Magnetic Resonance Imaging(MRI)-based Radiomics For Preoperative Grading And Molecular Classification Of Gliomas

PART1 RADIOMICS ANALYSIS OF DIFFUSION TENSOR IMAGING IN PREOPERATIVE GRADING OF GLIOMASObjective:Accurate gliomas grading plays an important role in the clinical management of patients and is also the basis of molecular stratification nowadays.The purpose of this study was to explore the value of radiomics analysis of fractional anisotropy（FA）and mean diffusivity（MD）maps derived from diffusion tensor imaging（DTI）in preoperative grading of gliomas.Methods:This retrospective study included 108 patients who had pathologically confirmed brain gliomas and DTI scanned.This cohort included 43 low grade gliomas（LGGs;all grade II）and 65 high grade gliomas（HGGs）,including 25 cases of WHO III and 40 cases of WHO IV.Firstly,two methods were used to segment the regions of interest（ROIs）on the B₀ images.The first method delineated all abnormal signal regions,including solid components of tumor,the necrotic or cystic components of tumor and peritumoral edema.The second method only delineated the solid components of tumor.Then a set of radiomic features of the two kinds of ROIs in the FA map and the MD map were respectively extracted from the first three and four layers of the pre-trained convolutional neural networks.Two classification tasks based on support vector machine（SVM）were established using the these radiomics features:（1）LGGs vs HGGs,（2）WHO III vs WHO IV gliomas.Finally,the area under the curve（AUC）,accuracy,sensitivity and specificity of the receiver operating characteristic（ROC）curve were used as the performance metrics using the leave-one-out cross validation method.Results:A prediction model based on DTI radiomics for preoperative grading of gliomas had been successfully established,and the optimal features mainly were derived from the deep features of pre-trained CNN.The MD map achieved the best predictive scores,with AUC=0.96,accuracy of 98%,sensitivity of 98%,specificity of 98%respectively in classifying LGGs from HGGs.While the multimodality（FA+MD）maps had the best predictive performance,with AUC=0.99,accuracy of 98%,sensitivity of 98%and specificity of 100%respectively in classifying WHO III and WHO IV.When the necrotic or cystic components of tumor and peritumoral edema were not included in the ROI,the predictive performance of FA map changed significantly.While classifying LGGs from HGGs,the predictive performance of FA features in the first three convolution layer decreased from AUC=0.92,accuracy of 93%,specificity of 88%to AUC=0.90,accuracy of 91%,specificity of 81%,respectively,except for the sensitivity;and the predictive performance of FA features in the first four convolution layer decreases from AUC=0.96,accuracy of97%,sensitivity of 98%,specificity of 95%to AUC=0.84,accuracy of85%,sensitivity of 86%,specificity of 84%,respectively.While classifying WHO III and WHO IV,the predictive performance of FA features in the first three convolution layer decreased from AUC=0.94,accuracy of 94%,sensitivity of 98%,specificity of 88%to AUC=0.88,accuracy of 83%,sensitivity of 95%,specificity of 64%,respectively;and the predictive performance of FA features in the first four convolution layer decreases from AUC=0.95,accuracy of 95%,sensitivity of 100%,to AUC=0.85,accuracy of 81%,sensitivity of 75%,respectively,except for the specificity.Conclusion:Radiomics based on FA and MD maps were useful for noninvasively grading of gliomas,in which the proposed deep convolutional radiomic features played an important role.The predictive performance of FA was susceptible to the ROI of tumor segmentation.PART2 PREDICTION OF 1P/19 Q CO-DELETION STATUS IN LOW-GRADE GLIOMAS USING MRI RADIOMICSObjective: The isocitrate dehydrogenase（IDH）mutation and 1p/19 q co-deletion status of low-grade gliomas（LGGs）are closely related to the prognosis and clinical treatment strategies.The purpose of this study was to investigate the radiomics based on T₂-FLAIR and apparent diffusion coefficient（ADC）map for noninvasive prediction of 1p/19q-codeletion status in LGGs with IDH mutation,in order to provide more information for the development of treatment plan and prognosis evaluation.Methods:The MRI data of 44 LGGs patients with IDH mutation confirmed by surgical pathology and molecular genetic tests were retrospectively analyzed,and their results of 1p/19 q codeletion were obtained.Firstly,a set of radiomics features from the T₂-FLAIR and ADC were extracted,including the traditional features,morphology,and the deep features extracted via the pre-trained convolutional neural（CNN）networks model.Then a machine learning model based on support vector machine（SVM）to predict 1p/19 q codeletion was established by using these features.The area under the curve（AUC）,accuracy,sensitivity and specificity of the receiver operating characteristic（ROC）curve were used as the performance metrics using the leave-one-out cross validation method.Results: A machine learning model based on SVM to predict the1p/19 q codeletion of LGGs was successfully established,and the best features were all selected from the deep features extracted by pre-trained CNN.According to ROC analysis,the model based on T₂-FLAIR had the best prediction effect,with AUC=0.95,accuracy of 95%,sensitivity of 96%and specificity of 95%;the model based on ADC had AUC=0.91,accuracy of 91%,sensitivity of 96%,specificity of 85%;the model based on multimodality（T₂-FLAIR+ADC）was the last,with AUC=0.91,accuracy of89%,sensitivity of 92%.and specificity of 85%.Conclusion: Radiomics based on T₂-FLAIR and ADC would be helpful in non-invasive prediction of 1p/19 q co-deletion in LGGs with IDH mutation,and the deep features were most predictive.