Preliminary Study Of Radiomics Based On Conventional MRI Images In Clinical Application Of Glioma

The first part:Clinical value of MRI radiomics for preoperative grading ofgliomas[Objective]To study the efficacy and clinical value of radiomics in preoperative grading of gliomas.[Method]The MR images(including T1WI,T2WI,FLAIR,OWI,ADC and post contrast-enhanced T1WI)of 144 glioma(WHO ? 45 cases,WHO ? 52 cases,WHO IV 47 cases)confirmed by surgical pathology are retrospectively ananlyzed.Cases originated from two hospitals with the same MRI scan method and pathological examination method.One case from one hospital was used as a training set(95 cases),and another case was used as a validation set(49 cases).The total level of the solid portion and edema part in MR images were selected as regions of interest(ROI)by a software named,"ITK-SNAP".4432 radiomics features were extracted by Pyradiomics.We deleted low variance features.Mann-Whitney U test was applied to remove the features without statistical significance.Receiver operating characteristic(ROC)curves of significant features were employed to confirm cutoff value of grading gliomas and to evaluate the performance of preoperative grading gliomas.Selecting several optimal features with high AUC and multi-arranged combinatorial tests with high diagnostic efficiency to construct a prediction model based on training set by tree model.Then we used the ROC curve to verify the model on the validation set.[Results](1)Identify HGGs and LGGs:skewness of DWI sequence(SkewDwi),standard deviation of the T1WI-CE sequence(SDT1WI-CE),long-run gray level emphasis,entropy,and median of the FLAIR sequence(EmpFlair,EntFlai,,MedFlair)were lower in low-grade gliomas(LGGs)than in high-grade gliomas(HGGs).A gradient tree boosting model which includes six radiomics features exhibited optimal performance in distinguishing HGGs from LGGs(trainging set,AUC=0.905,sensitivity=0.865,specificity=0.882;validation set,AUC=0.784,sensitivity=0.783,specificity=0.813)is better than MRI sighs(whether or not the tumor contains cystic necrosis,the degree of tumor enhancement,maximal diameter of edema traing set,AUC=0.668,0.849,0.767,validation set,AUC=0.708,0.712,0.570).(2)Identify grade ? and ? glioams:entropy of Flair(Entflair)was higher in grade ? gliomas versus grade III gliomas(trainging set,AUC=0.80,sensitivity=0.952,specificity=0.533;validation set,AUC=0.761,sensitivity=O.636,specificity=0.733).The efficiency of Entflair was better than MRI sighs(whether or not the tumor contains cystic necrosis,the degree of tumor enhancement,maximal diameter of edema,training set,AUC=0.626,0.809,0.669;validation set,AUC=0.716,0.703,0.559).(3)Identify grade ? and ? gliomas:standard deviation of T1WI-CE sequence(SDT1WI-CE)and entropy of T2WI sequence(EntT2WI)were higher in grade IV versus grade ? gliomas(P=0.005,P=0.037,P=0.012).A gradient tree boosting model including 3 radiomics features exhibited desirable performance(training set:AUC=0.801,sensitivity=0.765,specificity=0.714;validation set:AUC=0.850,sensitivity=0.778,specificity=0.824)is significantly better than MRI(the maximal diameter of edema,the degree of tumor enhancement,training set,AUC=0.609,0.675;validation set,AUC=0.534,0.533).[Conclusion]Radiomics can provide more quantification information while more accurately grading gliomas before surgery,not only can distinguish high and low grade gliomas,but also can distinguish grade ? and ? gliomas,grade ? and ? gliomas.The second part:Clinical value of MRI radiomics for forecasting overallsurvival of high-grade gliomas[Objective]To study the efficacy and clinical value of radiomics in forecasting overall survival of high-grade gliomas.[Method]The MR images and follow-up data of 25 glioma cases were restrospectively analyzed.The maximum level of the tumor in MRI images is selected as regions of interest(ROI)by a software named,"Image J".Texture analysis on extracted ROI of meningiomas is performed.Through histograms and grey-level co-occurrence matrices(GLCM),features including:maximum value,minimum value,standard deviation(SD),skewness,kurtosis,angular second moment,contrast,inverse different moment,entropy,and correlation are derived.Pearson correlation analysis was used to calculate the dependency between features and overall survival.Receiver operating characteristic(ROC)curves of significant parameters were employed to confirm the area under the ROC curve and cutoff value.Then Kaplan-Meier curve was used to calculate the mean overall survival,Log-rank test was used to compare overall survival.[Results]Age,enhancement degree of tumor and whether the tumor contains cystic and necrosis were negatively correlated with the overall survival of high-grade gliomas(r=-0.474,p=0.017;r=-0.671,p=0.00);Whether the tumor contains cystic necrosis was related to the overall survival of high-grade glioms(?2=4.314,p=0.038),AUC=0.753,0.6,0.65.The standard deviation,skewness and entropy were negatively correlated with the overall survival of high-grade gliomas,while maximum value,kurtosis and inverse different moment were positively correlated with the overall survival of high-grade gliomas.The average survival time of patients with SDT2WI below 36.7 was 599.09 days(95%CI:332.037?866.080),and the average survival time of patients above 36.7 was 109.5 days(95%CI:240.471?644.329)(P=0.003),AUC=0.74,sensitivity=0.571,specificity=1;the average survival of patients with ContFlair less than 6.514 was 854.286 days(95%CO:245.082?1445.489),and the average survival of patients with more than 6.514 patients was 260.647 days(95%CI:164.139?357.155)(p=0.016),AUC=0.621,sensitivity=0.786,specificity=0.50;the average survival time of patients with SkewFlar less than-0.174 was 604.429 days(95%CI:276.381-932.381),the average survival of patients with more than-0.174 patients was 188.6 days(95%CI:99.570?277.630)(p=0.008),AUC=0.743,sensitivity=0.714,specificity=0.80.SkewT1WI-CE less than-0.123 was 1083.8 days(95%CI:239.096?1824.504),the average survival of patients with more than-0.123 patients was 268.421 days(95%CI:182.556?354.286)(p=0.008),AUC=0.564,sensitivity=0.929,specificity=0.40.The combination of age,conventional MRI features and radiomics features can provide predictive diagnostic efficiency value with an AUC of 0.81,sensitivity is 0.778,specificity is 0.786.[Conclusion]Radiomics can provide more quantification information to more accurately forecast overall survival of high-grade gliomas.Combination radiomics,age and conventional MRI sighs can further improve the prediction of overall survival of HGGs.