CT Radiomics In Predicting Invasiveness And Ki-67 Expression Level Of Ground-glass Nodules

Background In clinical diagnosis work,the detection rate of ground glass nodules(GGN)in the lung has increased significantly.The malignancy rate of GGN lesions on thin-slice CT may be higher.To predicting the invasiveness and Ki-67 expression level accurately in GGN lung adenocarcinoma is important for clinicians to make better treatment decisions.Radiomics can deeply dig out the changes in the microscopic level of cells or proteins from CT imaging,and provide a reliable solution to the problem of tumor heterogeneity.Objective The purpose of this study was to develop and validate a CT radiomics model to explore its value of the differential diagnosis of benign and malignant ground-glass nodules(GGN),predicting the invasiveness and estimating the expression level of Ki-67 in GGN lung adenocarcinoma.Methods Thin-layer CT scan images and corresponding pathological and immunohistochemical results of 139 patients(161 GGN lesions)who underwent pathological confirmation at Renhe Hospital affiliated to Three Gorges University from July 2018 to December 2020 were collected for retrospective analysis and divided into benign(87 cases)and malignant(74 cases)groups according to the benignity of the lesions.Among the 106GGN-like lung adenocarcinomas,they were divided into pre-infiltrative lesion group(51cases)and infiltrative lesion group(55 cases)according to the degree of infiltration,they were also divided into Ki-67 low expression group(74 cases)and Ki-67 high expression group(32cases)according to the expression level of Ki-67.The region of interest(ROI)of all GGN was outlined along the nodal edges by two imaging physicians using automatic machine recognition followed by manual edge modification,and seven types of imaging histological features including first-order features,morphological features,and texture features were extracted.The radiomics features were filtered and downscaled by F Test,Pearson,L1 Based and Tree-Based methods.According to the experimental requirements,logistic regression models were used to classify the data,construct corresponding prediction models,and select the most valuable radiomics features.The diagnostic efficacy of CT radiomics models was evaluated by ROC curve and decision curve.Results(1)A total of 22 highly correlated radiomics features were screened in the radiomics model for diagnosing GGN benignity and malignancy,and the area under the curve(AUC)value of this model validation set for diagnosing GGN benignity and malignancy was0.916(95% CI: 0.873-0.959),with an accuracy of 0.845,sensitivity of 82.4% and specificity of 86.2%.Decision curve analysis showed that when the threshold probability was greater than 0.05,the CT radiomics model constructed in this study predicted GGN benignity and malignancy with greater benefit to patients,indicating that the model has better clinical application efficacy.(2)A total of 51 radiomics features were screened in the radiomics model for identifying PILs and ILs,and the area under the curve(AUC)value for the validation of this model focused on predicting the invasiveness of GGN lung adenocarcinoma was 0.926(95% CI: 0.878-0.974),with an accuracy of 0.849,sensitivity of 85.2% and specificity of84.6%,and the decision curve analysis showed that when the threshold probability was greater than 0.04,the CT radiomics model constructed in this study has greater benefit for patients with GGN lung adenocarcinoma and has better clinical application efficacy.(3)A total of eight most meaningful radiomics features were screened in the radiomics model for predicting Ki-67 expression level in GGN lung adenocarcinoma.The area under the curve(AUC)for predicting the expression level of Ki-67 in GGN lung adenocarcinoma in this model was 0.899(95% CI: 0.836-0.962),with an accuracy of 0.840,sensitivity of 93.8% and specificity of 79.7%.Decision curve analysis showed that when the threshold probability was greater than 0.12,the CT radiomics model constructed in this study predicted the Ki-67 expression level of GGN lung adenocarcinoma with greater benefit for patients with GGN lung adenocarcinoma,indicating that this model has better clinical application efficacy.Conclusion Radiomics based on CT images can provide a non-invasive detection method for the accurate preoperative diagnosis of GGN lung adenocarcinoma.The radiomics model established based on radiomics features can not only effectively identify the benign and malignant nature of ground-glass nodules(GGN),but also assess the degree of invasiveness and the expression level of Ki-67 in GGN lung adenocarcinoma.