Radiomics Model Based On Conventional MRI Combined With Diffusion Kurtosis Imaging To Predict The Pathological Grading Of Glioma

Objective:1.To discuss the ability of preoperative conventional MRI characteristics and diffusion kurtosis imaging(DKI)to evaluate the glioma’s pathological grading.2.To explore the value of a multiparametric radiomics model based on conventional MRI combined with diffusion kurtosis imaging(DKI)to predict the pathological grading of glioma.Method:1.Fifty-one patients with glioma were retrospectively collected from June 2014 to April 2021.The radiological data included of all patients included conventional MRI sequences(T1WI,T2WI,T2-FLAIR and CE-T1WI)and DKI sequences 2 weeks before surgery.Surgical pathological examination was used as the gold standard for glioma pathological grading.34 cases of high-grade gliomas and 17 cases of low-grade gliomas were included.The patients’ clinical data(age and gender)and characteristics of conventional MRI(tumor location,maximum tumor diameter,crossing middling,multifocality,border,cystic lesion,degree of necrosis,hemorrhage,volume and mode of enhancement,edema index,and abnormal signal area ratio of T1WI to T2-FLAIR images)were extracted.DKI sequences were post-processed to obtain mean kurtosis(MK)parameter maps.The MK of the solid part of the tumor was measured by outlining the region of interest(ROI)with reference to the CE-T1WI or T2WI sequences,and the nMK was obtained by normalizing the MK with reference to the contralateral normal appearing white matter(NAWMc).T-test,chi-square test or Fisher’s exact probability method were used to analyze whether there were statistical differences in clinical data and MRI characteristics between the high-and low-grade glioma groups.The capability of MK and nMK to predict the pathological grading of glioma was evaluated using receiver operating characteristic(ROC)curves.2.The DICOM format images of all patients were uploaded to the radiomics cloud platform(Huiying Medical Technology Co.,Ltd.)for radiomics analysis.Using CE-T1WI or T2WI sequences as reference,the region of interest(ROI)was manually outlined layer by layer for T1WI,T2WI,T2-FLAIR,CE-T1WI sequences and MK parameter maps for feature extraction,and 1409 radiomics features were extracted to each MRI sequence.The total sample set was randomly divided into five subsets of equal size,and the ratio of training set to test set samples is 4:1.,and the best feature set was obtained by using a 5-fold cross-validation method for feature selection.A support vector machine(SVM)algorithm was used to build a glioma grading prediction model with selected features.The ROC curves were plotted for single sequence and combined sequences radiomics model,respectively,and the sensitivity and specificity of the prediction models were calculated at the maximum Jorden index as the best cut-off value.The area under the curve(AUC)was used to evaluate the predictive efficacy of the radiomics models.Result:1.Clinical data:The mean age of the low-grade glioma group was lower than that of the high-grade glioma group,and the difference between the groups was statistically significant(P<0.05),and there was no statistical difference in gender composition between the two groups(P>0.05).2.Conventional MRI characteristics and DKI parameters:Among the 12 conventional MRI characteristics,six imaging characteristics involving tumor location,cystic,necrosis degree,enhancement volume,enhancement mode,and abnormal signal area ratio of T1WI and T2-FLAIR images were statistically different between the high-and low-grade glioma groups(P<0.05).The ROC curve showed that MK was more effective than nMK in predicting glioma grade,with an AUC value of 0.861,sensitivity of 0.706,and specificity of 0.941.3.After feature selection,5,9,5,4 and 15 radiomics features were selected from T1WI,T2WI,T2FLAIR,CE-T1WI and MK sequences,respectively.Among the prediction model based on a single MRI sequence,the highest prediction efficiency was achieved with the MK sequence,with a mean AUC of 0.864,sensitivity of 0.800,and specificity of 0.857 in the test set.The highest diagnostic efficacy of the combined model was achieved by combining multiple sequences,and the combined sequence consisting of T1WI+T2-FLAIR+CE-T1WI+MK had the highest diagnostic efficacy,with an AUC of 0.995,sensitivity of 0.867,and specificity of 0.905 in the test set.Conclusion:1.Conventional MRI characteristics and DKI parameters are of great value for preoperative prediction of glioma pathological grading.2.The radiomics model based on conventional MRI and DKI can effectively predict glioma pathological grading,and the model established by MK sequence in a single sequence has the highest predictive efficacy,better than the predictive efficacy of MK values in the traditional diagnostic model,and the combined model established by combined multiple sequences can improve the diagnostic efficacy,providing a non-invasive and efficient way to diagnose glioma.