Open Access
Add to BibSonomy
Abstract
At present, researchers are exploring biological tissue detection method using terahertz techniques. In this paper, techniques to inspect mouse liver injury by using terahertz spectroscopy were studied. The boxplots were applied to remove abnormal data, and the maximal information coefficient was employed to select crucial features from the absorption coefficient and refractive index spectra. Random Forests and AdaBoost were applied to recognize different levels of liver injury. We found that AdaBoost had better performance on low-level injury classification. This work suggests that terahertz techniques have the potential to detect liver injury at an early stage and evaluate liver treatment strategies.
© 2019 Optical Society of America under the terms of the OSA Open Access Publishing Agreement
Corrections
16 September 2019: A typographical correction was made to Fig. 1.
1. Introduction
The incidence of colorectal cancer has been increasing worldwide each year. Colorectal cancer has become a common malignant tumor disease since people's lifestyle and eating habits have changed in recent years with the economy growth, especially in developing countries [1]. At present, chemotherapy is a common treatment for colorectal cancer, which can improve the symptoms of patients and the clinical treatment effect. However, researches show that almost all chemical drugs used in cancer treatment can cause acute and chronic liver injury [2]. Especially, chemotherapeutic drugs can inevitably cause liver injury during the treatment, which called drug-induced liver injury (DILI). According to statistics from the World Health Organization (WHO), the anti-tumor drug ranks 5th in drug most frequently causing DILI [3,4], the anti-tumor drugs rank 2nd in acute liver failure caused by DILI [5]. The chemotherapy is often discouraged due to the effects of DILI which can affect the efficacy and cause a significant reduction in the survival of patients. Patients may suffer from liver failure or even death.
As the most important metabolic and detoxifying organ in the human body, liver is responsible for the metabolism and the removal of most chemotherapy drugs. Hence, liver injury caused by anti-tumor drugs will not have influence on the therapeutic effect on tumor. Also, this injury affects the patient's survival rate, and brings huge economic burden to patients and their families. Moreover, the anti-tumor drug toxicity is still the main factor limiting the application of drugs. It is usually hard to add human intervention to reduce or control liver injury since it is already medium or high level injury when it can be detected. So far, the blood test is the general medical test for liver injury. Although the blood test is convenient for patients, it is generally difficult to detect low-level liver injury. This is a problem that cannot be ignored in cancer treatment.
5-Fluorouracil (5-Fu) is widely used in the treatment of cancer, especially colorectal cancer [6–8]. 5-Fu is a metabolic drug which is converted to 5-fluorouracil deoxynucleoside in the body. The growth of cancer cells can be prevented by 5-fluorouracil deoxynucleoside which can inhibit thymidine nucleotide synthetase and DNA synthesis. The synthesis of DNA and ribosomal RNA exerts an anti-tumor effect. The drug has a killing effect on various stages of proliferating cells, and has a small molecular weight, good penetrating power to the interstitial space and cell membrane, and easy to penetrate the surface tissue of the tumor cells. However, 5-Fu also has strong toxic side effects on liver tissue as well as killing tumor cells [9]. Thus, in this paper, 5-Fu was used for injured liver sample preparation.
THz wave is an electromagnetic wave in the electromagnetic spectrum of 0.1-10 THz (1 THz = 1012 Hz), and its corresponding wavelength is 30-3000 μm. The frequency band has many unique features which make it a promising tool for biomedical applications. THz is harmless to biological samples since it is low-energy and non-ionizing. Also, it can provide a direct measurement of the field amplitude and phase, which are related to the optical and dielectric properties of the samples [10]. Terahertz wave is sensitive to polar molecules, especially to biomacromolecules as the vibration and rotation spectrum of ceramics and many macromolecules lie within this frequency band. Based on these unique properties, THz spectroscopy has broad application prospects in fields such as nondestructive testing and medical diagnostics [11–16].
In addition, terahertz technology has become a critical nondestructive testing method for biological tissues in the last decade [17–21]. Zhang et al. applied composite multiscale entropy (CMSE) method to identify fresh porcine skin and muscle tissues based on terahertz signals and better performance was verified [17]. Fresh rat glioma tissue samples were detected using a reflection terahertz time-domain spectroscopy system. Both the refractive index and absorption coefficient of tumor tissues were higher than those of normal tissues as recently described by Yamaguchi et al [18]. Also, the potential of terahertz imaging for breast tissue recognition within the underexplored 0.3–0.6 THz range was investigated. The dielectric response could provide contrast for breast tissue identification within this frequency range [19]. As liver is considered relatively homogeneous in terms of tissue optical properties in microscale, research on liver tissue inspection using THz spectroscopy has increased [20–22]. Since dealing with fresh tissues in terahertz biomedical experiments has traditionally been a challenging problem, Png et al. investigated the influence of terahertz properties on rat liver tissues with different water contents [21]. Furthermore, Sy et al. reported their measurements of healthy and cirrhotic liver tissues using terahertz reflection spectroscopy [22].
With this in mind, we applied terahertz time-domain spectroscopy (THz-TDS) to mouse liver injury recognition that could provide a method for low-level liver injury detection. In this study, an identification method to classify mouse liver tissues into different injury groups by employing THz-TDS was proposed. The statistical method, boxplots were used to remove outliers, and the distinct frequency points from both absorption coefficient and refractive index spectra were selected by the maximal information coefficient (MIC) method. In addition, the relationships between the absorption and refractive index features and different levels of mouse liver injury were analyzed by Random Forests (RF) and AdaBoost. The results suggested that this methodology could have the potential to achieve automatic discrimination of mouse injured liver tissues and then to identify low-level injury.
2. Experiment
2.1 THz-TDS measurement
In this paper, we applied the commercial THz-TDS system for the experiments, which was developed by Zomega Terahertz Corporation in USA. The details of this system were previously reported in our earlier work [23]. To do the experiments, we designed a sample holder showed in Fig. 1. There were a circular holes with a radius of 1.0 and 0.5 cm opened on the plastic sheet and the copper sheet respectively. The size of the hole was a little bigger than the spot size of our THz beam since we didn’t want to cut any part of THz wave. The copper sheet was used for controlling the THz wave spotted on the tissue slices whose size were about 0.5 cm radius. In addition, the sample slices were placed on the plastic wrap for easy placement and movement of sample slices. Also, we had tested that almost all THz waves could come through the plastic wrap. The THz experiments were conducted at room temperature (20°C) and the room humidity was controlled at about 2%. An available frequency range of 0.1 to 2.0 THz was obtained, though the effective range tended to be around 0.2-1.0 THz due to signal attenuation by the sample.
2.2 Sample preparation
In this study, the chemotherapeutic drug 5-FU was purchased from Shanghai Xudong Haipu Pharmaceuticals Co.,Ltd (Shanghai, China). It was diluted in 0.9% sterile saline. All drugs were prepared immediately prior to use. The 25 Female BALB/C nude mice at a 5-week age were purchased from Shanghai SLAC Laboratory Animal Co.,Ltd (Shanghai, China). Animals were maintained under constant temperature and humidity and 12 hours of light and dark cycle with food and water available ad libitum. All animal experiments were performed according to the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. They were randomly divided into five groups after one week of adaptation. Since there were two mice that died from other factors, 23 mice were tested. The number of mice divided into the five groups were varied from 3 to 6. The placebo group, 0.9% saline was given by intraperitoneal injection once a day for five consecutive days. The other four groups were intraperitoneal injected 25, 50, 75, and 100 mg/kg of 5-Fu per body weight, once daily, for five consecutive days. After three weeks, the mice were respectively sacrificed in individually ventilated cages with carbon dioxide. One week and three weeks after the last medication, blood was taken from the orbit of the mice and serums were obtained by centrifugation after overnight fasting. Serum ALT and AST were determined by HITACHI 3100 automatic analyzer series (Hitachi, Ltd., Tokyo, Japan). The serum ALT and AST values of five groups after three weeks of 5-Fu injection were shown in Table 1. Since the injected dose 50 mg/kg was too toxic for mice, this resulted in the mice of 50 mg/kg, 75 mg/kg and 100 mg/kg groups not surviving three weeks. The serum ALT and AST values of the three groups were used by the test data before death. From the results of the blood tests, the liver tissues injected with different 5-Fu concentrations were verified to be at different levels of liver injury. Liver tissues were collected and stored at −80 °C.
Table 1. Serum ALT and AST values of five groups after three weeks of 5-Fu injection.
Furthermore, liver tissues were cut into slices by a cryotome. We obtained 8-15 sample slices for each tissue. The thickness of each liver tissue slice was all approximately 350 μm. The slices were used in terahertz spectrum detection. Our experiments had been conducted for 3 days. We tested 1-3 points on each tissue slice and each point was tested twice to make sure of the reproducibility of the measurements. The number of spectra from the five groups were 64, 62, 114, 88, and 50, respectively. In total, 378 spectra from tissue slices were studied and analyzed.
2.3 Methodology
2.3.1 Liver injury recognition method
Aiming to explore the identification results on mouse injured liver tissue, we proposed a recognition algorithm as shown in Fig. 2. For preliminary data processing, the statistical technique boxplot was applied to remove outliers for terahertz time-domain spectra of each liver tissue slice first since there were some abnormal spectra caused by experimental mistakes and systematic instability. To delete the most abnormal datapoints, two parameters which were $\Delta E$ and the time-delay $\Delta t$ of the peak.
These two parameters were selected because they were both critical characteristics of THz-TD spectra, which had been verified to be the important features for biological sample recognition [24]. Then, the time-domain spectrum was transferred to absorption coefficient and refractive index by Fast Fourier Transform (FFT). As for feature extraction process, distinct frequency points of both absorption coefficient and refractive index were selected by the MIC method. After that, PCA method was performed to extract crucial features of the dataset, and the first five principal components were taken as the data features. Finally, the data features were used in RF and AdaBoost classifiers. In this study, the machine learning classification methods RF and AdaBoost were compared to differentiate five injured liver groups.
2.3.2 Classification analysis by RF
The RF algorithm is now a widely used method for regression and classification problems [25–26]. Since RF has great generalization ability and is suitable for small size dataset classification, we considered the use of this method to identify and detect liver tissues with different injury levels. RF algorithm randomly extracts an N size dataset from the total dataset D by means of bootstrap at first. Then, using the best split mode, a decision tree is constructed without pruning. In this paper, the Gini impurity which was applied to select the best feature for each node of the decision tree was calculated as Eq. (1),
where $gini(D)$ was the Gini impurity of the dataset D, n denoted the number of groups of samples.${p_i}$ denoted the probability of group i data in the dataset D. After that, the two previous steps are repeated m times to construct the RF [27]. The output of RF is the type with the highest average probability value among all decision trees. The probability value is shown as follows: One of the characteristics of RF is that the relative weight of each distinct frequency point could be given, and the contribution of the distinct frequency points for liver tissue recognition can be expressed to some extent. RF is a nonlinear integrated machine learning model composed of multiple decision trees, and the introduction of random feature selection during the training process of the decision tree can enhance the independence between each tree. This will increase the generalization of the model.2.3.3 Classification analysis by AdaBoost
The AdaBoost algorithm is a classic adaptive ensemble learning algorithm. The basic principle is to integrate weak learners to generate a highly learnable model to achieve a significant improvement in classification results. At present, the algorithm has been successfully applied to applications such as pattern recognition and regression prediction [28–29].
Initially, each data in the training set is given the same weight, and a basic weak learner f1 is trained and generated under the original weight distribution. An initial weight w1 is assigned to the weak learner according to the training result. The training data of the subsequent weak learner f2 depends on the classification result of f1, that is, the data that has been erroneously classified by f1 would appear in the new weak learner with a larger weight. As a result, the weak learner f2 is also given a weight w2. Therefore, the algorithm trains and generates a plurality of weak learners by repeatedly calling the training data weight adjustment. The final recognition result can be obtained according to the weighted fusion method, showing as follows.
2.3.4 Distinct frequency points identification using MIC
We attempted to find about the features with high relevance to improve the recognition rate of liver injury in the terahertz region. We considered this analysis method because it has the ability to identify relationships between two variables [30–31]. Especially, in this paper, MIC was able to find the correlation between the frequency points of absorption coefficient and refractive index spectra with the injury level of the mouse liver tissues. MIC is based on the idea that if a relationship exists between two variables, then a grid can be drawn on the scatterplot of the two variables that partition the data to encapsulate that relationship. Let variable X denote a frequency point of the data and variable Y denote the level of liver injury, in MIC, the ordered X values and Y values are divided into a bins and b bins, respectively, which results in an a-by-b grid G. The distribution of the values in X-Y space located in the cells of G is denoted as (X,Y)|G. Different grid partitions lead to different distributions. The statistic MIC is the maximum value of the characteristic matrix M(X,Y) defined as follows:
Where $\max I((X,Y){|_G})$ represents the maximal mutual information of $(X,Y){|_G}$ over all possible grids. The MIC of X and Y are shown as Eq. (5), where $n$ is the number of samples, $B(n)$ imposes an upper bounds on the sizes of G for searching the MIC value. In this paper, $B(n) = {n^{0.6}}$ was used. A larger value of MIC meant this feature was more relevant to liver injury tissues.3. Results and discussion
3.1 Results for preliminary data processing
Figure 3 showed the average time domain spectra of two liver tissues injected with 75 and 50 mg/kg 5-Fu with error bars, respectively. Since the size of each liver tissue was slightly different, the number of tissue slices and test number of each tissue would vary. Figure 3(a) represented the average THz time domain spectra of the 4th liver tissue in 75 mg/kg group which had no clear abnormal spectrum, while the average spectra of the5th liver tissue in the 50 mg/kg group were shown in Fig. 3(b) along with two abnormal spectra. There were two main reasons for them. First, they were due to the incorrect placement of the sample slices during the experiments, resulting in no corresponding tissue pieces being detected. Secondly, the presence of air bubbles on the sample slices caused an uneven illumination plane. Each time domain spectrum was used as a rat liver tissue corresponding to the liver injury groups. Therefore, it was necessary to remove abnormal spectral data. In Fig. 3(a), it could be seen that each spectrum had a relatively similar structure and trend, and the peak value and position had good consistency as well. This indicated that all the experimental points of this tissue slice had little deviation and could be taken as valid data. However, Fig. 3(b) was taken as an example where the peak value of two spectra were significantly lower, and the peak positions were different from most of the remaining spectra.
Fig. 3. The average terahertz time domain spectra of two liver tissues. (a) The average terahertz time domain spectra of the 4th liver tissue in 75 mg/kg group without clear abnormal data. (b) The average terahertz time domain spectra the 5th liver tissue in the 50 mg/kg group with clear abnormal data.
Hence, the boxplot was introduced to remove the outliers. Figure 4 showed the boxplots of the outlier data recognition for all tissues in 25 mg/kg and 50 mg/kg groups according to the parameters $\Delta E$ and time-delay $\Delta t$, respectively. The x-axis represented the tissue number at the drug injected concentration, which represented we obtained four and six valid tissues in 25 mg/kg and 50 mg/kg groups, respectively. To improve the detection efficiency of outliers, these two parameters were introduced. The outlier removal was first performed on the time domain spectra of each tissue according to the parameter $\Delta E$, and then the boxplot would be conducted on the remaining data according to the parameter $\Delta t$. In addition, it was mentioned in the literature [24] that these two parameters were crucial parameters in biological detection and had high sensitivity to biological samples. However, the outliers found in the first step was significantly less than the number of outliers detected when the parameter $\Delta t$ was employed. The slight differences of the tissue composition corresponding to each tested point of the same tissue had greater impact on the peak-to-peak value. Thus, the range of the boxplots were larger than that of the boxplots from $\Delta t$, which caused the parameter $\Delta t$ to have better performance on the removal of outliers. According to the parameter $\Delta E$, the shape of the box diagram of several tissues was similar to the 3rd tissue of 25 mg/kg group, that is, the range was relatively high, with no outliers were recognized. According to $\Delta t$, the range of each tissue's box diagram was greatly reduced, and the outliers were far from the box. The main reason was because our experiments had been carried out for 3 days, and the daily reference spectrum would be slightly different due to environmental factors. $\Delta t$ had to be derived from related reference spectrum, so this parameter was dependent and could effectively eliminate the outliers introduced by external interference. After two boxplot processes, a total of 38 abnormal data were removed. 10 outliers were deleted according to the peak-to-peak value, while 28 abnormal data were removed by the parameter $\Delta t$.
Fig. 4. The represent boxplots showing the abnormal spectra in each tissue in one group. (a) The boxplot result of 25 mg/kg group with the parameter $\Delta E$. (b) The boxplot result of 50 mg/kg group with the parameter $\Delta t$.
Figure 5 showed the absorption coefficient and refractive index spectra of the 2nd tissue in 100 mg/kg group obtained after FFT process. Figures 5(a) and (c) were the absorption coefficients and refractive indices of original data. While Figs. 5(b) and (d) represented the absorption coefficient and refractive indices with outlier removal process. In Fig. 5(a), there were two absorption coefficient spectra that were significantly higher than most of the spectra, and in Fig. 5(c) there were several data having higher refractive indices than 3.0 or close to 2.0 at 0.2 THz. These could basically be determined as abnormal data. After the preliminary data processing, most of the data with conspicuous error had been removed, while the absorption coefficient and refractive index fluctuation range of the remaining data spectra were reduced, and the fluctuation trend of the spectra was similar.
Fig. 5. Representative absorption coefficient and refractive index spectra of the 2nd tissue in 100 mg/kg group. (a) Absorption coefficient spectra without boxplot process. (b) Absorption coefficient spectra with boxplot process. (c) Refractive index spectra without boxplot process. (d) Refractive index spectra with boxplot process.
3.2 Results for feature extraction
The average absorption coefficient and refractive index spectra of five groups were in Fig. 6. Although the error bars of each group overlapped part of the range and became more and more coincident as the frequency increased in Figs. 6(a) and (b), it could still be seen from Fig. 6(a) that the absorption coefficient of 0 mg/kg group was generally the lowest. And as the level of injury increases, the amplitude of the absorption coefficient also increased. It’s worth noting that in Fig. 6(b), the 100 mg/kg group had the greatest refractive index. The relationship between other groups were better differentiated in low frequency band. Figures 6(c) and (d) represented the average spectra without error bars. The relationship between liver tissue and terahertz optical parameters with different levels of injury could be easily seen. In Fig. 6(d), the trend of average refractive index spectra of each group was consistent with the five groups. The spectra of 0 mg/kg and 25 mg/kg groups were almost the same since the liver injury of the tissues in the 25 mg/kg group was relatively low which could be verified in Table 1. The original features used for tissue group classification were 100, including 50 features for both absorption coefficient and refractive index in 0.2-0.8 THz. All features were normalized before dimensional reduction and distinct feature extraction. The 40 features with more correlation were then extracted by the MIC technique, which were mainly in the range of 0.2-0.6 THz in the absorption coefficient spectrum and 0.2-0.3 THz in the refractive index spectrum. These coincide to the overlap of the spectra in the corresponding frequency range in Figs. 6(c) and (d).
Fig. 6. Average spectra of both absorption coefficient and refractive index of five groups. (a) Average absorption coefficient spectra of each group with error bars. (b) Average refractive index spectra of each group with error bars. (c) Average absorption coefficient spectra of each group. (d) Average refractive index spectra of each group.
3.3 Results for tissue group identification
In this study, we applied two classification algorithms, namely RF and AdaBoost. First, in order to verify the feasibility of the proposed method, the data of injection drug concentration 0 mg/kg, 50 mg/kg and (75 + 100) mg/kg were selected into three groups, namely the control group, medium-dose group, and high-dose group. The level of liver injury of the 25 mg/kg group was relatively small compared with the control group, and the 50 mg/kg group had clear liver injury. The 75 mg/kg and 100 mg/kg groups were combined because the discrimination between the two concentrations was remarkable, and it could increase the amount of data of high-dose group. Therefore, we considered to classify and identify these three groups first. Figure 7 showed the confusion matrix of three groups’ classification. The x-axis suggested the predicted group of the data, and y-axis indicated the true group of the data. The recognition results by RF were shown in Figs. 7(a)–(c), and Figs. 7(d)–(f) were the results using AdaBoost. The features applied in the first column was the original 100 features after preprocessing. PCA was used to reduce the dimension of the original dataset and the result was shown in the second column. The third column indicated the result with the dataset processed by MIC and PCA. In the confusion matrix, the percentages on the diagonal indicated that the truly predicted percentage of data, and the other grids represented the percentage of data that were assigned to the wrong groups. It could tell that according to the different feature extraction methods, the two algorithms mainly improve the recognition rate of the control group and medium concentration groups. Both RF and AdaBoost could reach high precision. The specific classification results were shown in Table 2.
Fig. 7. Confusion matrix of three groups’ classification. (a) Classification results using RF with original features. (b) Classification results using RF with PCA features. (c) Classification results using RF with MIC + PCA features. (d) Classification results using AdaBoost with original features. (e) Classification results using AdaBoost with PCA features. (f) Classification results using AdaBoost with MIC + PCA features.
Table 2. Classification results of three groups based on different features.
Then, RF and AdaBoost methods were used to detect the five liver injury groups. The classification confusion matrix and the specific results were shown in Fig. 8 and Table 3, respectively. It could be clearly seen from the confusion matrix that the results of PCA after preprocessing were significantly better than those of the first column, whether from low-concentration groups or high-concentration groups. In addition, the results of data after MIC and PCA processed were significantly improved compared with the second column. As for RF, it could be seen in Figs. 8(b) and (c) that the algorithm had a better recognition effect on injured tissues of 25 mg/kg and 100 mg/kg groups. Comparing with In Figs. 8(e) and (f), obviously, the overall identification effect by AdaBoost was more effective than RF. However, the liver tissue in the 25 mg/kg group had a higher probability of misclassification. In summary, from the results in Table 3, it could be clearly seen that for the identification of mouse liver tissues with different injury levels, the AdaBoost algorithm yielded better results than RF, especially after (MIC + PCA) process. This might be because it could recognize more low-level injury correctly. The identification accuracy could achieve close to 90%.
Fig. 8. Confusion matrix of five groups’ classification. (a) Classification results using RF with original features. (b) Classification results using RF with PCA features. (c) Classification results using RF with MIC + PCA features. (d) Classification results using AdaBoost with original features. (e) Classification results using AdaBoost with PCA features. (f) Classification results using AdaBoost with MIC + PCA features.
Table 3. Classification results of five groups based on different features.
4. Conclusion
In this paper, we proposed a liver injury level identification method which was aimed at identifying five liver injury levels. We employed the boxplot method to search for abnormal sample spectra to achieve data cleaning. Based on the MIC method, the distinct frequency points of both absorption coefficient and refractive index spectra were investigated to improve the recognition rate of liver injury. And the first 40 distinct features were selected from both spectra though most of the features were from low frequency range of absorption coefficient spectra, which suggested that absorption coefficient could be more sensitive with liver injury. Moreover, RF and AdaBoost methods had similar recognition rate at about 92% for three groups’ classification. However, AdaBoost showed better performance for the five groups’ classification since it could be more sensitive to low-level liver injury. The reason for this might because AdaBoost paid more attention on the samples which were misclassified. The proposed method might be promising for improve the reliability of mouse lesion tissue inspection and even for human liver tissues. It was able to promote the development of terahertz technology in clinical applications as well. Yet, although a satisfactory recognition rate was reached by current model, it still requires more samples to be trained to improve its feasibility.
Funding
National Natural Science Foundation of China (61473255, 61873234); Fundamental Research Funds for the Central Universities (2019FZA5007); National Human Genetic Resources Sharing Service Platform (2005DKA21300).
Acknowledgments
We would like to acknowledge the members in Cancer Institute of the Second Affiliated Hospital for sample preparation.
Disclosures
The authors declare that there are no conflicts of interest related to this article.
References
1. Catherine A. O’Brien, A. Pollett, S. Gallinger, and J. E. Dick, “A human colon cancer cell capable of initiating tumour growth in immunodeficient mice,” Nature 445(7123), 106–110 (2007). [CrossRef]
2. Y. Zhou, L. Yang, Z. Liao, X. He, Y. Zhou, and H. Guo, “Epidemiology of drug-induced liver injury in China: a systematic;analysis of the Chinese literature including 21 789 patients,” Eur. J. Gastroenterol. Hepatol. 25(7), 825–829 (2013). [CrossRef]
3. W. Bernal and J. Wendon, “Acute liver failure,” N. Engl. J. Med. 369(26), 2525–2534 (2013). [CrossRef]
4. E. Björnsson and R. Olsson, “Suspected drug-induced liver fatalities reported to the WHO database,” Dig. Liver Dis. 38(1), 33–38 (2006). [CrossRef]
5. M. Petronijevic and K. Ilic, “Associations of Gender and Age with the Reporting of Drug-Induced Hepatic Failure: Data from the VigiBase,” J. Clin. Pharmacol. 53(4), 435–443 (2013). [CrossRef]
6. A. Schalhorn and M. Kühl, “Clinical pharmacokinetics of fluorouracil and folinic acid,” Semin. Oncol. 19(2 Suppl 3), 82 (1992).
7. B. V. Triest, C. J. V. Groeningen, and H. M. Pinedo, “Current chemotherapeutic possibilities in the treatment of colorectal cancer,” Eur. J. Cancer 31(7-8), 1193–1197 (1995). [CrossRef]
8. D. B. Longley, D. P. Harkin, and P. G. Johnston, “5-Fluorouracil: mechanisms of action and clinical strategies,” Nat. Rev. Cancer 3(5), 330–338 (2003). [CrossRef]
9. T. Aloia, M. Sebagh, M. Plasse, V. Karam, F. Levi, S. Giacchetti, D. Azoulay, H. Bismuih, D. Castaing, and R. Adam, “Liver Histology and Surgical Outcomes After Preoperative Chemotherapy With Fluorouracil Plus Oxaliplatin in Colorectal Cancer Liver Metastases,” J. Clin. Oncol. 24(31), 4983–4990 (2006). [CrossRef]
10. N. Bajwa, S. Sung, D. B. Ennis, M. C. Fishbein, B. N. Nowroozi, D. Ruan, A. Maccabi, J. Alger, M. A. St. John, W. S. Grundfest, and Z. D. Taylor, “Terahertz Imaging of Cutaneous Edema: Correlation with Magnetic Resonance Imaging in Burn Wounds,” IEEE Trans. Biomed. Eng. 64(11), 2682–2694 (2017). [CrossRef]
11. H. Chen, S. Ma, X. Wu, W. Yang, and T. Zhao, “Diagnose human colonic tissues by terahertz near-field imaging,” J. Biomed. Opt. 20(3), 036017 (2015). [CrossRef]
12. H. Cheon, H. Yang, S. H. Lee, Y. A. Kim, and J. H. Son, “Terahertz molecular resonance of cancer DNA,” Sci. Rep. 6(1), 37103 (2016). [CrossRef]
13. K. I. Zaytsev, K. G. Kudrin, V. E. Karasik, I. V. Reshetov, and S. O. Yurchenko, “In vivo terahertz spectroscopy of pigmentary skin nevi: Pilot study of non-invasive early diagnosis of dysplasia,” Appl. Phys. Lett. 106(5), 053702 (2015). [CrossRef]
14. A. J. Fitzgerald and E. Al, “Terahertz pulsed imaging of human breast tumors,” Radiology 239(2), 533–540 (2006). [CrossRef]
15. C. B. Reid, A. Fitzgerald, G. Reese, R. Goldin, P. Tekkis, P. S. O. Kelly, E. Pickwell-MacPherson, A. P. Gibson, and V. P. Wallace, “Terahertz pulsed imaging of freshly excised human colonic tissues,” Phys. Med. Biol. 56(14), 4333–4353 (2011). [CrossRef]
16. Y. Cao, P. Huang, X. Li, W. Ge, D. Hou, and G. Zhang, “Terahertz spectral unmixing based method for identifying gastric cancer,” Phys. Med. Biol. 63(3), 035016 (2018). [CrossRef]
17. R. Zhang, Y. He, K. Liu, L. Zhang, S. Zhang, E. Pickwell-MacPherson, Y. Zhao, and C. Zhang, “Composite multiscale entropy analysis of reflective terahertz signals for biological tissues,” Opt. Express 25(20), 23669 (2017). [CrossRef]
18. S. Yamaguchi, Y. Fukushi, O. Kubota, T. Itsuji, T. Ouhci, and S. Yamamoto, “Brain tumor imaging of rat fresh tissue using terahertz spectroscopy,” Sci. Rep. 6(1), 30124 (2016). [CrossRef]
19. C. Quentin, A. Amel, M. Laven, H. Philipp, G. Janusz, M. Gaëtan, Z. Thomas, G. Jean-Paul, and M. Patrick, “Pilot study of freshly excised breast tissue response in the 300–600 GHz range,” Biomed. Opt. Express 9(7), 2930 (2018). [CrossRef]
20. M. Singh, S. Singh, and S. Gupta, “A New Quantitative Metric for Liver Classification from Ultrasound Images,” IJCEE 4(4), 605–607 (2012). [CrossRef]
21. G. M. Png, J. W. Choi, B. W. Ng, S. P. Mickan, D. Abbott, and X. Zhang, “The impact of hydration changes in fresh bio-tissue on thz spectroscopic measurements,” Phys. Med. Biol. 53(13), 3501–3517 (2008). [CrossRef]
22. S. Sy, S. Y. Huang, and Y. Xiang, “Terahertz Spectroscopy of Liver Cirrhosis: Investigating the Origin of Contrast,” Phys. Med. Biol. 55(24), 7587–7596 (2010). [CrossRef]
23. D. Hou, X. Li, J. Cai, Y. Ma, X. Kang, P. Huang, and G. Zhang, “Terahertz spectroscopic investigation of human gastric normal and tumor tissues,” Phys. Med. Biol. 59(18), 5423–5440 (2014). [CrossRef]
24. L. H. Eadie, C. B. Reid, A. J. Fitzgerald, and V. P. Wallace, “Optimizing multi-dimensional terahertz imaging analysis for colon cancer diagnosis,” Expert Systems with Applications 40(6), 2043–2050 (2013). [CrossRef]
25. L. Breiman, “Random Forests,” Machine Learning 45(1), 5–32 (2001). [CrossRef]
26. L. R. Iverson, A. M. Prasad, S. N. Matthews, and M. Peters, “Estimating potential habitat for 134 eastern US tree species under six climate scenarios,” For. Ecol. Manage. 254(3), 390–406 (2008). [CrossRef]
27. A. Cutler, D. R. Cutler, and J. R. Stevens, “Random forests,” Machine Learning 45(1), 157–175 (2011). [CrossRef]
28. S. Avidan, “Ensemble Tracking,” IEEE Trans. Pattern Anal. Machine Intell. 29(2), 261–271 (2007). [CrossRef]
29. J. Zhu, A. Arbor, and T. Hastie, “Multi-class AdaBoost,” Statistics & Its Interface 2(3), 349–360 (2009). [CrossRef]
30. D. N. Reshef, Y. A. Reshef, H. K. Finucane, S. R. Grossman, G. McVean, P. J. Turnbaugh, E. S. Lander, M. Mitzenmacher, and P. C. Sabesi, “Detecting Novel Associations in Large Data Sets,” Science 334(6062), 1518–1524 (2011). [CrossRef]
31. Y. Sun, P. Du, X. Lu, P. Xie, Z. Qian, S. Fan, and Z. Zhu, “Quantitative characterization of bovine serum albumin thin-films using terahertz spectroscopy and machine learning methods,” Biomed. Opt. Express 9(7), 2917 (2018). [CrossRef]
