Gliomas Grading Methods Based On PET Images And Genetic Information

Gliomas are the most common primary intracranial tumors,accounting for 32%of all central nervous system tumors and 81% of intracranial malignant tumors.Key gene biomarkers,such as MGMT,Ki67 and IDH1/2,are the main indicators to evaluate the malignant degree of gliomas.Accurate preoperative diagnosis is very important for the choice of treatment and prognosis prediction.Because of the special location of the brain,the operation and biopsy methods used in clinical practice are high-risk and expensive to obtain the status of biomarkers of related genes,which bring great pain to patients.PET-CT imaging can provide non-invasive information on tumor metabolism and has become a common imaging method for the diagnosis of glioma.Based on the use of PET-CT image data,relevant clinical and genetic information,we use machine learning methods to predict the state of gene biomarkers in patients with gliomas.This will help achieve accurate treatment of glioma patients,reduce the risk of surgery,and reduce medical costs.In this study,based on the theory of radiomics,we achieved accurate preoperative glioma grading by noninvasive prediction of key gene biomarkers MGMT,Ki67 and IDH1/2.During the research process,based on individual radiomics features,individual clinical features,and a combination of both,a radiomics model,a clinical model,and a combined model are established respectively.First,the radiomics features are extracted and calculated from the region of interest of PET-CT image,and the critical radiomics features were determined using the feature selection algorithm(U-test,elastic net and SVM-RFE),and the radiomics model was constructed by SVM algorithm.In the construction of clinical model,we first use the logistic regression algorithm with AIC as the evaluation index to screen the clinical features,and then use the screened features and logistic regression to build the clinical model.Third,the prediction probability of the radiomics model is taken as a key feature,and after combining with the clinical feature,the feature selection of the combined model is performed through the logistic regression algorithm of AIC as theevaluation index.Then,the screened features are used to construct a combined model by logistic regression.Finally,the Kaplan-Meier curve is drawn according to the prognostic information of the patient,and the prognosis of the patient is analyzed.And the difference of overall survival time of glioma patients in different states of three gene biomarkers(MGMT,Ki67 and IDH1/2)was obtained.The experimental results show that the method proposed in this study has good prediction performance,and the radiomics features significantly improve the performance of classification model.Among the classification models that predict whether the MGMT promoter is methylated,the AUC(area under receiver operating characteristic curve)values of the training and test datasets of the radiomics model reached 0.94 and 0.86,both higher than the clinical models(0.64 and 0.69).In the classification models that predict Ki-67 expression,the AUC of the radiomics model in the training and test datasets reached 0.88 and 0.76,both higher than the clinical models(0.84 and 0.67).Among the classification models that predict whether IDH1/2is mutated,the AUC of the combined model in the training dataset and test dataset reached 0.91 and 0.80,both higher than the clinical models(0.71 and 0.66).In order to facilitate clinical application,this study used Nomogram to visually predict the classification model.The results of the calibration curve and the decision curve show that the final classification model built using nomograms has similar performance in the training set and test set,proving the effectiveness of these models.