Value Of BOLD-MRI Texture Analysis On Predicting The Pathological Grades Of Renal Clear Cell Carcinoma

Objective:To explore the predictive value of blood oxygen-level dependent magnetic resonance imaging(BOLD-MRI)based texture analysis for grading clear cell renal cell carcinoma(ccRCC).Methods:Retrospective analysis was performed on the data of 84 patients with ccRCC confirmed by pathology in the Third Affiliated Hospital of Soochow University from April 2011 to July 2019.All patients were divided into the low-grade group(FuhrmanⅠ+Ⅱ,n=56)and the high-grade group(Fuhrman Ⅲ+Ⅳ,n=28).BOLD-MRI was performed within two weeks before the surgery.Imaging analysis was performed on the T2*Maps at the section with the largest cross-sectional area of renal mass by two radiologists using the double-blind method.The regions of interest(ROI)were manually delineated along the outline of the renal masses.A total of 396 texture features were calculated for each ROI based on the shape,histogram,gray-level co-occurrence matrix(GLCM),gray-level run length matrix(GLRLM),gray-level size zone matrix(GLSZM),and gray-level dependence matrix(GLDM).The intra-observer and inter-observer reproducibility of all the texture features were assessed by the calculation of the intraclass correlation coefficient(ICC).Differences in the texture features between the high-and low-grade ccRCCs were determined using the Mann-Whitney U test.The univariate logistic regression analysis was performed to select the texture features with differential diagnosis value for the differentiation of the high-from low-grade ccRCC(P<0.05).Ten texture features were selected using the minimum redundancy maximum relevance(mRMR)method.The multivariable logistic regression analysis was performed to develop the BOLD-MRI based texture model for the prediction of the high grade ccRCC.The diagnostic performance of the texture model in the differentiation of the high-from the low-grade ccRCC was assessed by receiver operating characteristic(ROC)curve analysis.The data was randomly grouped into training and test sets at a 7:3 ratio,and the leave group out cross validation(LGOCV)was performed to test the reliability of the model.P<0.05 was considered to indicate a significant difference in all the statistical tests.Results:A total of 396 BOLD-MRI based texture features were extracted from each case.The ICC of all the texture features for inter-observer reproducibility were 0.76-0.92,and the ICC for intar-observer reproducibility were 0.80-0.95.After the Mann-Whitney U test,164 features that showed significant differences between the low-and high-grade ccRCCs(all P<0.05)were selected.120 features with differential diagnosis value for the differentiation of the high-from low-grade ccRCC(all P<0.05)were selected using the univariate logistic regression analysis.Finally,ten features were selected using the mRMR method.Feature 3(wavelet_LHH_glcm_MaximumProbability),4(wavelet_HLH_glszm_SmallAreaEmphasis),7(wavelet_HHH_glszm_SmallAreaEmphasis),9(wavelet_LLH_glcm_Idmn),10(wavelet_LHH_glszm_SizeZoneNonUniformity)were independent predictors,and the odds ratio were 0.37,2.15,1.80,1.78,and 2.78.The area under ROC curve of the prediction model for differentiating high-from low-grade ccRCC was 0.84,with the cutoff value of 0.80,73.81%accuracy,85.71%sensitivity,and 67.86%specificity.The mean accuracy,sensitivity,and specificity was 82.26%,82.54%,and 82.50%for the training sets,and 75.01%,77.11%,and 71.00%for the test sets.Conclusion:BOLD-MRI texture analysis can be a noninvasive approach for preoperative prediction of ccRCC Fuhrman grading,and provide a basis for the optimizing operation scheme and clinical prognostic prediction.