large amount of spectral data, the eigenvalues drop to significant quantities after the first few. The spectra of a given set of samples may have contributions only from a limited number of factors, and except for the first few factors, many of the eigenvalues may be close to zero with no practical contribution to the spectra. A simple plot of eigenvalues can also be used to informally indicate the number of significant components. In our analysis, the eigenvalues, the contribution to total variance, the eigenvectors, and other parameters were used to determine the significant number of factors required for a correct analysis.

Figures 3a and 3b , respectively, show the eigenvalues for the factors and total percentage variance (i.e., total percentage contribution to the variation spectra with increasing number of factors) of 60 standard oral samples (20 normal, 20 malignant, and 20 premalignant). The plots indicate that three factors are sufficient to explain 99.94% of the variance in the total data set. And thus the feature vector of length 6 could be reduced to three components. With many types of data sets, it is common for the first few PCs to possess most of the variance of the original data. Table 2 shows the variance associated with first three PCs. We notice that first PC itself covers over 95.79% of variance. From Table 2, we can see that the eigenvalues decrease very rapidly and are almost zero after three factors. Therefore, in the present analysis, the first three PCA components, which show difference in total percent variance values among normal, premalignant, and malignant cases, were retained for further analysis and remaining higher order PCs were discarded. Thus, the feature space, a $6\times 143$ matrix, was reduced dramatically to a $3\times 143$ matrix, and hence classification became computationally more efficient.

Fig. 3

Plots of calculated (a) eigenvalues and (b) total percent variance of a PCA decomposition of the 60 oral tissue (20 normal, 20 malignant, and 20 premalignant) spectral data.

Table 2

Eigenvalues and percent variance values of a standard set of 60 (20 normal, 20 malignant, and 20 premalignant).

Figure 4 shows a cluster/scatter plot obtained by plotting the scores of the first two PCs of 143 samples [ $(20+40)$ normal, $(20+37)$ malignant, and $(20+06)$ premalignant]. We clearly see from the plot that all samples diagnosed as normal, malignant, and premalignant by pathological examination in the calibration set clustered in three distinct regions. This shows that the calibration set of samples used in this study are almost free of outliers. Also, when all test normal, premalignant, and malignant samples are used, almost all normal and premalignant samples are clustered around their respective calibration set centroid vectors. Though some of the malignant and normal test samples are overlapped, malignant samples are scattered more compared to normal and premalignant samples. Only three test malignant samples overlapped with the normal and premalignant regions. The specificity and sensitivity of this technique is thus found to be quite good.

Fig. 4

Cluster/scatter plot in log mode between the scores of first two PCs of 143 spectra of oral normal, malignant, and premalignant samples: NC, normal calibration; NT, normal test; MC, malignant calibration; MT, malignant test; PMC, premalignant calibration; PMT, premalignant test.

3.2.

$k$ -NN Classification

NN methods can detect a single or multiple numbers of NNs. A single NN method is primarily suited for recognizing data where we have sufficient confidence in the fact that class distributions are nonoverlapping and the features used are discriminatory. In most practical applications, however, the data distributions for various classes are overlapping and more than one NNs are used for majority voting. ^{36, 49, 50, 51, 52} The NN method works well with data where features are statistically independent. From the score plot (Fig. 4), it is evident that the data points of training set of 60 spectra (20 normal, 20 malignant, and 20 premalignant) are clustered in three distinct groups without overlapping. Thus, in this study, we used the single NN method to classify the oral lesions with different pathological conditions (normal, malignant, and premalignant).

3.2.4.

Classification results using $k$ -means NN technique

Once the k-NN was trained with input feature values of 20 normal, 20 malignant, and 20 premalignant samples, the technique was ready to predict any unknown new data. A MATLAB-based program was executed to predict any new data. The classification of test samples was made based on the algorithm already mentioned. In this study, we have features from the same 40 normal, 37 malignant, and 6 premalignant oral tissue spectra as used in PCA. When this classification technique was used on calibration model, all 60 calibration samples (20 normal, 20 malignant, and 20 premalignant) were classified to their respective groups (as clustered in three distinct regions of cluster plot of PC scores). In our analysis using the $k$ -NN technique, all 40 normal test spectra were classified as normal, 34 out of 37 malignant spectra were classified as malignant, and all 6 test premalignant spectra were classified as premalignant. The specificity, sensitivity, and accuracy obtained in this classification technique were 100, 94.5, and 96.17%, respectively.

Figure 5 shows the plot of the Euclidean distance against sample number for the 143 samples ( $20+40$ normal, $20+37$ malignant, and $20+06$ premalignant). In Fig. 5, the “normal centroid” is taken as the reference point for plotting Euclidean distances of all samples. From the plot, it is clear that the Euclidean distances of premalignant samples lie in between the Euclidean distances of normal and malignant samples. The calibration set of samples (normal, premalignant, and malignant) cluster in three distinct groups without any overlap. The test normal and premalignant samples also cluster along with the corresponding calibration set samples, showing 100% discrimination of these samples. Three of the malignant test samples were overlapped with the normal calibration set cluster, and this may be due to from the normal sites of the malignant samples. A closer observation of Fig. 5 shows that only a very small number of malignant samples fall outside the general range of malignant species, indicating the probability of malignant samples being in the respective cluster is about 97% and finding them out of the cluster is only 3%. Thus, the plot clearly shows discrimination of the different classes of oral spectral samples used in this study.

Fig. 5

Plot of the Euclidean distance against sample number for the 143 samples ( $20+40$ normal, $20+37$ malignant, and $20+06$ premalignant): NC, normal calibration; MC, malignant calibration; PC, premalignant calibration; NT, normal test; MT, malignant test; and PT, premalignant test.

Table 3

Performance evaluation of k -NN classification and comparison with performance of other conventional techniques PCA and ANN.

Classifier	Specificity (%)	Sensitivity (%)	Accuracy (%)	Reference
$k$ -NN	100	94.5	96.17	Current study
PCA	100	92.9	96.5	Biopolymers (Ref. 27)
ANN	100	96.5	98.3	Biopolymers (Ref. 27)

The $k$ -NN performance parameters are shown in Table 3 and are compared with the performance parameters of other analyses done on the same set of spectral data using the PCA and ANN analysis.²⁷ The performance parameters, specificity, sensitivity, and accuracy in PCA match/no match analysis were 100, 92.9, 96.5%, respectively, whereas in case of ANN analysis, these parameters were 100, 96.5, 98.3%, respectively.²⁷ In this study, with the MATLAB PCA-based $k$ -NN analysis, the specificity, sensitivity, and accuracy are found to be 100, 94.5, and 96.17%, respectively. When compared in terms of sensitivity, ANN and MATLAB-based $k$ -NN techniques are found slightly better than the conventional PCA (match/mismatch) technique. And when compared in terms of accuracy, PCA and $k$ -NN analyses were similar in nature, whereas ANN analysis is found to be better than PCA and $k$ -NN analyses.

As already noted, we showed that by forming standard calibration sets of pathologically certified samples of normal, premalignant, and malignant tissue samples one can match a new sample with these standard sets. This leads to a more reliable diagnosis and makes it possible to do periodic screening, monitoring of effectiveness of therapy, surgical boundary check, and early detection of recurrence, if any, without the necessity of repeated biopsy. The preparation of the standard sets requires only a central well-equipped hospital facility. Once the standard sets are ready, they can be made available to any clinical setup even in small hospitals. With a relatively inexpensive instrument, the screening process can thus be carried out across large cross sections of the rural population, for whom such screening will not be easily available otherwise. To the best of our knowledge, none of the earlier workers have done this $k$ -NN analysis on oral tissue fluorescence spectra.

The great advantage of the optical method combined with statistical data analysis is that cases that show anomalous behavior can be reevaluated with a second scan with more sampling points. Furthermore, because optical methods do not require tissue removal, they may provide better tools for experts to direct diagnostic biopsies, screen for early detection, and determine tumor margins.