| Background:Lung cancer is the most common malignant tumor and the leading cause of death from cancer.The prognosis of patients with advanced lung cancer is poor,and the prognosis of patients with early lung cancer is good.Therefore,early detection and identification of patients with malignant pulmonary nodules are important secondary preventive measures to reduce the risk of lung cancer death and disease burden.At present,the physical examination of pulmonary nodules through electronic computer tomography(CT)screening is becoming more and more popular,but discriminating the nature of pulmonary nodules based on CT requires doctors to have a wealth of clinical medical knowledge and medical imaging knowledge,for basic medical doctors and young medical workers,these are very difficult.It is mainly based on CT images to discriminate the nature of pulmonary nodules to assist clinicians to accurately identify malignant pulmonary nodules,which is of great significance for early detection and early treatment of lung cancer patients and improving the prognosis of patients.Non-small cell lung cancer is the most common type of lung cancer,and its prognosis is poor.Tumor-Node-Metastasis(TNM)staging system has its limitations,and the survival time of patients with the same stage varies greatly.Besides,most existing prognostic prediction models are not suitable for patients to predict their prognosis before treatment,because these models rely heavily on the results of pathological examination,it is necessary to obtain pathological tissue after an invasive examination or even surgery.At present,lung cancer patients undergo CT examinations.If CT images can be used to predict prognosis accurately,appropriate treatment options can be selected for patients to improve patient survival rates and prognosis.Radiomics is a method of quickly extracting a large number of features from medical images.Further analysis of the high-dimensional data can develop discriminant and prognostic prediction models.However,omics data has the characteristics of the high dimension,many confounding and noises,and non-linear correlation.The models that can handle high-dimensional omics data will be discussed and compared in this study.Purpose:This paper aims to discuss the construction of a discriminant model of pulmonary nodules in lung cancer screening and a prognostic prediction model of non-small cell lung cancer and development of clinical easy-to-use tools based on radiomics data to assist doctors in selecting appropriate treatment options and improving clinical decision-making efficiency,thereby improving lung cancer survival rate.Materials and Methods:In the study of the discriminant model of pulmonary nodules,875 patients with pulmonary nodules were included.Radiomics features were extracted from the CT images of each patient.A random forest discriminant model was constructed with routinely available variables and all radiomics features;radiomics features were screened by Least Absolute Shrinkage and Selection Operator(LASSO)and radiomics scores were calculated;a logistic discriminant model was constructed with routinely available variables and radiomics scores.The two models are evaluated for the training set and the validation set by the Area Under Curve(AUC).A simple application tool was developed based on the model with higher AUC.In the study of the prognostic model of non-small cell lung cancer,1302 patients with non-small cell lung cancer were included.Radiomics features were extracted from the CT images of each patient.A random survival forest model and a deep survival neural network model were constructed with routinely available variables and all radiomics features;radiomics features were screened by the LASSO-COX method and radiomics scores were calculated;a COX Proportional risk model was constructed with routinely available variables and radiomics scores.The Concordance index(C-index)is used to evaluate the three models for the training set and the validation set,and a simple application tool was developed based on the model with the highest C-index.Results:CT images of 875 patients with pulmonary nodules and 1302 patients with non-small cell lung cancer were collected,the region of interest(ROI)was segmented,and 1288 radiomics features were extracted for every patient.In the study of the discriminant model of nature of pulmonary nodule,a random forest discriminant model was constructed with the AUC and its 95%confidence interval of 0.762(0.716,0.808)on the training set,and the AUC and its 95%confidence interval of 0.755(0.674,0.836)on the validation set;17 features were selected from 1288 radiomics features,and radiomics scores were calculated.A logistic discriminant model was constructed using a stepwise regression method and the radiomics scores,age,and gender were screened out.The AUC and its 95%confidence interval are 0.837(0.797,0.877)on the training set,and the AUC and its 95%confidence interval are 0.804(0.733,0.875)on the validation set.After comparing and evaluating the two models,the logistic discriminant model has a better effect,and it is used to draw a discriminant nomogram of the pulmonary nodule,and the calibration curve is in good agreement.In the study of the prognostic model of non-small cell lung cancer,a random survival forest model was constructed with the C-index and 95%confidence interval of 0.811(0.771,0.850)on the training set and the C-index and 95%confidence interval of 0.810(0.762,0.858)on the validation set;a deep survival neural network model was constructed,the C-index and its 95%confidence interval are 0.820(0.784,0.857)on the training set,and the C-index and its 95%confidence interval are 0.823(0.775,0.871)on the validation set;9 imaging omics features were selected from 1288 imaging omics features,and imaging omics scores were calculated.The COX model was constructed using a stepwise regression method and the radiomics scores,age,smoking,and family history of lung cancer were selected.The C-index and its 95%confidence interval are 0.840(0.806,0.874)on the training set,and the C-index and its 95%confidence interval are 0.849(0.810,0.888)on the validation set.After evaluation and comparison,the prognostic prediction effect of the COX model is the best,and the prognostic nomogram is drawn with it,and the calibration curve agrees well.Conclusion:Radiomics features are of great significance for discriminating the nature of pulmonary nodules and predicting the prognosis of non-small cell lung cancer.The logistic model for discriminating the nature of pulmonary nodules based on the radiomics features is well,and the discriminant nomogram obtained from it is simple and easy to use.The COX model for predicting the prognosis of non-small cell lung cancer based on radiomics features has a good effect,and the prognostic prediction nomogram is simple and easy to use. |
