2.1.

Sample Preparation

Formalin-fixed, parafin-embedded tissues from colon adenocarcinoma patients were received from the histopathology files of Soroka University Medical Center (SUMC). The tissue samples selected for this study had normal, adenomatous polyp (pre-malignant) and malignant areas. Two thin (10 μm) parafin sections were cut from each biopsy; one was placed on a zinc–selenium slide and the other on a glass slide. The slide to be used for microscopic FTIR measurements was deparaffinized and dehydrated using a standard procedure.¹⁶ The second slide was stained with hematoxylin and eosin for a histology review.

2.2.

FTIR Microspectroscopy

Microscopic FTIR measurements were performed in transmission mode using the FTIR microscope IRscope II with a mercury–cadmium–telluride (MCT) detector coupled to the FTIR spectrometer (Bruker Equinox model 55/S, OPUS software). All three kinds of cells (normal, polyp, and malignant) in the tissue samples were obtained from each patient. Measurements were performed on normal, polyp, and cancerous epithelial cells from intestinal tissues using microscopic inspection in parallel with a detailed pathological analysis of the tissue architecture. During each measurement, the area of the sites measured was about 50×50 μm ² , it had only a few cells. The spectra measured cover a wave number range of 600–4000 cm ⁻¹ in the mid-IR region. The number of co-added scans was increased to 128 to achieve a high signal to noise ratio. A data base of about 300 measured spectra at a resolution of 4 cm ⁻¹ was created. The spectra were normalized to an amide I peak after baseline correction for the entire spectrum using the rubber band method (a polynomial function). Spectral analysis was done using OPUS software. The integrated absorbance were calculated by a band fitting procedure that was reported previously.^{17 18} The error bar (standard error) was calculated for each sample (normal or malignant) separately for all patients reported in this article.

2.3.

ANN Analysis of the FTIR Spectra of Normal, Polyp and Malignant Cells from Biopsy Tissue Samples of Colon Cancer Patients

Artificial intelligence (AI) is a well established area of computer science that focuses on the creation of software capable of performing highly sophisticated intelligent computational work similar to that of the brain. AI has two different classes, namely, the expert systems that simulate human experience using a set of rules and the other comprises systems modeling the function of the human brain.²¹ The human brain is the best recognition system in the world. In the recognition process, the biological neuron in a specific area of the brain identifies a given object by linking it with previous experience through past learning. ANN can be a computational model derived from hundreds of single units (like neurons) connected in a network. Each model [Figure 1(a)] is a processing element with weighted (the degree of importance) inputs, a transfer function (prescribed task) and the output. As the neurons are interconnected in the brain, the power of the ANN mainly depends on the number of connections between artificial neurons. ANNs can process huge amounts of data and make accurate predictions. There are various types of NNs (neural networks or models) designed that have different transfer functions and network connections. The applications of ANNs in medicine and biology are increasing rapidly.^{19 20} A detailed account of ANNs can be obtained from two excellent books^{21 22} and a review article.²³ In our case, X(i) are the FTIR spectra, W(i) are the weights and Z is the assignment (prediction) given by the model f(w,x).

Figure 1

(a) Schematic of an artificial neuron. (b) Supervised network with a backpropagation learning rule.

Multilayer perception (MLP) has commonly been used in many NN applications due to its simple implementation. The network consists of many processing elements connected in several layers. The output of one processing element is connected to the input paths of other processing elements through connection weights. When presented repetitively with the input and the desired output the MLP organizes internally, gradually adjusting the weights to achieve the input/output mapping desired. Given enough data it is possible to design and teach a MLP with one hidden layer to reproduce the behavior of any process, linear or nonlinear. The training process is shown in Figure 1(b), where the network is trained or “taught” by showing a set of examples, observing the network answers and correcting the “wrong” answers by changing or adapting the network weights.

2.4.

Multiscale Decomposition by Fast Wavelet Transform

An appropriate description of a feature is considered to be one of the most important components of classification procedures. The result of feature extraction is a more concise description that still retains most of the spectrum characteristics. The use of features rather than the signal itself makes the classification process easier and faster. In this case, we have examined the performance of the NN based classifiers by performing multiscale decomposition of the spectrum.²⁴ The wavelet transform provides an important tool for signal analysis and feature extraction.²⁵ It provides a good local representation of the signal in both the time domain and the frequency domain. Shown in Figures 2(a) and 2(b) is the Fourier transform (FT), which is global and provides a description of the overall regularity of the signals, whereas the wavelet transform looks for the spatial distribution of singularities. While the FT decomposes the signal into a series of sinusoidal components, the fast wavelet transform (FWT) decomposes the signal into a series of wavelets of different scales and positions.

Figure 2

Graphic representation of a one-dimensional signal decomposition by (a) Fourier and (b) wavelet transforms.

In this article, the FWT, proposed by Mallat and Zhong,²⁵ was applied. The sampled spectrum is decomposed into an orthogonal set of wave forms that are the dilations, translations, and modulations of the Coiflet wavelet (the mother wavelet). The Coiflet wavelet was chosen because in practice it showed better results than other common wavelets that were tested. The wavelet transform is computed by convolving the spectrum with these dilated wavelets. The wavelets’ coefficients of the different scales offer a compact representation of the spectrum signal. It is evident that the transform involves differentiation and progressive smoothing. Details of the peaks are gradually lost as the downward slopes of the wave are picked up at higher scales. The number of scales is chosen by searching for the optimal signal representation. It was found that scales higher than the first five do not add significant information about the spectrum. In the present case, the performance of the MLP based classifier for different sets of wavelet coefficient features was examined (Table 1). The wavelet coefficients were located around the two main peaks of the spectra.

Table 1

The data employed in this work was extracted from 25 patients. The data sets consist of three groups: the cancer group (83 records extracted from 11 patients), the control group (109 records extracted from 6 patients) and the polyp group (106 records extracted from 12 patients). A total of 298 records was available. The samples obtained were classified with the help of an expert pathologist and confirmed by clinical diagnosis using standard pathological methods. In this study, training and test sets were selected randomly from the same data sets. Seventy percent of each set was employed for training and the remainder was for testing. In addition, the simulations were repeated 100 times, with the same networks parameters but with different sets of randomly selected training vectors, and then the results were averaged.

3.1.

Morphological Changes for Normal, Polyp and Malignant Lesions in Human Colonic Tissue

A histological cross section of formalin-fixed human colonic tissue is shown in Figure 3. The encircled areas in Figures 3(a)–3(c) show the site covered by FTIR-Monte Carlo (MC) measurements in this study. Figure 3(a) shows the crypts of the normal colonic mucosa. The epithelium surrounds a center opening and the lumen of the crypt. The normal epithelial cells show a single row of regular cells with small nuclei in the periphery of the cell and abundant cytoplasm. The nuclear to cytoplasm ratio is low. The extracellular matrix (lamina propria) around the epithelium consists of scattered lymphocytes and granulocytes (inflammatory cells) (dark dots). Figure 3(b) shows the neoplastic epithelial lesion (tubular adenoma), which shows a slight change from the normal architecture, with a slightly irregular crypt, and it has crowded nuclei. The nuclear to cytoplasm ratio is increased compared with that of normal colon tissue, while the lumen becomes narrow. Figure 3(c) depicts the adenocarcinoma of the colon, showing stratification of hyperchromatic nuclei with marked atypical cells. The glands become irregular and are back to back. The size of the lumen is uneven and contains cellular debris. There is a desmoplastic stromal reaction with a secondary inflammatory reaction.

Figure 3

Histological image cross sections of formalin-fixed human colonic tissue stained with hematoxylin teosin: (a) normal, (b) polyp, and (c) malignant. Circled areas are those selected for FTIR-MC measurements.

3.2.

FTIR Microspectroscopy (FTIR-MC) of Normal, Adenomatous Polyp and Malignant Human Colonic Tissues

Microscopic mid-IR spectra of normal, adenomatous polyp and malignant cells from biopsy tissue samples of three different patients of varying disease stages are shown in Figure 4. Patients (a), (b), and (c) are classified as those with early, moderate and advanced stages of malignancy, respectively, as assigned by group pathologists. Due to the limited availability of all three kinds of cells (normal, polyp, and malignant) in any given tissue sample, only three cases are given. The absorption due to normal tissue was higher than polyp and cancerous cells in this entire region of the spectrum in all three patients. In the early case [patient (a)], the absorbance of adenomatous polyp was higher than in the malignant cells. This trend was reversed in the advanced case [patient (c)]. This effect was more pronounced in the phosphate region compared to in other regions of the spectrum. There was a distinct change in the pattern in the region between 1000 and 1200 cm ⁻¹ for normal and abnormal tissue samples, which include polyp and malignant tissues. In the spectrum of normal cells, splitting can clearly be observed, whereas it disappeared in the case of adenomatous polyp and malignant cases. This has been observed in most of the cases studied by our group. No shifts in frequency were observed over the entire region (600–2000 cm ⁻¹ ) in any of the three cases.

Figure 4

FTIR microspectroscopy of normal, polyp and malignant tissue samples in the range of 900–1600 cm ⁻¹ from three patients. (a), (b), and (c) early, moderate and advanced stages of malignancy, respectively. All spectra are consistently normalized to amide I.

3.3.

Phosphate Content

The variation of phosphate levels measured by integrating the absorbance between symmetric (1000–1150 cm ⁻¹ ) and asymmetric (1170–1310 cm ⁻¹ ) bands for 24 cases is presented in Figure 5. On average the phosphate level of polyp and malignant tissue samples was lower than that of the controls. The average content of polyp and malignant samples was equal. It is also clear from Figure 5 that the diversity among polyps was larger than for the malignant and the controls.

Figure 5

Phosphate as a biological marker derived from FTIR spectra for 24 patients. The phosphate content is calculated as the sum of the integrated absorbance of the symmetric and asymmetric bands of the phosphate group. PeakFit software was used.

3.4.

RNA/DNA

The ratio of absorbance at 1121/1021 provides an estimate of the RNA/DNA ratio in cells and it is found that this ratio increases with the stage of malignancy.²⁶ Our analysis presented in Figure 6 indicates that the RNA/DNA ratio was lower for normal cells in comparison to that for polyp and malignant cells. Also, it is interesting to note that the RNA/DNA ratio was nearly constant for normal and polyp samples. But, there was a marked variation among the malignant samples obtained from different patients.

Figure 6

Absorbance ratio at 1121/1020 as the RNA/DNA ratio for all 24 patients.

3.5.

Two-dimensional Carbohydrate Versus Phosphate Content

It is interesting to find the correlation between different metabolites among the different categories of cells. The two-dimensional (2D) plot shown in Figure 7 gives the relationship between carbohydrate and phosphate content in normal, polyp, and malignant cells. It is clear from Figure 7 that there is a linear correlation between these two metabolites. As was reflected in the earlier spectra, the normal samples had the highest carbohydrate and phosphate contents. There was no distinction between polyp and malignant samples. The perfect mixing of these two cells categorizes them into a single family.

Figure 7

2D plot of carbohydrate vs the phosphate content for 24 patients. The carbohydrate content is measured as the ratio of the absorbance at 1080/1545.

3.6.

Classification of Normal, Polyp and Malignant Samples by ANN

A summary of the results obtained with the proposed classifier for five sets of features (Table 1) is shown in Table 2. The feature combinations were selected according to the biomarkers in their respective spectral regions. For example, the spectral region between 1055 and 1719 cm ⁻¹ includes the phosphate (both symmetric and asymmetric stretching bands) and total protein contents of the cells. The best results were obtained for the MLP with a set composed of 13 input feature coefficients located between 2465 and 2963 cm ⁻¹ which correspond to the absorption of phospholipids (set 1, Table 1). The MLP-based classifier resulted in sensitivity values of 89.4, 83.1, and 81.3 for normal, cancer, and polyp tissues, respectively. Although the sensitivity values obtained using this classifier were high, a careful analysis of Table 3 revealed that the false positive for normal was up to 10.6 which might be classified as either cancer (2.0) or polyp (8.6). On the other hand, the sum of the false negative and false positive for the polyp group was about 18.6 (about 9.7 of the polyp cases might be classified as normal while up to 8.9 might be classified as cancer). It is important to note that compared with the results obtained by the local density approximation (LDA)¹⁷ method for classification of normal and malignant samples from colon cancer patients, these results show good agreement and have greater consistency.

Table 2

Table 3