2.1.

Spectrum Acquisition

The cervical cell samples used in this study were obtained from the Gribbles Pathology Laboratory, Petaling Jaya, Selangor, Malaysia (a private provider of diagnostic laboratory services in performing tests for all major disciplines of pathology). The acquired samples were taken from tissue biopsies of women undergoing routine cervical cancer screening. The samples collected from ThinPrep^® solution (PreservCyt; Cytyc) along with their cytology diagnostic results were classified according to the Bethesda System 2001. In this work, we have obtained 650 normal cases, 160 LSIL cases, and 40 HSIL cases of FTIR spectra from individual cervical cells.

The cervical cell FTIR spectra were obtained by placing a small amount, $\sim 0.005\mathrm{ml}$, of liquid ThinPrep samples in a circular KRS5 window (an IR transparent cell). The liquid samples were then dried using a dryer for 2 to 3 min before the samples are induced by IR light.

After preparing the cervical cell in the KRS5 window cells, the cervical cell spectra were collected using Spectrum BX II Fourier Transform Spectrometer (Perkin Elmer type 2000) equipped with a deuterated telluride triglycine sulphate detector in mid IR region between 400 and $4000{\mathrm{cm}}^{-1}$.

The FTIR spectroscopy software was employed to manipulate the original spectrum received from the instrument. The purpose of manipulating a spectrum is to enhance its appearance.³¹ In this work, automatic baseline correction, smoothing, and normalization were applied. The spectrum was submitted to the automatic baseline correction process before it was smoothed using the smoothing package within the FTIR software. According to Quintero et al.,³² during the acquisition process, noise may affect the spectrum more than once. Thus, smoothing was required after the acquisition process to improve the appearance of spectrum.

2.2.

Features Extraction

Figure 1 shows various spectrum patterns with different prominent peaks. The prominent peaks represent the absorption bands of biochemical compounds. Based on the previous study done by Wong et al.,⁶ there are seven biochemical compounds detected in the cervical cell FTIR spectra that could be used for the classification purpose (Fig. 1).

Fig. 1

The differences of blue line with small dot, black solid line, and red line with big dot cervical cell FTIR spectra in prominent bands using FTIR spectroscopy. Note that A, B, C, D, E, F, and G refer to Amide I, Amide II, Proteins, NA II, Carbohydrates, NA I, and Glycogen, respectively. $m$ and $n$ are examples of peak region as tabulated in Table 1. $k$ and $l$ are examples of two base points of the peaks as tabulated in Table 2.

Table 1

The range of wavenumbers for the peak absorbance value determination of the biochemical components of the cervical cell FTIR spectrum.

Table 2

The proposed base points for the biochemical components of the cervical cell FTIR spectrum.

The biochemical compounds are as follows.

However, most of the acquired signal suffers from noise, which further complicates the feature extraction process. Thus, a smoothing filter is proposed.

The coefficient filter (i.e., $b_k$) must fulfill three conditions.^33,34

Condition (1) ensures the conservation of the peak area and a constant background. Meanwhile, conditions (2) and (3) avoid a phase shift between input and output data and avoid undesired oscillations at both sides of the peak, known as the wing effects. The Savitzky–Golay (SG) filter is currently being used widely among chemists for the smoothing and differentiation of the spectroscopy spectra.³⁵ Almost all spectroscopic software packages contain this standard smoothing technique. However, the last condition of the coefficient filter is not completely obeyed by the SG filters, as the last coefficients $(N_p,-N_p)$ are always negative. These negative coefficients introduce some small oscillations at the peak sides.³³ As a result, the SG smoothing algorithm can cause false-negative signals at the shoulders of each vibrating band.³⁵ In addition, the SG smoothing algorithm can lead to the loss of weak signals and the reduction of spectral resolution.³⁶

The previous problems can be solved by using binomial smoothing filters.³⁶ In other work, we used the quadratic of half ellipse (QHE) filter as a smoothing filter.³⁷ The QHE filter also fulfills the conditions in which the coefficients of the QHE filter are obtained by

However, when implementing the direct-form filter, error surface could possibly occur.³⁸ Thus, Williamson conducted research to consider implementing the cascade-form filter as a transformation of the direct-form filter.³⁸ The cascade-form filters have been proven to have better performance than the direct-form filters.^39,40 In addition, the cascade-form filters can construct low-cost systems due to their less physical modifications.^39,40 Thus, inspired by the improvement of the smoothing filter, this paper uses cascade-form filters based on the QHE and the binomial filters.³⁶

These cascade-form filters are also inspired based on analysis of the equation and the geometry of their ellipse curve. As shown in Fig. 2, when the QHE coefficient filter is plotted in $x$- and $y$-axes, it is observed that the curve is similar to that of the binomial coefficient. Both curves show similar patterns; thus, by cascading these two coefficients, it is believed that the cascade of binomial and QHE filters will produce a smoother signal, which could be used as a good filter in this work.

Fig. 2

The coefficient filters of the QHE and the binomial filters, which are plotted in $x$- and $y$-axes.

Based on the previous study, we developed a features extraction algorithm for the automated screening system, where the preprocessing is considered as the features extraction process (Fig. 3). In the previous studies, the range of the wavenumber used for analyzing between normal, LSIL, and HSIL lies in the 950 to $1800{\mathrm{cm}}^{-1}$ region.¹⁹ As plotted in Fig. 1, different types of cervical cells show different spectrum patterns with different prominent peaks in the specific bands.

Fig. 3

The developed features extraction algorithm for the automated screening system.

To extract those aforementioned features of the FTIR spectra, this study employed a peak-corrected area-based features extraction (PCABFE) algorithm, as presented in Fig. 4.⁴¹ The PCABFE extracts three primary features: the height of specific peaks, the height of slope between amide I and amide II, and the corrected area in specific regions. The features are calculated using

Fig. 4

Description of the PCABFE algorithm procedures.

For evaluation of the automatic feature extraction performance, a correlation test was conducted to determine the capability of the proposed PCABFE algorithm as compared to the manual extraction by using the FTIR software.

2.3.

Feature Selection

The extracted features are reduced by discriminant analysis (DA) and principal component analysis (PCA) techniques. Both techniques are employed to determine the dominant features, where the irrelevant or unrelated features that could deteriorate the generalization performances of artificial neural network (ANN) are eliminated to avoid a large dimensionality problem.⁴²

Using the Wilks’ lamda method for DA, the optimum features are selected based on the null hypothesis and the $p$-value. The null hypothesis is the equality of several classes of parameters, while the $p$-value is the probability of observing the given sample result under the assumption that the null hypothesis is true.²³ Significant level ($\alpha$) chosen for admitting or rejecting the null hypotheses is 0.05. If the $p$-value is less than $\alpha$, then the null hypothesis is not rejected. On the other hand, the null hypothesis is rejected when the $p$-value is more than $\alpha =0.05$.

For PCA, a scree plot of all the PCs of input features of the cervical cell dataset is presented. The appropriate numbers of principal components to be used are selected by considering the result from the scree plot and eigenvalues. The features with higher eigenvalue in eigenvectors of the appropriate principal components are selected as the optimum features.

2.4.

Classification of Cervical Precancerous Fourier Transform Infrared Spectrum Using Neural Network

After the signal acquisition, signal smoothing, features extraction, and features selection process, the dominant features are then fed as input data to the intelligent classification stage. One of the aims of this study is to classify the cervical cell FTIR spectra into three classes (normal, LSIL, and HSIL). In this paper, the hybrid multilayered perceptron (HMLP) network is trained with the modified recursive prediction error algorithm for the classification purposes.

During the training process of the HMLP network, this study employs a 10-fold cross-validation method. The detail on the 10-fold cross-validation method can be found in the previous study.⁴³ The data is partitioned into 10 sized segments or folds. Ten run iterations of training sets (i.e., 585 normal, 144 LSIL, 36 HSIL for each fold) and testing sets (i.e., 65 normal, 16 LSIL, and 4 HSIL for each fold) phases are performed with different sets in each run. A different fold of the data is used for testing, whereas the remaining nine folds are used for training in each run.

The datasets with selected features based on DA only and DA–PCA techniques are tested to obtain the better system for the automated system. The confusion matrixes are presented for evaluation purposes. In this paper, the comparison of performance is done based on accuracy result between this study and related published work. The cervical cell spectra from the FTIR spectroscopy were compared with cytology (the gold standard). Therefore, the confusion matrix is important to be presented in this paper to present an actual condition

In this study, our system uses validity measures of screening with normal and abnormal (LSIL and HSIL) classes. True positives and true negatives are obtained when the abnormal (LSIL and HSIL) and normal cervical cells are correctly classified. False positive (FP) is obtained when the normal cervical cell is incorrectly classified as an abnormal cervical cell (LSIL and HSIL). False negative (FN) is obtained when the abnormal cervical cell (LSIL and HSIL) is incorrectly classified as a normal cervical cell.

3.1.

Features Extraction Results

The primary features (i.e., CA amide 1, CA amide 2, CA proteins, CA NA II, CA NA I, PH amide 1, PH amide 2, PH proteins, PH NA II, PH carbohydrate, PH NA I, PH glycogen, and height of slope between amide 1 and amide 2) create a very strong linear relationship with correlation test results more than 0.95 (approaching one), as presented in Table 3.

Table 3

Results of correlation test of primary features of cervical cell FTIR spectra.

Generally, all features have a strong linear relationship (correlation test achieved more than 0.8, as presented in Table 4) compared to those extracted manually by using FTIR spectroscopy software. The results show that the PCABFE algorithm and the manual extraction using the FTIR spectroscopy software have constructed a linear curve for all features. Based on the evaluation of the PCABFE performance, a total of 32 features are extracted, as listed in Table 4.

Table 4

Results of the correlation test of 32 possible features of the cervical cell FTIR spectra for classification purpose.

3.2.

Features Selection Results

Based on the 32 possible features, as shown in Table 4, the DA and PCA techniques are employed to determine the dominant features for the classification process in the next stage. Table 5 tabulates the results attained from DA of 32 features. Based on the results, 11 features show insignificant effect or low impact to the classification process as the $p$-values distribution obtained are more than 5% (as made bold in Table 5). The features with $p$-value distribution less than 5% are said to have high impact on the classification process. Thus, based on this argument, 21 of 32 features have been selected as dominant features for the DA process.

Table 5

Results for stepwise method of DA technique for 32 extracted features.

Afterward, the 21 features of the cervical cell FTIR spectra are further analyzed using the PCA technique. By considering the results from the scree plot and eigenvalues, the appropriate number of principal components to be used is four. The four principal components [first principal component (PC1), second principal component (PC2), third principal component (PC3), and fourth principal component (PC4)] are used in the features selection. Table 6 lists the variables that have strong relationship with PC1, PC2, PC3, and PC4.

Table 6

Variables that have strong relationship from PC1 to PC4.

From the result of PC1, the variables that tend to have strong relationship are $x_3(r_{3c})$, $x_5(r_{5c})$, $x_7(r_{7c})$, $x_{13}(r_{3p})$, $x_{14}(r_{4p})$, $x_{15}(r_{5p})$, $x_{16}(r_{6p})$, $x_{17}(r_{7p})$, $x_{20}(r_{10p})$, $x_{22}(r_{12p})$, and $x_{32}(h_{1s})$. In PC2, the variables which have strong relationship are $x_9(r_{9c})$, $x_{12}(r_{2p})$, $x_{23}(r_{13p})$, $x_{24}(r_{14p})$, and $x_{26}(r_{16p})$. In PC3 and PC4, the variables which tend to have strong relationship are $x_2(r_{2c})$ and $x_8(r_{8c})$, and $x_4(r_{4c})$ and $x_6(r_{6c})$, respectively. Therefore, the feature selection process result reveals that the important or dominant input features consist of 20 variables: $x_2(r_{2c})$, $x_3(r_{3c})$, $x_4(r_{4c})$, $x_5(r_{5c})$, $x_6(r_{6c})$, $x_7(r_{7c})$, $x_8(r_{8c})$, $x_9(r_{9c})$, $x_{12}(r_{2p})$, $x_{13}(r_{3p})$, $x_{14}(r_{4p})$, $x_{15}(r_{5p})$, $x_{16}(r_{6p})$, $x_{17}(r_{7p})$, $x_{20}(r_{10p})$, $x_{22}(r_{12p})$, $x_{23}(r_{13p})$, $x_{24}(r_{14p})$, $x_{26}(r_{16p})$, and $x_{32}(h_{1s})$.

3.3.

Intelligent Classification Results

For the intelligent classification results, the datasets using 21 features from the DA and 20 features from the DA–PCA processes are tested, respectively, to determine the most stable system. These dominant features from the DA and the DA–PCA processes are individually fed into HMLP classifier for classification purpose. The HMLP classification results of the DA datasets with 21 dominant features are presented in Table 7. The HMLP with the DA datasets could detect 634 normal from 650 totals normal, 102 LSIL from 160 totals LSIL, and 24 HSIL from 40 totals HSIL cervical FTIR spectra.

Table 7

Confusion matrix of DA datasets and DA–PCA datasets for distribution of those individuals screened by screening status and cytology condition status.

Meanwhile, the HMLP classification results of the DA–PCA datasets with 20 dominant features are presented. As shown in Table 7, the HMLP could detect 621 normal from 650 totals normal, 106 LSIL from 160 totals LSIL, and 27 HSIL from 40 totals HSIL cervical FTIR spectra.

Overall, the results tabulated in Tables 7 demonstrate that the HMLP shows a good performance for classifying cervical cell FTIR spectra into normal, LSIL, and HSIL classes. However, when the DA–PCA datasets were used, higher FP values were achieved than the HMLP with DA datasets. As shown in Table 7, the FP value was given as 29 (which is 26 normal cells incorrectly classified as LSIL, and three normal cells are incorrectly classified as HSIL cases) for DA–PCA dataset. While the FP values from DA datasets were significantly lower with 16 normal cases incorrectly classified as LSIL cases, and the normal cells were not classified as HSIL class. These results occurred because, in fact, the HSIL cells are high stages of abnormality, and their characteristics exhibit apparent differences from the normal cells. Meanwhile, the FN values for the HMLP with DA datasets are higher than the HMLP with DA–PCA datasets given in detail in Table 7. The FN values are 52 LSIL cells incorrectly classified as normal class for DA–PCA datasets. No HSIL cells are incorrectly classified as normal class. Meanwhile, the FN values of DA datasets obtained 54 LSIL cells incorrectly classified as normal class, as tabulated in Table 7. The HSIL cell is also not classified as normal class. These results occurred because the LSIL cells, in fact, only affect the surface of the cervical tissue. The majority will regress back to normal spontaneously.⁴⁴ Over time, a small proportion will continue to develop into true cancer. The HSIL cells cannot be recovered to be normal cells. Based on the FP and FN values for both datasets results in Table 7, the system can significantly differentiate between normal and HSIL cells, and the LSIL and normal cells can also be distinguished. However, small portions of the LSIL are incorrectly classified as normal cells, and part of normal cells are incorrectly classified as LSIL cells. Similarly, these results are expected since most LSIL cells will regress back to normal.⁴⁴ Therefore, our system produced consistent results with acceptable accuracy to classify the cervical precancerous cells.

The result of the HMLP with the DA dataset shows relatively better performance in term of stability. Therefore, based on the 21 selected features (from DA datasets), the HMLP classifier shows a good performance for classifying cervical cell FTIR spectra into normal, LSIL, and HSIL classes with 92% of accuracy. The promising results obtained in this stage are utilized to develop an automated screening system for cervical cancer. The results of the proposed system are elaborated in detail in Sec. 3.4.

3.4.

Automated Screening System for Cervical Cancer

The proposed screening system contains the automatic features extraction and intelligent screening. Figure 5 shows the interfacing of the system. A user is only required to input the cervical cell FTIR spectra. The smoothing spectrum, the features of cervical cells FTIR spectrum, and the case and class of the cervical cell FTIR spectra will automatically be displayed. This procedure could possibly produce faster screening results and decrease the dependency on human experts, thus reducing the workload of pathologists.

Fig. 5

Interfacing of the proposed automated screening system for cervical cancer.

To date, several researchers have developed cervical cancer screening tools based on the spectroscopy approaches. Our proposed system can be compared to the other developed system using FTIR spectroscopy.¹⁹ This system used only five features, which are obtained from the ratios of the peak height values (1) glycogen/NA I, (2) NA I/carbohydrates, (3) NA I/amide II, (4) proteins/amide I, and (5) NA I/proteins to differentiate the different types of cervical cell spectra. We also include the developed system that uses Raman spectroscopy²⁷ in our comparison. The comparison results are shown in Table 8, where the three systems of A¹⁹, B²⁷, and C (proposed system) are tabulated.

Table 8

Comparison between the published systems and the proposed system results.

Table 8 suggests that our proposed system achieved the best performances in term of accuracy with 92%. This is likely because our proposed system used more dominant features (21 features from DA datasets) to differentiate between three classes of the cervical cells (normal, LSIL, and HSIL cells).

Therefore, we suggest, based on the aforementioned explanation, that our system simultaneously has better results to differentiate the cervical cells due to the proposed signal smoothing filter, PCABFE algorithm to extract features from the cervical cell FTIR spectra, DA to select the optimum features (21 features), and HMLP network for classification.