Prediction Of Ki-67 Expression In Glioma Of High Grade Based On MRI Imagomics

Objective: To explore the value of conventional MRI based imaging omics in preoperative prediction of Ki-67 expression in high-grade glioma,and to provide evidence for clinical treatment and prognosis prediction.Methods: A total of 105 patients with pathologically confirmed high-grade glioma were retrospectively collected from July 2019 to March 2022,including 54 patients with high expression of Ki-67 and 51 patients with low expression of Ki-67.They were divided into training group(n=73)and verification group(n=32)according to the ratio of 7: 3.The basic clinical information,including gender and age,was recorded.The common imaging signs of MRI plain scan and enhanced scan were evaluated,including the location of the lesion,whether the lesion was regular in shape,whether it was multiple,the degree of peritumoral edema,whether the midline structure was shifted,and whether it was peritumoral enhancement.Independent sample t test was used to compare the difference of age between the two groups,and chi-square test or Fisher’s exact test was used to compare the difference of gender and imaging signs between the two groups.The ROI was manually delineated using IBEX software based on MATLAB.Five feature extracting group,including ID(Intensity Direct)、IH(Intensity Histogram)、GLCM(Gray-Level Cooccurence Matrix)、GLRLM(Gray-Level Run-Length Matrix)、NIDM(Neighbor Intensity Difference Matrix).A total of 2208 radiomics features were extracted from T2 WI,T2-FLAIR and enhanced T1 WI sequences.Intra-class correlation coefficient(ICC)was used to evaluate the intra-researcher agreement.The parameters with statistically significant differences were screened by one-way analysis of variance,and then the dimension was reduced by LASSO.The best radiomics features obtained after dimension reduction were incorporated into the decision tree(DT),logistic regression(LR),support vector machine(SVM)and adaptive enhancement(Adaboost)learner respectively,to construct the radiomics model.The validation set was used to test the four prediction models established.T The receiver operating characteristic curve(ROC)and the area under the curve(AUC)of the training set and the validation set were drawn to evaluate the sensitivity,specificity and accuracy of each model.Results: One-way analysis of variance found that there was no statistically significant difference in the general clinical data and MRI images collected between the two groups for high-grade glioma Ki-67 high expression and Ki-67 low expression Six radiomics features were selected after dimensionality reduction.The imaging features were used to construct LR,SVM,RF and DT models and evaluate their effectiveness respectively.ROC curve results showed that the AUC value of the LR model training set was 0.817,the sensitivity was 76.9%,the specificity was 74.3%,and the accuracy was 75.7%.The AUC value of the validation set was 0.713,the sensitivity was 73.3%,the specificity was 75.0%,and the accuracy was 74.2%.The AUC value of the DT model training set was 0.871,the sensitivity was 71.8%,the specificity was85.7%,and the accuracy was 78.4%.The AUC value of the validation set was 0.731,the sensitivity was 86.7%,the specificity was 56.3%,and the accuracy was 71%.The AUC value of SVM model training set was 0.808,the sensitivity was 84.6%,the specificity was 54.3%,and the accuracy was 70.3%.In the validation set,the AUC value was 0.662,the sensitivity was 53.3%,the specificity was 68.8%,and the accuracy was 61.3%.The AUC value of Adaboost model training set was 0.718,the sensitivity was 69.2%,the specificity was 74.3%,and the accuracy was 71.6%.The AUC value of the validation set was 0.646,the sensitivity was 66.7%,the specificity was 62.5%,and the accuracy was 64.5%.Conclusion: 1.Imaging based on conventional T2 WI,T2-FLAIR and enhanced T1 WI has certain value in predicting the expression status of Ki-67 in high-grade glioma before surgery,and has a promising application prospect,to provide a basis for clinical treatment and prognosis prediction..2.In this study,compared with logistic regression,adaptive enhancement and support vector machine model,decision tree model has the best predictive performance.