Study On The Predictive Value Of MRI Radiomics In The Pathological Grading Of Meningioma

Objective(s): To investigate the predictive value of conventional MRI radiomics in the pathological grading of meningiomas.Methods: The clinical data and MRI images of 174 patients with meningiomas confirmed by histopathology and WHO grade in the Third Affiliated Hospital of Kunming Medical University were retrospectively analyzed.There were 125 cases of low-grade meningiomas(WHO grade 1)and 49 cases of high-grade meningiomas(WHO grade 2 or 3).According to the ratio of 7:3,these cases were randomly divided into training set(122 cases,87 cases of low level,35 cases of high level)and test set(52 cases,38 cases of low level,14 cases of high level).The general clinical data and conventional MRI imaging features of the patients were analyzed,including age,gender,left and right cerebral hemispheres,location,shape,whether the growth across the midline,whether the edge was clear,vascular flow void phenomenon,peritumoral edema,whether the enhancement was uniform,the degree of enhancement,meningeal tail sign,tumor invasion,cystic degeneration,calcification,hemorrhage and necrosis.Independent sample t test was used to compare the differences in age between the two groups,and chi-square test or Fisher’s exact test was used to compare the differences in gender and clinical characteristics.The axial T1 WI,T2WI and T1WI-CE sequences were selected,and the lesions were manually delineated by 3D Slicer software.Different sequences were used to extract radiomics features.The radiomics features included Shape2 D,Shape3D,First Order,GLCM,GLDM,GLRLM,GLSZM,NGTDM features.Least absolute shrinkage and selection operator(LASSO)and 10-fold cross-validation were used for dimensionality reduction.The best performance radiomics features were included in six machine learning classifiers,including logistic regression(LR),support vector machine(SVM),decision tree(DT),random forest(RF),gradient boosting decision tree(GBDT)and extreme gradient boosting(XGBoost).The model with the best performance was selected.The statistically significant conventional MRI features were included to construct a clinical diagnostic model.All the selected radiomics features of the three sequences were included to construct a radiomics model.A combined diagnostic model was constructed by combining conventional MRI imaging features and radiomics features.The receiver operating characteristic curve(ROC)was drawn,and the area under the ROC curve(AUC),sensitivity,specificity,precision and F1-score were used to evaluate the performance of the model.Results: There was no significant difference in age among the general clinical data of the patients.The gender difference was not statistically significant in the training set,but was statistically significant in the test set(p<0.05).The number of senior male was significantly higher than that of senior female.Among the conventional MRI imaging features,Whether the tumor showed vascular flow void phenomenon(training set VS test set p=0.000 VS 0.016),whether there was peritumoral edema(training set VS test set p=0.000 VS 0.020),and whether the tumor was enhanced evenly(training set VS test set p=0.018 VS 0.026)and whether the tumor invaded the surrounding tissues(training set VS test set p=0.000 VS 0.004),the differences between the low grade and high grade groups were statistically significant.1316 features were extracted from T1 WI,T2WI and T1WI-CE sequences,respectively.After dimension reduction,26 features were selected from T1 WI,25features from T2 WI and 18 features from T1WI-CE.Among the T1 WI sequence models,SVM model had the best performance,with an AUC value of 0.873(95%CI:0.813-0.933),sensitivity of 75.0%,specificity of 86.8 %,accuracy of 73.5%,F1-score74.0% in the test set.Among the T2 WI sequence models,the RF model had the best performance,with an AUC of 0.879(95%CI: 0.817-0.942),a sensitivity of75.0%,a specificity of 78.9%,a precision of 76.9%,and a F1-score of 75.7% in the test set.Among the T1WI-CE sequence models,SVM model had the best performance,with an AUC value of 0.883(95%CI: 0.818-0.948),sensitivity of 73.1%,specificity of 78.9%,accuracy of 74.4%,and F1-score of 73.6% in the test set.Among the three sequences,the SVM model had the highest AUC value and the best performance,and the T1WI-CE sequence had better performance than T1 WI and T2 WI sequences.For the model based on SVM machine learning classifier,the AUC value of the test set of clinical diagnosis model was 0.871(95%CI: 0.815-0.928),the sensitivity was 65.4%,the specificity was 57.9%,the accuracy was 78.5%,and the F1-score was 67.2%.The AUC value of the radiomics model test set was 0.908(95%CI: 0.869-0.947),the sensitivity was 75.0%,the specificity was 86.8%,the accuracy was 73.5%,and the F1-scorewas 74.0%.The AUC value of the test set of combined diagnostic model was 0.916(95%CI: 0.881-0.951),the sensitivity was80.8%,the specificity was 84.2%,the precision was 81.8%,and the F1-score was81.2%.Conclusion(s): 1.Among the conventional MRI imaging features,the vascular flow void phenomenon inside the tumor,the edema of the brain tissue around the tumor,the enhancement uniformity inside the tumor and the tumor invasion of the surrounding tissue structure have certain clinical reference value for the WHO pathological grading diagnosis of meningioma.2.The model based on MRI radiomics features can assist clinical preoperative prediction of WHO pathological grade of meningioma,and provide high operability and excavability.With the continuous accumulation of sample size and the continuous improvement of machine learning algorithms,it can provide a quantitative basis for clinical diagnosis and effectively predict the preoperative grading of meningioma,which has good application value and development prospects.3.The performance of the model based on MRI radiomics features combined with conventional imaging features was improved,and the predictive performance of the combined diagnostic model was better than that of the single model.