| Objective 1.To investigate the value of radiomic models based on CE-T1WI and T2FLAIR in predicting the early development of high-grade glioma after standard treatment.2.To investigate the value of histogram analysis model based on Dynamic Contrast Enhanced MRI(DCE-MRI)and Dynamic Susceptibility Contrast Enhanced MRI(DSC-MRI)in predicting the development of non-measurable diseases(NMD)of high-grade glioma after standard treatment.Material and Methods 1.Data were collected from Center 1(from January 2012 to May 2021),Center 2 and 3(from January 2018 to January 2000)follow-up MRI of HGG patients after standard treatment,including CE-T1WI and T2FLAIR.A total of 335 patients were enrolled,including 257 in the center 1,divided into training group and internal validation group in a 7:3 ratio and 78 patients in center 2 and Center 3 used as external validation group.Shape feature,first order feature,texture feature and the first-order features and texture features of the image processed by six filters were extracted.A total of 107 original features and 1209 filter-based features were collected for each sequence of each patient.Four feature selection methods including ANOVA,KW,Relief and RFE were used to further screen features.The first 20 important features are applied to the model using the backward method.We used 10 machine learning classifiers to model the training group:AB,AE,DT,GP,LDA,LR,LRLasso,NB,RF and SVM.Two radiologists scored the findings of each imaging feature.Decision curve analysis was used to measure the clinical utility of radiometrics models and visual evaluations by radiologists.The diagnostic performance of the developed model was evaluated using receiver operating characteristic(ROC)curves in the training,internal and external validation groups.2.Follow-up MRI,including structural MRI,DCE-MRI and DSC-MRI,were collected from HGG patients who underwent standard treatment in our hospital from January 2016 to October 2021.A total of 98 patients were enrolled,including 55 in the recurrence group and 43 in the non-recurrence group.The histogram features of DCE(Ktrans,Kep,Ve and Vp)and DSC(rCBV)of NMD were extracted,and binary logistic regression was used to construct a fusion model of single perfusion histogram parameters and multiple perfusion parameters to predict recurrence in the next followup.Operator characteristic curves are used to compare the prediction performance of different models.Results 1.A total of 40 models were established by different feature selection methods and classifiers,and the combination of feature selection method ANOVA+GP had the best performance.The AUC of training and internal validation groups was 0.925 and 0.843,respectively.The model used 10 radiomics features includeing 8 texture features(including 2 GLCM,3 GLRLM,and 3 GLDM)and 2 first-order features.Decision curve analysis showed that radiomic models consistently provided a better net benefit than the radiologist’s visual assessment in predicting high-grade glioma progression.2.The number of the histogram features of Kep,Vp,Ktrans,Ve and rCBV that could predict the development of NMD were 15,10,10,13 and 14,respectively.The fusion model of DSC-rCBV(90Percentile+Kurtosis+RootMeanSquared)in prediction of recurrence and non-recurrence group showed the AUC of 0.915.The fusion model AUC of VP,Kep and Ktrans was 0.915,0.913 and 0.917 respectively.Ve showed the lowest AUC for prediction of the recurrence of non-measurable lesions(AUC=0.884).Fusing all DCE parameters and DSC-rCBV parameters,the model consists of Kep_Energy,Vp_Minimum and rCBV_RootMeanSquared,whose AUC value reaches the maximum(0.970).Conclusion 1.The machine learning radiomics model based on structural MRI sequence can accurately predict whether HGG will relapse next time after standard treatment.2.The histogram features based on DCE and DSC can predict the development of NMD after HGG standard treatment,and the multi-parameter fusion model can further improve the prediction efficiency. |
PDF Full Text Request