| Lung cancer is a kind of malignant tumor with high mortality.Non-Small Cell Lung Cancer is a subtype of lung cancer.According to the epidemiological reports,NSCLC patients consist of 80% of lung cancer population.Comparing with the Small Cell Lung Cancer,the diffusion velocity of NSCLC is slower,but due to the high variability between tumor cells of across NSCLC patients,there are significant individual-specific differences in the development of the cancer.Statistical report shows that a large number of patients with NSCLC have not been able to receive appropriate treatment in time due to the lack of reliable predictor,and resulted in the high mortality rate of up to 75% in patients with NSCLC.Therefore,effective patients' survival time prediction models are needed in the disease treatment and review strategy selection and for the improvement of cure or survival rate of NSCLC patients.Radiomics is a new research field in medical science,its generation is inseparable with the good performance of radio-genomics in disease study and the great potential of medical imaging in the disease diagnosis and treatment.The radiomics mapped the information of tumor areas to a high dimensional feature space,then constructing a prognostic model to predict the future development of the disease by machine learning,and further to guide the selection of disease treatment and review strategy.In view that the data acquisition by CT are simple and the imaging results are easy to compare,CT was used as an important modality in the radiomic database and be widely used in radiomic researches.In this paper,we included 127 NSCLC data from the MAASTRO NSCLC Lung 1 dataset in accordance with the restriction of patients' survival time and divided the data into shorter group(less than 400 days)and longer group(more than 700 days)based on the subjects' survival time,in which the longer group.85 cases were randomly selected from the whole data included in this study to form the training set to construct the prognostic model,and the remaining 42 cases were used as the test set to evaluate the performance of the prediction model.According to the labels marked by radiologists,we conducted semi-automatic segmentation on all the tumors enrolled using the 3D-Slicer platform and Grow-Cut algorithm.Then,we wrote a tumor feature extraction toolbox for extraction of the morphological,texture and first order histogram-based and wavelet transformation-based tumor characteristics.In this paper,535 tumor features were defined and extracted.During the construction of prognostic model,we firstly adopted 11 feature selection methods to sort the features according to their correlation with survival time and then reduce the dimension of the feature space.Secondly,9 classification models were used in the training process of the prognostic model,meanwhile,the number of features in the training process was changed to find an optimal construction strategy for each selector-classifier combination.Then we synthetically considered the results of feature selection method and selected the best model constructing strategy based on the global optimal features.Finally,the reliability and accuracy of the prognostic models are evaluated quantitatively based on the feature selectorspecific or the global optimal features.Results shows that the reliability of the prognostic model is closely related to the number of features used in the model training process.Therefore,the number of characteristics for model training should be tested before the construction of the prediction model.Moreover,the prediction accuracy and the AUC of the prediction model based on global optimal feature is higher than that of the prediction model based on the optimal feature of the specific algorithm,which indicates that it may bring better results to take into account the characteristics of different feature selection methods in the construction of radiomics prognostic model. |
PDF Full Text Request