Radiomics Based Pathological Grading Prediction Of Glioma

Objective: A random forest model was constructed by using radiomics and machine learning to interpret MRI images of gliomas,and to explore the feasibility and effectiveness of preoperative prediction of glioma pathological grade by MRI images.Methods: First,the DICOM image of the BraTS training set is obtained,and all the imaging datasets have been segmented manually.Feature extraction of all images using the Pyradiomics package under the Python platform.All samples were randomly divided into training groups according to 70%,and 30% were assigned to the test group.The training group(198 cases)was then divided into 5 subgroups(40,39,40,39,40 samples in each subgroup),and 5-fold cross-validation fitting was performed using the R.Using the Gini coefficient as the branching rule of the decision tree in the random forest,the best parameter of the random forest is mtry(Number of variables to possibly split at in each node)and min.node.size(that is,the size of the smallest node).The optimal parameters obtained are used to construct the final random forest model and output the feature importance ranking.Finally,images of the test group were interpreted,and classifier performance was evaluated.The confusion matrix,accuracy,sensitivity,specificity,accuracy,Fl value were calculated,as well as evaluation of consistency,and the ROC curve was plotted.Results: In the training set,a total of 210 high-grade gliomas and 75 low-grade gliomas were obtained.Each case contained 4 sequences of T1,T2,Flair,T1 ce and 1 pre-segmented ROI sequence.Three of the 75 cases of LGG were excluded because some features could not be extracted.The remaining 72 cases of LGG and all 210 cases of HGG extracted 14 shape features,18 first-order features,24 GLCM features,14 GLDM features,16 GLRLM features,16 GLSZM features,and 5 NGTDM features.The 5-fold cross-validation yielded a random forest with a mtry optimized value of 48 and a min.node.size of 7,with an AUC of 0.8958746,a sensitivity of 0.6254545,and a specificity of 0.9324138.The random forest model feature variables fitted by the training group and the sequence of their belongings ranked in order of importance are: Surface Volume Ratio,GLDM-Small Dependence Low Gray Level Emphasis-T1 ce,glrlm-Run Entropy-T1,Sphericity,GLSZM-Large Area High Gray Level Emphasis-Flair,GLDM-Large Dependence High Gray Level Emphasis-T1 ce,GLDM-Dependence Non Uniformity-T1,GLDM-Dependence Non Uniformity-T1 ce,GLSZM-Large Area High Gray Level Emphasis-T2,NGTDM-Coarseness-Flair.The random forest model constructed with the optimal parameters was used to interpret the MRI images of the test group.The accuracy of the final pathological grade was 91.67%(95% CI was 83.58%-96.58%),the sensitivity was 95.24%,and the specificity was 90.48%.The precision was 76.92%,the non-information prediction rate was 75%,and the accuracy was greater than non-information prediction rate by P<0.001.The consistency test Kappa value was 0.7941 and the F1 value was 0.8511.Conclusion: The classifier constructed in this study has high sensitivity and specificity,and the consistency test shows that its classification ability is highly consistent with pathological diagnosis.Combining multiple evaluation data,it can be considered that the use of the random forest classifier constructed in this study for the pathological grading prediction of glioma has good feasibility,and the classifier can play a role in assisting preoperative diagnosis to a certain extent.