Texture Analysis Based On MRI In Cerebral Glioma Grading And Prognosis Prediction

Part ?:The application value of histogram parameters based on diffusion weighted imaging in cerebral glioma gradingObjective To investigate the application value of histogram parameters based on diffusion weighted imaging(DWI)in cerebral glioma grading.Methods A retrospective analysis was made of 30 low grade glioma(grade ?)and 97 high grade glioma(grade ? 46 cases,grade ? 51 cases)confirmed by pathology.MaZda was used to extract histogram parameters,including mean,variance,skewness,kurtosis,Pere.l%,Pere.10%,Pere.50%,Pere.90%and Pere.99%,of the highest signal region of tumor parenchyma in all patients' DWI images.Mann-Whitney U and independent-samples t test was used to compare histogram parameters between the two groups.Using receiver operating characteristic curve(ROC)to analyse the classification efficiency of the parameters with significant difference between LGG and HGG.Logistic regression(LR)classifier model was established by the parameters with significant difference between the two groups and ROC was used to evaluate the efficiency of the LR classifier model.Results The mean,variance,Pere.1%,Pere.10%,Pere.50%,Pere.90%,Pere.99%of histogram parameters were significantly different between LGG and HGG[(144.198±47.133)vs(185.609±40.341),(28.101 ±39.529)vs(160.143±211.832),(134.233±43.673)vs(162.577±40.478),(138.100±44.970)vs(172.81 4 ±39.384),(144.400±47.211)vs(186.247±40.473),(149.833±49.537)vs(197.443±42.977),(152.333±50.384)vs(201.361±43.720),all P<0.001)]and variance between ?/? was also statistically significant[(91.018±150.360)vs(160.143±211.832),P=0.001],while no significant difference in skewness and kurtosis between the two groups[(-0.322±0.499)vs(-0.369±0.542),P=0.669;(-0.171±0.587)vs(-0.135±0.973),P=0.440].When the variance was 29.23 as the threshold between the two groups,the classification efficiency was the highest,and the corresponding sensitivity,specificity and area under the curve(AUC)were 72.16%,76.67%and 0.793,respectively.The sensitivity,specificity and AUC of the LR classifier model established by the seven histogram parameters,which showed significantly different between LGG and HGG,were 61.86%,86.67%and 0.807,respectively.Conclusion The histogram parameters based on DWI can effectively differentiate LGG from HGG before operation,and the variance has high classification efficiency.Part ?:The application value of texture analysis based on MRI in cerebral glioma gradingObjective The purpose of this study was to investigate the application value of texture analysis(TA)based on magnetic resonance imaging(MRI),including diffusion weighted imaging(DWI),T1 weighted imaging(T1WI),T2 weighted imaging(T2WI)and contrast-enhanced T1 weighted imaging(T1W-CE),in differenting low grade gliomas(LGG)from high grade gliomas(HGG).Methods A retrospective analysis was made of 29 LGG(grade ?)and 62 HGG(grade? 31 cases,grade ? 31 cases)confirmed by pathology.MaZda was used to extract texture features,including gray histogram(GH),gray-level co-occurrence matrix(GLCM)and gray-level run length matrix(GLRLM),of tumor parenchyma in all patients' DWI,T1WI,T2WI and T1W-CE images.Feature parameters of grade ? and ? gliomas from four sequences(DWI,T1WI,T2WI,T1W-CE)were respectively compared with LGG(grade ?),then the parameters with statistical significance in the two comparisons were selected.After that,PCA was used to reduce the dimension of the selected feature parameters of each sequence and grid searching method was applied to reduce the number of PC.At last,logistic regressionlogic(LR)classifier models were established and receiver operatingcharacteristic curve(ROC)curve was drawn to evaluate the efficiency of classifier models.Results Between grade ? and grade ?,grade ? and grade ? gliomas,the number of feature parameters with statistical significance was larger in T1W-CE than the other three sequences,and the proportion of GLCM parameters was the largest in each sequence.The number of feature parameters with statistical significance of T1WI,T2WI and T1W-CE in grade ?/? was larger than that in grade ?/?.The number of overlapping feature parameters with statistical significance in the two comparisons of four sequences(DWI,T1WI,T2WI,T1W-CE)were 42,25,19 and 54 respectively,accountting for a large proportion in grade ?/?,which was 55.6%,95.0%,80.6%and 85.6%respectively.Among the LR classifier models established by PC of four sequences,T1W-CE had the best classification efficiency with sensitivity,specificity,accuracy,and area under the curve(AUC)of 82.8%,79.0%,80.8%,and0.869,respectively.After integrating PC of the four sequences,the efficiency of the classifier model was further improved,corresponding sensitivity,specificity,accuracy,and AUC were 81.0%,83.9%,82.5%,and 0.883,respectively.By grid searching method,the number of PC for modeling from T1WI sequence was reduced from 5 to 4,and only AUC slightly increased to 0.809.All 12 PC of the four sequences were reduced to 6 through grid searching method,and the performance of LR classifier model was further improved,with the specificity and accuracy increased to 90.3%and 85%while the sensitivity and AUC slightly reduced to 79.3%and 0.850 respectively.Conclusion TA based on MRI can effectively distinguish HGG from LGG before operation,and the LR classifier model based on texture features of T1W-CE sequence has the best performance.PCA is an effective way to reduce dimensions,combining with grid searching method,it my further simplify the model and improve the classification efficiency.Part ?:The application value of texture analysis based on MRI in predicting the time to recurrence of high grade glioma of brainObjective To investigate the application value of texture analysis(TA)based on magnetic resonance imaging(MRI)in predicting the time to recurrence(TTR)of high grade glioma(HGG)of brain.Methods A retrospective analysis was made of 48 HGG(grade ? 22 cases,grade ?26 cases),confirmed by pathology,who were divided into short-term group(TTR?6 months,n=23)and long-term group(TTR>6 months,n=25)according to the length of TTR.MaZda was used to extract texture features,including three features of gray histogram(GH),gray-level co-occurrence matrix(GLCM)and gray-level run length matrix(GLRLM),of tumor parenchyma in all patients' diffusion weighted imaging(DWI),T1 weighted imaging(T1WI),T2 weighted imaging(T2WI)and contrast-enhanced T1 weighted imaging(T1W-CE).Support vector machine-recursive feature elimination(SVM-RFE)was used for texture features selection,and support vector machine(SVM)classifier model is used for classification.Receiver operating characteristic curve(ROC)was drawn to evaluate the efficiency of SVM classifier model.Finally,Kaplan-Meier and log-rank test was used to compare TTR of long-term group and short-term group based on SVM classifier model.Results After preliminary comparison,SMO classifier model had the best performance among the three SVM classifier models(LibLinear,LibSVM,SMO).When the feature subset contained the top 23 feature parameters,the accuracy of SMO classifier model were the highest,corresponding sensitivity,specificity,accuracy and area under the curve(AUC)were 87.0%,92.0%,89.6%and 0.899,respectively.After optimizing the subset by grid searching method,the number of parameters in the optimal subset was reduced to 15,and the efficiency of SMO classifier model was further improved with the corresponding sensitivity,specificity,accuracy and AUC of 87.0%,96.0%,91.7%and 0.921,respectively.Finally,the TTR of the long-term and short-term groups divided by the optimal SMO classifier model was statistically significant(P<0.001).Conclusion Texture features,based on DWI,T1WI,T2WI and T1W-CE sequences,which can objectively evaluate the influence of HGG heterogeneity on TTR,has great potentiality for evaluating the prognosis of HGG.