| Accurate diagnosis in early stage of colorectal cancer and injury identification during chemotherapy possess significant research value and promising application prospect.At present,traditional clinical technologies of cancer diagnosis have problems,such as invasive detection in tissue biopsy,high false negative rate in early-stage diagnosis when using videography and endoscopy,and insufficient specificity in blood examination for patients with mild chemotherapy injury,which adversely affect the treatment and the prognosis of cancer patients.Whereas,terahertz(THz)spectroscopy,which benefits from its being non-destructive and needless of label or reaction substrate,has attracted extensive attention and has been substantially studied in various fields including security inspection,structural defect inspection and biological lesion detection,especially in cancer diagnosis and chemotherapy injury identification.In this dissertation,a series of methods based on machine learning is proposed for colon tumor detection and chemotherapy injury identification using THz time-domain spectroscopy(THz-TDS).The primary works and innovations of this dissertation are as follows:(1)A qualitative method for colon cancer cell line detection based on THz-related characteristics in hydration is proposed.The sensitivity of three THz frequency-domain parameters,including absorption coefficient,refractive index and dielectric loss tangent,in the detection is investigated by comparing the THz response of normal cell lines and colon cancer cell lines within a certain cell concentration range.The results indicate that it is the difference in picosecond-scale hydration kinetics between the cell lines that leads to their distinct THz responses.In culture medium environment,the absorption coefficient shows a strong correlation with colon cancer cell lines.The model using it as the input feature gives better prediction of 93.8% accuracy.This method confirms the feasibility of using terahertz spectroscopy for cancer cell detection.(2)A classification method of cancer cell lines based on terahertz-related pathological information is proposed.Similarity and difference of the frequencydomain characteristics among cell lines at six different cell concentrations are studied.Feature fusion of terahertz-sensitive parameters is implemented for mining effective information from the samples,maximum information coefficient(MIC)is used to select the characteristics with strong pathological correlation,and the optimal feature dimension of the modeling is explored.Based on these studies,a classification model of colon cancer cell lines is constructed.The research shows that some stronglycorrelated characteristics match the terahertz fingerprint responses of certain biomarkers in the colon cancer cells,such as tryptophan,leucine,and threonine,enhancing the specificity and interpretability of this model,which reaches an accuracy of 80.1% in the validation dataset.This study provides the theoretical and experimental fundament for qualitative and quantitative detection of cell lines based on terahertz spectroscopy.(3)An evaluation method of liver injury due to chemotherapy of colon cancer based on extraction of specific principal components in terahertz spectrum is proposed.Outliers resulting from uneven sample surface and non-uniform illumination are filtered using time-domain parameters like peak-peak value and time delay.Principal component analysis(PCA)is implemented to extract features from frequency-domain terahertz spectrum and avoid the influence from their collinearity.Three models that employ Adaboost,random forest,and support vector machine,respectively,are used to evaluate liver injury and their performance are compared.It is shown that terahertz wave is capable of distinguish different degrees of liver injury based on their difference in water content of liver and in structure and metabolites of the injured parts;and Adaboost gives better desirable result by overweighting the misclassified samples.This study verifies the feasibility of to use terahertz technology for the detection of liver tissue injury.(4)A detection method of mild liver injury tissue based on terahertz-related pathological information of injured tissue is proposed to overcome the lack of clinical approaches for this purpose.The time-delay parameter is used for data preprocessing,the maximum information coefficient is used for deep mining of THz frequencydomain characteristics of residual chemotherapeutic drugs and abnormal metabolites,and the optimal feature dimension is selected.An evaluation model is constructed using Adaboost,with t-distributed stochastic neighbor embedding applied in feature extraction to facilitate the performance when strongly-correlated local features carry crucial characteristics of the sample.Meanwhile,in order to explore the terahertz characteristic response of the chemotherapeutic drug 5-FU,its molecular and crystal structure are simulated and are compared with the experiment results.Good matching between some strongly-correlated features and the terahertz fingerprint responses of certain abnormal metabolites in injured liver tissue,including glutathione,proline,Lcystine and 5-FU,are found,showing specificity and considerable potential of this method in mild liver injury detection.In summary,this dissertation,having its subject based on practical needs in medical detection and supported by Cancer Institute of Zhejiang University School of Medicine,studies the terahertz response of colon cell lines,both normal and cancerous,as well as injured tissues in chemotherapy of colorectal cancer.It explores the feasibility of analyzing biomedical lesions from sample’s terahertz spectrum and introduces a highly self-adaptive and generalizable method for lesion detection based on terahertzrelated pathological information.This work offers innovative ideas and techniques for future development of terahertz-based detection of biomedical lesions. |
