Home

Search

Collections

Journals

About

Contact us

My IOPscience

Terahertz spectroscopic investigation of human gastric normal and tumor tissues

This content has been downloaded from IOPscience. Please scroll down to see the full text. 2014 Phys. Med. Biol. 59 5423 (http://iopscience.iop.org/0031-9155/59/18/5423) View the table of contents for this issue, or go to the journal homepage for more

Download details: IP Address: 158.42.28.33 This content was downloaded on 19/05/2015 at 08:17

Please note that terms and conditions apply.

Institute of Physics and Engineering in Medicine Phys. Med. Biol. 59 (2014) 5423–5440

Physics in Medicine & Biology doi:10.1088/0031-9155/59/18/5423

Terahertz spectroscopic investigation of human gastric normal and tumor tissues Dibo Hou1, Xian Li1, Jinhui Cai2, Yehao Ma1, Xusheng Kang1,3, Pingjie Huang1 and Guangxin Zhang1 1

  Department of Control Science and Engineering, Zhejiang University, Hangzhou, People’s Republic of China 2   China Jiliang University, Hangzhou, People’s Republic of China 3   Zhejiang University City College, Hangzhou, People’s Republic of China E-mail: [email protected] Received 7 December 2013, revised 3 July 2014 Accepted for publication 9 July 2014 Published 28 August 2014 Abstract

Human dehydrated normal and cancerous gastric tissues were measured using transmission time-domain terahertz spectroscopy. Based on the obtained terahertz absorption spectra, the contrasts between the two kinds of tissue were investigated and techniques for automatic identification of cancerous tissue were studied. Distinctive differences were demonstrated in both the shape and amplitude of the absorption spectra between normal and tumor tissue. Additionally, some spectral features in the range of 0.2~0.5 THz and 1~1.5 THz were revealed for all cancerous gastric tissues. To systematically achieve the identification of gastric cancer, principal component analysis combined with t-test was used to extract valuable information indicating the best distinction between the two types. Two clustering approaches, K-means and support vector machine (SVM), were then performed to classify the processed terahertz data into normal and cancerous groups. SVM presented a satisfactory result with less false classification cases. The results of this study implicate the potential of the terahertz technique to detect gastric cancer. The applied data analysis methodology provides a suggestion for automatic discrimination of terahertz spectra in other applications. Keywords: transmission time-domain terahertz spectroscopy, gastric tumor, absorption coefficient spectra, classification (Some figures may appear in colour only in the online journal)

0031-9155/14/185423+18$33.00  © 2014 Institute of Physics and Engineering in Medicine  Printed in the UK & the USA

5423

D Hou et al

Phys. Med. Biol. 59 (2014) 5423

1. Introduction Gastric cancer is a leading cause of cancer death worldwide. In China, it is estimated that around 400000 new cases are diagnosed and 200000 people die from gastric cancer each year, making it the most common malignancy for both sexes (Fujioka et al 2004). If the disease is found and removed at an early stage, a high cure rate of up to 90% can be achieved. Unfortunately, the symptom of precancerous lesions is occult and hard to be detected. Only when the symptom becomes overt does the clinic method become efficient. Currently, gastric cancer is usually diagnosed by endoscopic screening. Once abnormal areas are located, tissue samples are taken and sent to histopathology to determine whether the margin is tumor or not, which is time-consuming and labor-intensive. Thus, a quick and accurate diagnostic method has key importance. Terahertz radiation, which lies between the microwave and infrared regions and which covers frequencies of 0.1–10 THz (1 THz = 1012 Hz), has some practical advantages as a sensitive analytical technique in medical diagnosis. For example, it is non-ionizing and causes no harm to living tissue because of its low energy. Furthermore, it is less susceptible to scattering in comparison with optical and infrared methods because of its longer wavelength. Above all, many biomolecules and polar molecules have distinctive absorption spectra in the terahertz frequency range, which makes the terahertz radiation a viable tool for identification and characterization of biological tissue (Fitzgerald et al 2006). In the past decade, THz time-domain spectroscopy and imaging techniques have been generally proposed in biomedical applications, particularly cancer detection. The initial studies on cancer diagnosis by THz techniques were focused by investigating the differences of terahertz properties between healthy and diseased tissue from human skin (Woodward et al 2002, Pickwell et al 2004, Wallace et al 2006, Joseph et al 2011) and breast (Fitzgerald et al 2006, Wallace et al 2006, Ashworth et al 2009, Chen et al 2011). The absorption coefficient and refractive index in the terahertz frequency range were found to be higher for the freshly excised tissue that contained cancer compared to normal tissue by terahertz pulsed spectroscopy (Pickwell et al 2004, Wallace et al 2006, Ashworth et al 2009). Simulating the interaction of terahertz radiation with normal and cancerous tissue has been made to help understanding the origin of contrast in terahertz properties of cancerous tissue (Pickwell et al 2005, Ashworth et al 2009). A distinct contrast was also shown in THz imaging of diseased skin tissue compared to that of normal tissue (Woodward et al 2002, Joseph et al 2011). The shapes of tumor regions in THz imaging of breast cancers corresponded well with those derived from histology (Fitzgerald et al 2006). The terahertz techniques have also been suggested for a number of other cancer diagnostic purposes, such as liver cancer (Enatsu et al 2007, Sy et al 2010, Chen et al 2013) and colorectal cancer (Reid et al 2011, Eadie et al 2013). Similarly, a higher water content and absorption coefficient have been found in diseased freshly liver tissue than normal tissue (Sy et al 2010). The investigation of a terahertz technique for possible use in ex vivo colon cancer diagnosis showed sensitivities of 90–100% and specificities of 86–90% in classification between normal tissue and cancer tissue by using a series of THz parameters and optimized processing methods (Eadie et al 2013). Since water has strong absorptions at THz frequencies, variation in water content is considered as a likely contributing factor to the contrast mechanism between normal and abnormal tissue. However, the still noticeable contrasts in the studies on terahertz properties of dehydrated paraffin-embedded tissues (Loffler et al 2002, Enatsu et al 2007, Nakajima et al 2007, Jung et al 2011, Hassan et al 2012) indicate that there are additional contrast-contributing factors, such as the increased vasculature of cancerous tissue, decreased lipid content, increased cell density, and the presence of certain proteins (Fitzgerald et al 2006). The higher 5424

D Hou et al

Phys. Med. Biol. 59 (2014) 5423

Table1.  Information of investigated patients and representative samples.

Case No. Gender Age(y)

Lesion type

Sample thickness (mm) Tumor grade Normal Tumor

1 2 3 4 5

Gastric adenocarcinoma Gastric adenocarcinoma Gastric adenocarcinoma Gastric adenocarcinoma Gastric adenocarcinoma

I II III II II

F M F M M

45 50 58 53 56

0.863 0.546 0.754 0.725 0.750

0.840 0.599 0.899 0.736 0.745

concentration of water in cancer tissue may cover the effects of these biomarkers in the performance of terahertz radiation in cancer detection. Terahertz techniques have been considered to have potential for revealing biological changes in diseased tissue during the process of carcinogenesis. However, the precise origin for the contrast of terahertz response in diseased tissue is still unknown and there have been few researchers proposing efficient data analysis methodology for classification between normal and cancerous tissue based on the terahertz absorption coefficient spectra. Therefore, in this study we investigated the differences of terahertz absorption characteristics between dehydrated normal and cancerous gastric tissue samples by employing transmission time-domain terahertz spectroscopy (TTDTS). Furthermore, statistical analysis was conducted on the distribution of terahertz absorption peaks in an attempt to figure  out the general terahertz spectral features and their biological sources for cancerous gastric tissue. In addition, we proposed a clustering methodology to categorize the measured samples into normal and cancerous groups. The satisfactory classification result suggests that methodology can be also applied to automatic discrimination of terahertz spectra for other classification tasks by terahertz techniques. 2.  Materials and experimental details 2.1.  Sample preparation

The tissue samples investigated in this study came from five patients who were undergoing gastric resection surgery at the Second Affiliated hospital of Zhejiang University School of Medicine. All of the patients (aged 45–58, mean age 52±4; 3 males and 2 females) were diagnosed with gastric adenocarcinoma but not with the same tumor grade. The detailed information of the patients is listed in table 1. The tumor grade I to III corresponds to tumor malignant degree from low to high. A pair of normal and cancerous gastric tissues were retrieved from each patient and treated by the standard pathology procedure for histopathology examination. At first, the freshly extracted tissues were fixed with 10% neutral buffered formalin for 24 h and they were then washed with water to flush out the fixative in the tissues. Next, the tissues were successively immersed into 70%, 80%, 95%, 100% ethanol to be dehydrated and kept for 1–2 h at every purity until complete dehydration. The dehydrated tissues were then treated with xylene to make the tissues transparent. They were finally embedded in paraffin blocks. The processed tissues were dissected and cut into slices by a cryotome. After early trials to slice tissues with different thicknesses, the finally sectioned slices have a thickness of around 0.7 mm. Except for patient 3, from which the tumor tissue has a higher thickness due to handling issues, the thicknesses of the normal sample and tumor sample are comparable for almost all of the patients. Finally, all the tissue slices were trimmed to an area of 5 mm × 7 mm. 5425

D Hou et al

Phys. Med. Biol. 59 (2014) 5423

Beam Splitter Ti:Sapphire Femtoscecond Laser 80 MHz 56fs

M1 Probe Beam λ/2 Plate M2

Off Axis Parabolic Mirror Si-Lens

M3 Mortorized Delay Line

Lens Sample ITO GaAs

λ /4 Plate

Lens M4

M5

Computer

M7

Polarizer

ZnTe

Wollaston Prism M6

A

Lock-in Amplifier

Figure 1.  The schematic diagram of the terahertz transmission time-domain system.

2.2.  Experimental setup

The THz-TDS system used in this experiment is developed by Zomega Terahertz Corporation in USA. The schematic diagram of the whole system is shown in figure 1. The system uses photoconductive methods to generate and detect terahertz pulses in transmission mode. Optical excitation was achieved by a commercial mode-locked Ti: sapphire laser (Coherent Company in USA) which produces 56 fs pulse at center wavelength around 800 nm with repetition rate of about 80 MHz and an average power of 960 mW. The system gives a usable frequency range of 0.1–3 THz and a terahertz beam with a diameter of 1 mm. A more detailed description of the system is presented in the article by Liu et al (2013). During the measurements, the THz beam path was purged with dry nitrogen to avoid the absorption of water vapor. The humidity was maintained at approximately zero and the temperature was kept at room temperature. 2.3.  Parameter extraction

The light intensity difference of probe beam and pump beam is proportional to the THz electric field intensity. So, based on the measured transmitted THz electric field pulses of the samples, the terahertz optical properties of gastric tissues could be derived. The reference measurement was created by focusing the beam through no sample. Fourier transformation was performed on the THz time signal and the absorption coefficient α(ω) was calculated according to the following expressions (Dorney et al 2001): c φ (ω ) n(1) +1 (ω ) = ωd 5426

D Hou et al

Phys. Med. Biol. 59 (2014) 5423

(a)

(b)

(c)

2mm

1um

Figure 2. (a) Visible image of gastric tumor from patient 4. The other samples have the

same size. H&E stained histopathology of gastric normal tissue (b) and tumor tissue (c).

2κ (ω)ω 2 4n(ω) (ω ) = α = ln (2) c d A(ω)(n(ω) + 1)2

where A(ω) and φ(ω) are the amplitude ratio and phase difference of a reference and sample signal, which can be directly obtained from the transformed pulses. ω represents the frequency, d is the sample thickness, and c is the velocity of light in vacuum. n(ω) is the calculated refractive index. 2.4.  Data acquisition

The gastric tissue sample under investigation was mounted on a customized specimen holder without any coverings. Gold markers on the specimen holder guaranteed that every sample was fixed at the same position. Then, the specimen holder was moved in the vertical plane, with a step size of 0.2 mm, which enabled the sample to be scanned by the terahertz beam. Transmitted THz electric field pulses were measured and 10 scans were averaged at each point. During the measurement process of the sample, a false appearance was found in the absorption coefficient spectra measured at the edge of the sample where the terahertz spot could not be totally covered by the sample, which was proved to have no correlation with the terahertz absorption characteristics of the sample. Therefore, we extracted the data obtained at five points, which were located at the middle part of the sample as the efficient signals describing the terahertz characteristics of the tissue sample. The presentation of the false appearance and selection of measurement points within a sample is given in the appendix. To investigate reproducibility, the same tissue sample was measured repeatedly at different periods and several slices from the same gastric tissue were tested. After measurement with TTDTS, all tissue sections were sent for routine pathological H&E staining and examination in the hospital to check the consistency within the samples. Slices exhibiting sufficiently homogenous were chosen as investigated samples, from which two slices, one cancerous and one healthy, were picked as representative samples to analyze for each patient. The thicknesses of the ten representative samples are displayed in table 1 and the variation within a sample is less than 6%. The H&E stained histopathology images of normal and tumor gastric tissue are presented in figure 2. To ensure that the influence of the paraffin on the end result is insignificant, it has been taken out from the samples prior to the measurement, it has also been reported as being transparent at THz frequencies (Wahaia et al 2011). The working frequency of the absorption coefficient spectra ranged from 0.2 THz to 2 THz, which escaped from the strong noise associated with higher frequencies. By aggregating the data acquired from the chosen points within the 10 representative tissue samples, 50 absorption 5427

D Hou et al

Phys. Med. Biol. 59 (2014) 5423

coefficient spectra of gastric tissues were calculated in all. A total of 25 spectra were from normal tissue samples and 25 spectra corresponded to tumor tissues. 3.  Absorption coefficient spectra of gastric tissues 3.1.  Absorption coefficient spectra of gastric tissues from one patient

Figure 3 shows the average absorption coefficient spectra of a pair of gastric tumor and corresponding normal tissues that were obtained from the same patient (patient 4). The error bars highlight the spread of measurements and the variation between different tested points within the tissue samples. In figure 3(a), the distinctions between the normal and tumor gastric tissue can be recognized. For most of the frequency range, the difference in the amplitude between the two tissue types is clear. Only at a few frequency points do the error bars overlap. There is also a difference in the shape of the curve between the normal and cancerous tissue. A characteristic absorption peak at the range of 1.3~1.4 THz is exhibited in the average spectrum of a gastric tumor, which is missing in that of normal tissue. Additionally, there is a peak drift at lower frequencies between the tumor and normal tissue. These contrasts may be associated with the physiologic changes of the tissue during the process of carcinogenesis. The normalized percentage difference plot shown in figure 3(b) demonstrates that at lower THz frequencies the differences between the average absorption coefficient of tumor and normal tissue are greater, which agrees with other THz studies of human tissue (Ashworth et al 2009, Reid et al 2011). The differences of the average absorption coefficient spectra between paired tumor and normal tissues from the other four patients are also distinct, although the shapes of the absorption spectra of normal tissues from different patients have a relatively great diversity due to the variety of the tissues. The greater differences between normal and tumor tissues from other patients fall into lower THz frequencies. However, at higher THz frequencies the tumor tissue displays either greater or weaker absorption than normal tissue, which varies from patient to patient. A similar appearance can be seen in the study on paraffin-embedded liver cancer tissues (Nakajima et al 2007). 3.2.  Absorption coefficient spectra of gastric tumors from different patients

Figure 4 displays the averaged terahertz absorption coefficient spectra of gastric tumors from other patients. The error bars showing the standard error of the averaged values represent the variation over different chosen points within a tumor sample. For each sample, there is high similarity between the terahertz absorption spectra measured at different chosen points. A small variation exists in the amplitude between these spectra and the maximum root mean square error of the differences within a sample for the five patients are all less than 7.8 cm−1. The good homogeneity of the terahertz absorption characteristic within a sample reflects similar constituents, which are sensitive to the terahertz radiation at different points. By comparing the terahertz absorption coefficient spectra in figure 4, it should be highlighted that tumor tissues from different patients have some spectral features in common, despite the differences in their sample information. There are two apparent absorption peaks in every spectrum of gastric tumors, which are, respectively, located in the range of 0.2~0.5 THz and 1~1.5 THz. An obvious concave point in the range of 0.5~1 THz is derived due to these two peaks. These appearances indicate that the gastric tumor tissues from different patients 5428

D Hou et al

Phys. Med. Biol. 59 (2014) 5423

(a)

50 normal tissue tumor tissue

45

Absorption Coefficent/cm−1

40 35 30 25 20 15 10 5 0 0.2

Normalized percentage difference in absorption

(b)

0.4

0.6

0.8

1 1.2 Frequency/THz

1.4

1.6

1.8

2

0.4

0.6

0.8

1 1.2 Frequency/THz

1.4

1.6

1.8

2

70 60 50 40 30 20 10 0 −10 −20 0.2

Figure 3. (a) The average absorption coefficient spectra, with standard error for gastric

normal tissue (blue) and tumor (red) from patient 4. (b) The percentage difference of the average absorption coefficient for normal and tumor tissue.

undergo a similar physiological change, which can be observed by the terahertz measurement during the process of carcinogenesis. In addition, the absorption peaks are likely to be attributed to some certain constituents in tumor tissues. A discussion of the distribution of the spectral features and their assignments will be presented in the following sections. These 5429

D Hou et al

Phys. Med. Biol. 59 (2014) 5423

60 Patient 1

Absorption Coefficent/cm−1

Absorption Coefficent/cm−1

60

40

20

0 0.2

0.4

0.6

0.8

1

1.2

1.4

1.6

1.8

Patient 2

40

20

0

2

0.5

Frequency/THz

1

1.5

2

1.5

2

Frequency/THz

Absorption Coefficent/cm−1

Absorption Coefficent/cm−1

60 80

Patient 3

60

40

20

0

0.5

1

Frequency/THz

1.5

2

Patient 5

40

20

0

0.5

1

Frequency/THz

Figure 4.  The average absorption coefficient spectra with the standard error of tumor

tissues from different patients.

visually identified spectral features lead to the assumption that the gastric tumor may have a terahertz ‘fingerprint’ absorption spectrum. The significantly higher absorption after 1 THz of the gastric tumor from patient 3 compared with those from other four patients is considered as the combined effect of two factors. One is the largest thickness of the tumor sample for patient 3, which results in fast attenuation of the measured signals in the frequency domain that influences the tendency of the absorption spectrum of the sample. The other factor is related to the tumor grade of patient 3. The content of some intrinsic biomarkers varies in the tumor tissue as the tumor grade progresses; for example, increased tryptophan metabolism via the serotonin pathway has been linked to malignant progression in breast cancer (Juhasz et al 2012, Kamson et al 2013). In addition, the tryptophan shows an absorption peak at 1.435 THz (Wahaia et al 2011). So, it is reasonable to speculate that the different cancer stages may also account for the differences in the amplitude of the terahertz absorption spectra between tumor tissues from different patients, as seen in figure 4. Further investigations to clarify this point are currently under way. 3. 3. Reproducibility of the measurements

To ensure the reliability of the measured THz electric signals, the reproducibility of the measurements was investigated. The same slice sample was measured repeatedly at different periods with two days interval and several slice samples from the same gastric tissue were tested. 5430

D Hou et al

Phys. Med. Biol. 59 (2014) 5423

90

6

3 60

0

Difference/cm−1

Absorption Coefficent/cm−1

the first period the second period difference

30 −3

0 0.2

0.4

0.6

0.8

1 1.2 1.4 Frequency/THz

1.6

1.8

−6 2

Figure 5.  The comparison of the averaged terahertz absorption coefficient spectra with

the standard error of the gastric tumor from patient 3, measured in different periods.

The experimental conditions, such as the temperature and the position of the sample and so on, were kept the same at each measurement. Both the absorption coefficient spectra of the same sample measured at different periods and the spectra of several slices from the same gastric tissue had good consistency. The trend of the average spectra and positions of the absorption peaks in the spectra for each gastric tissue were almost kept the same. The amplitude of the absorption coefficient spectra that was measured at different periods for the same sample changed a little, while a relatively larger variation was found in the amplitude of the absorption coefficient spectra for several slices from the same gastric tissue. A comparison of the terahertz absorption coefficient spectra of the gastric tumor from patient 3 measured at different periods is presented in figure 5. The maximum deviation of the two spectra is less than 6 cm−1 and the root mean square error of the differences across the working frequency range is 1.89 cm−1. These values are far less than the average absorption coefficient. Generally speaking, the reproducibility of the measurements in this study is satisfactory. 4.  Statistical analysis of spectral features To investigate the feature distribution contrast of the absorption coefficient spectra between normal and tumor tissue in detail, and also to make sure that the presented spectral features are specific to a gastric tumor, the statistical analysis was taken on the locations of the absorption peaks and the concave point. The spectral features and their appearance probability for different kinds of tissue are given in tables 2 and 3. The precise locations of the spectral features for the normal and tumor tissue from every patient are shown in figure 6. Solid and hollow symbols correspond to tumor and normal tissue. The different markers represent features in different frequency bands. 5431

D Hou et al

Phys. Med. Biol. 59 (2014) 5423

Table 2.  Features and their appearance probability in the absorption coefficient spectra of normal tissues.

No.

Feature

Range/THz

Position/THz

Probability (%)

1 2 3 4 5 6

Peak Peak Peak Peak Peak concave

0.240~0.336 0.456~0.553 0.769~0.818 1.273~1.442 1.634~1.827 0.624~0.890

0.288 0.528 0.793 1.418 1.682 0.745

60 56 40 36 28 20

Table 3.  Features and their appearance probability in the absorption coefficient spectra of tumor tissues.

No.

Feature

Range/THz

Position/THz

Probability (%)

1 2 3 4

Peak Peak Peak Concave

0.336~0.433 0.456 1.202~1.490 0.624~0.890

0.361 0.456 1.442 0.769

88 20 100 100

It can be seen that the absorption bands of normal gastric tissue samples are distributed loosely while the population for tumor tissue samples keeps good consistency. For normal tissue, there are few common absorption bands among the samples from different patients, which may be attributed to the diversity of human tissues. The spread of the first absorption peak before 0.6 THz, whose appearance probability is larger than 50% for normal tissues, appears less concentrated than for tumors. By contrast, in the respective ranges of 0.3~0.54 THz and 1.2~1.49 THz, broad characteristic absorption peaks are exhibited in every spectrum of the gastric tumor samples, which results in an apparent concave point at the range of 0.62~0.9 THz. These features have not been found simultaneously in any spectra of normal gastric tissues, which can be clearly recognized in figure 6. The combination of these features, especially the presence of the peak in the range of 1.2~1.49 THz and the derivative concave point among the range of 0.62~0.9 THz, provides a powerful tool for the identification of a gastric tumor, despite the variety of tissues from different patients. 5.  Tissue classification and identification of gastric tumors For the systematic identification of a gastric tumor, a proposed data analysis methodology on the terahertz absorption spectra of gastric tissue was carried out in this study. Principal component analysis (PCA) combined with a t-test was used to find the significant principal components indicating the best distinction between normal and cancerous gastric tissues. In addition, two kinds of general approach, K-means and support vector machine (SVM), have been tested to automatically categorize the processed terahertz data into normal and cancerous groups. 5.1.  PCA combined with t-test

To extract valuable information out of the multidimensional spectra, PCA combined with a t-test has been taken on the observed absorption coefficient spectra. By applying the singular value decomposition, the original absorption coefficient data matrix can be projected into a low dimensional principal component (PC) space, where PCs are set to make the largest variance between data points (Nakajima et al 2007). After the original data was transformed into 5432

D Hou et al

Phys. Med. Biol. 59 (2014) 5423

Figure 6.  The positions of the peaks (circle showing the peak before 0.6 THz, triangle

showing the peak after 1 THz) and concave point (square) in the absorption coefficient spectra for normal (hollow, n = 5) and tumor (solid, n = 5) tissue from each patient.

Table 4.  P-values of the PCs with most significant difference.

PC

2

3

5

7

6

13

P-value

0.0000

0.0004

0.0064

0.0468

0.1266

0.3642

the score matrix, a t-test was used to evaluate the degree of difference between normal and abnormal tissues for individual PCs. The PCs were then sorted by increasing the P-values, and those with the most significant difference were chosen to categorize the samples into different types. P 

Terahertz spectroscopic investigation of human gastric normal and tumor tissues.

Human dehydrated normal and cancerous gastric tissues were measured using transmission time-domain terahertz spectroscopy. Based on the obtained terah...
1MB Sizes 0 Downloads 4 Views