Conventional MRI Texture Analysis For Predicting IDH1 Mutations In Gliomas And Prognosis

Objective:The mutation status of IDH1 in glioma was predicted by texture analysis of contrasted-enhanced T1WI,T1WI and T2WI MRI sequences,and the correlation between different texture features and survival rate and Ki-67 was explored.Materials and methods:A total of 63 patients with histopathologically confirmed gliomas,including 28 patients carrying IDH1 mutation and 35 with wild-type IDH1were enrolled in this retrospective study.The preoperative MRI data?contrasted-enhanced T1WI,T1WI and T2WI?,clinical data,postoperative pathology?including histology,Ki-67 and IDH genes?,progression free survival?PFS?and overall survival?OS?were collected.The regions of interest?ROIs?covering the entire tumor and edema were manually delineated on axial slices using O.K.?Omni Kinetics,GE Healthcare,China?software;and the texture features?a total of 29 features?.Statistical package for social science?SPSS?22.0 was used for data analysis.Intra-class correlation coefficients?ICC?were used for the consistency assessment between radiologists.For comparison of texture features between groups that max-intensity,skewness,quantile95 in contrasted-enhanced T1WI,inverse difference moment in T1WI,GLCM-energy in T2WI were conformed to normal distribution,student t-test was used;and Mann-Whitney U test was used for other features.Kaplan-Meier method was used for survival analysis between the two groups.Receiver operating characteristic?ROC?curve was plotted to assess the differential diagnostic efficiency of the significant features.R language?RStudio Version 1.0.1432009-2016 RStudio,Inc.?was used to complete feature dimension reduction and model construction.In order to establish the model,70%of each group was classified as training group and the remaining 30%as validation group.Lasso method and10-fold cross-validation were used to reduce the dimensionality of texture features and select high efficiency features.Multivariate logistic regression analysis was used to model.Confusion matrix was used to analyze the accuracy of the model.ROC was used to evaluate the diagnostic efficiency of each model.Spearman correlation analysis was used to analyze the correlation between texture features and PFS,OS,Ki-67.Results:1.Difference of texture features between two groups of contrasted-enhanced T1WI,T1WI and T2WI sequences?1?In texture analysis of contrasted-enhanced T1WI features,max-intensity,standard deviation,variance,skewness,quantile95,inertia,cluster shade and cluster prominence were significantly decreased in IDH1 mutation group compared to IDH1wild group;while uniformity,correlation of IDH1 mutation group were significantly increased?all P<0.05?.According to ROC analysis,the uniformity?cutoff value=0.80,area under the curve?AUC?=0.809?was considered the best parameter for the diagnosis of two groups,with the sensitivity of 67.9%and specificity of 91.4%,respectively.?2?In texture analysis of T1WI features,the IDH1 mutation group lower values of GLCM entropy,inertia and cluster prominence were significantly decreased compared to IDH1 wild group,and the IDH1 mutation group revealed higher values for GLCM energy,correlation and inverse difference moment?all P<0.05?.According to ROC analysis,the cluster prominence?cutoff value=514.188,area under the curve?AUC?=0.698?was considered the best parameter for the diagnosis of two groups,with the sensitivity of 65.7%and specificity of 78.6%,respectively.?3?In texture analysis of T2WI features,standard deviation,variance,GLCM entropy,inertia,cluster shade and cluster prominence were significantly decreased in IDH1mutation group compared to IDH1 wild group;while kurtosis,uniformity,GLCM energy,correlation,inverse difference moments of IDH1 mutation group were significantly increased?all P<0.05?.According to ROC analysis,the GLCM entropy?cutoff value=0.039,area under the curve?AUC?=0.728?was considered the best parameter for the diagnosis of two groups,with the sensitivity of 78.6%and specificity of 68.6%,respectively.2.Results of logistic regression model of contrasted-enhanced T1WI,T1WI and T2WI sequences?1?Modeling formulas:Using Lasso method and 10 fold cross-validation,the high-efficiency parameters of T1WI enhancement??skewness,SK?,?inertia,IN?,?correlation,CO?,?cluster shade,CS??,T1WI??min-intensity,MI?,?inertia,IN?,?cluster prominence,CP??and T2WI signal intensity??uniformity,UN?,IDM?inverse difference moment,IDM??for predicting IDH1 mutation were obtained respectively.Multivariate logistic regression was used for modeling.The modeling formula of T1WI enhancement,T1WI and T2WI were as follows:f_T1WI+C=-0.4901+3.8697×SK+0.7731×IN-18.6774×CO-0.0501×CS f_T1WI=-1.377750+0.002105×MI+0.123742×IN+0.000885×CPf_T2WI=13.75-12.39×UN-7.78×IDM?2?Diagnostic efficiency of contrasted-enhanced T1WI modeling:The AUC of the contrasted-enhanced T1WI model?training group?was 0.914?cutoff value=0.649,sensitivity of 83.3%and specificity of 89.5%?,and the accuracy was 86.0%.The AUC of model?verification group?was 0.869?cutoff value=0.649,sensitivity of81.8%and specificity of 88.9%?.The accuracy was 85.0%.?3?Diagnostic efficiency of T1WI modeling:The AUC of the T1WI model?training group?was 0.726?cutoff value=0.537,sensitivity of 66.7%and specificity of89.5%?,and the accuracy was 76.7%.The AUC of model?verification group?was0.707?cutoff value=0.537,sensitivity of 63.6%and specificity of 77.8%?.The accuracy was 70.0%.?4?Diagnostic efficiency of T2WI modeling:The AUC of the T2WI model?training group?was 0.785?cutoff value=0.584,sensitivity of 66.7%and specificity of84.2%?,and the accuracy was 74.4%.The AUC of model?verification group?was0.737?cutoff value=0.584,sensitivity of 54.5%and specificity of 88.9%?.The accuracy was 70.0%.3.The correlation between texture features and PFS,OS,Ki-67 of contrasted-enhanced T1WI,T1WI andT2WI sequences?1?Texture features of contrasted-enhanced T1WI:The max-intensity,standard deviation,variance,range,relative deviation,skewness,uniformity,energy,entropy,quantile95,cluster shadow,cluster prominence were significantly correlated with Ki-67?P<0.05?.There was no significant correlation between texture features and PFS,OS?P>0.05?.?2?Texture features of T1WI:The skewness and cluster shadow are significantly correlated with PFS and OS?P<0.05?;kurtosis,GLCM energy,GLCM entropy,inverse difference moment were significantly correlated with Ki-67?P<0.05?.?3?Texture features of T2WI:The min-intensity,skewness and cluster shadow were significantly correlated with PFS?P<0.05?;the texture features of T2WI were not significantly correlated with OS?P>0.05?;the min-intensity,skewness,kurtosis,uniformity and cluster shadow were significantly correlated with Ki-67?P<0.05?.Conclusion:1.MRI texture analysis could be used as a new noninvasive method for identification of gliomas with IDH1 mutation.2.The contrasted-enhanced T1WI,TIWI and T2WI model can improve the diagnostic efficiency,and the contrasted-enhanced T1WI model has the highest diagnostic efficiency and the best stability.3.Some texture features?skewness,cluster shadows,etc.?of conventional MRI sequences could be used as prognostic indicators.