The Value Of Radiomics Based On MR Non-Gaussian Diffusion Models In Predicting IDH Status Of Glioma

Background and PurposeAs a malignant tumor,glioma has a high incidence in adults,and the prognosis of patients varies widely due to differences in genetic phenotypes.The isocitrate dehydrogenase(IDH)1/2 gene status is an important molecular marker for predicting the outcome of glioma patients.It was added to the 2016 World Health Organization(WHO)classification of tumors of the central nervous system.Magnetic resonance imaging(MRI)is an important part of examinations of glioma patients before treatment,which can provide critical information for tumor characterization.With perfect tissue contrast and multiple parameters,it can well display the location and volume of tumor,as well as the peritumoral edema from multiple directions.The advantage of MR diffusion imaging is that it does not require a contrast agent and can reflect the microstructure of the tumor.The clinically used routine diffusion weighted imaging(DWI)is based on monoexponential model.It can provide information about tissue microstructure by quantifying the diffusion of water molecules in tissue,and its quantitative parameter,apparent diffusion coefficient(ADC)has been used to predict IDH status of glioma,but the results are unfavorable.This is because of the assumption of routine DWI that the diffusion of water molecules in vivo follows a Gaussian distribution,which is opposite to reality.In fact,due to the existence of various barriers such as cell membranes and myelin sheaths in vivo,the diffusion of water molecules follows a non-Gaussian distribution.In order to better characterize the diffusion of water molecules in vivo,various non-Gaussian diffusion models,for example,diffusion kurtosis imaging(DKI),neurite orientation dispersion and density imaging(NODDI)and mean apparent propagator(MAP)were developed and have been used for predicting IDH status of glioma.However,the results are still conflicting.This may result from the heterogeneity of tumor.In previous studies that predict IDH status of glioma with diffusion models,only the minimum,mean/median,and maximum values or histogram parameters and texture features of quantitative parameters were used,which could not mirror the heterogeneity of tumor adequately,and thus may affect the diagnostic performance in predicting IDH status of glioma.Radiomics can extract large quantities of high-dimensional quantitative features from images,providing much more information than histogram parameters and texture features,better reflecting the heterogeneity of tumors,and improving the accuracy in predicting IDH status of gliomas.In this study,patients with glioma were divided into IDH wild-type group and IDH mutant group according to their IDH status.Radiomics features of diffusion parameters were extracted from the lesions and different radiomics models were constructed for predicting IDH status of glioma.The purpose of this study was to assess the diagnostic value of radiomics models based on non-Gaussian diffusion models(DKI,NODDI and MAP)in predicting IDH status of glioma and to compare the diagnostic efficiency of these models and that of routine DWI-based model.Materials and Methods1.Patients:One hundred and forty-three histopathologically confirmed glioma patients with available IDH sequencing results were enrolled in this study.The ages of patients ranged from 18 to 74 years old,the average age was 46.86±11.73 years,there were 7 7 males and 66 females.There were 73 cases in IDH wild-type group and 70 cases in IDH mutant group.All patients were divided into training set(n=100,IDH wild-type/IDH mutant=51/49)and validation set(n=43,IDH wild-type/IDH mutant=22/21)for building and validating radiomics models.2.MRI scanning:All MRI examinations were performed using a 3.0 T scanner(Magnetom Prisma;Siemens Healthcare,Erlangen,Germany)with an integrated 64-channel head and neck coil.The sequences including routine MRI(T1WI,T2WI,T2-dark-fluid and routine DWI),multi-b-value(b=0,500,1000,1500,2000,2500 s/mm2)diffusion MRI and contrast-enhanced MRI.3.Image post-processing:The ADC map was derived from routine DWI(b=1000 s/mm2)by the MRI workstation automatically.Parametric maps of DKI[fractional anisotropy(FA),mean diffusivity(MD),axial diffusivity(AD),radial diffusivity(RD),mean kurtosis(MK),axial kurtosis(AK)and radial kurtosis(RK)],NODDI[orientation dispersion index(ODI),intracellular volume fraction(ICVF)and isotropic volume fraction(ISOVF)]and MAP[return to the origin probability(RTOP),return to the axis probability(RTAP),return to the plane probability(RTPP),non-Gaussianity(NG),axial non-Gaussianity(NGax),radial non-Gaussianity(NGrad),mean square displacement(MSD)and q-space inverse variance(QIV)]were generated from the multi-b-value diffusion MRI using an in-house-developed post-processing software,NeuDilab,based on DIPY(http://nipy.org/dipy).4.Radiomics process:After scanning,T2-dark-fluid images of all patients were imported into ITK-snap software(version 3.8.0,http://www.itksnap.org)for region of interest(ROI)delineation manually.The ROI was defined as the largest area with abnormal signal,including tumor and peritumoral edema across all slices on T2-dark-fluid images.Each parametric map was registered to T2-dark-fluid images with ITK-snap software.With FAE(version 0.4.2)software,851 radiomic features were extracted within the ROI from each parametric map.Then normalization and feature reduction were performed on the features.Before building the model,recursive feature elimination(RFE)was applied for features selection,and then support vector machine(SVM)was employed as the classifier for building radiomics models based on routine DWI,DKI,NODDI and MAP,respectively.To determine the hyper-parameter(e.g.the number of features)of models,cross validation with 5-fold was applied on the training data set.The hyper-parameters were set according to the model performance on the validation data set.The performance of a model was assessed using receiver operating characteristic(ROC)curve analysis.The area under the ROC curve(AUC)was conducted to measure the diagnostic performance.The accuracy,sensitivity,specificity,positive predictive value(PPV),and negative predictive value(NPV)were calculated at a cutoff value that with the maximum Yorden index.5.Statistical analyze:The clinical characteristics of IDH wild-type group and IDH mutant group were analyzed using SPSS 21.0 software(IBM,SPSS,version 21.0),the age and sex of patients with different IDH status were compared using independent sample T-test and chi-square test,respectively.When testing categorical variables and continuous variables of radiomics features,chi-square test and Mann-Whitney U test were used,respectively.When comparing the diagnostic performance of radiomics models based on different diffusion models,MedCalc software(Version 20.015)was used to perform the DeLong test to compare the ROC curves of different models.P<0.05 was considered statistically significant.Results1.Correlation of sex and age with glioma grade and IDH status:There were significant differences in age between IDH wild-type group(51.80±11.04 years)and IDH mutant group(41.71±10.15 years)and between high-grade(WHO grade IV)group(52.25±12.09 years)and low-grade(WHO grade II,III)group(43.69±10.32 years)(P<0.001,respectively).There was no significant difference in gender between IDH wild-type and IDH mutant groups and between high-grade and low-grade groups(P=0.918 and 0.593,respectively).The ratio of number of high-grade cases/number of low-grade cases was statistically different between the IDH wild-type and IDH mutant groups(P<0.001).2.Diagnostic performance of each radiomics model in predicting glioma IDH status:The AUC of the routine DWI-based radiomics model in prediction IDH status of glioma was 0.699,and the corresponding accuracy,sensitivity,specificity,positive predictive value,and negative predictive value were 0.674,0.864,0.476,0.633,and 0.769,respectively.The AUC of the DKI-based radiomics model in prediction IDH status of glioma was 0.870,and the corresponding accuracy,sensitivity,specificity,positive predictive value,and negative predictive value were 0.814,0.682,0.952,0.938,and 0.741,respectively.The AUC of NODDI-based radiomics model in prediction IDH status of glioma was 0.913,and the corresponding accuracy,sensitivity,specificity,positive predictive value,and negative predictive value were 0.861,0.909,0.810,0.833,and 0.895,respectively.The AUC of the MAP-based radiomics model in prediction IDH status of glioma was 0.920,and the corresponding accuracy,sensitivity,specificity,positive predictive value,and negative predictive value were 0.884,1.000,0.762,0.815,and 1.000,respectively.3.Comparisons of AUC among all models in predicting IDH status of glioma:The results of DeLong test showed that there was no significant difference in AUC among the three radiomics models based on non-Gaussian diffusion models(DKI,NODDI and MAP)(all P values were>0.05).When the DKI,NODDI,and MAP-based models were compared with the routine DWI-based models respectively,the differences in AUC were statistically significant(all P values<0.05).ConclusionThe radiomics models based on non-gaussian diffusion models(DKI,NODDI and MAP)can effectively predict the IDH status of glioma,and perform better than routine DWI-based radiomics model.