Applied Research On Prediction Of Ki-67 Expression In Glioma Based On Radiomics Of MRI

Objective:To investigate the application value of radiomics features based on T₂WI andenhanced T₁WI sequence in preoperative prediction of Ki-67 expression in patients with glioma.Materials and Methods:1.General information and image acquisitionRetrospective analysis of 81 glioma patients in our hospital,including 48 males and 33 females,aged 18 to 94 years,with an average age of（55±13.62）years.All patients underwent surgery and obtained pathological results.According to thepathological results,they were divided into 37 cases of glioblastoma,10 cases ofanaplastic astrocytoma,9 cases of diffuse astrocytoma,6 cases of glioblastoma（IDH-1Wild type）,6 cases of anaplastic oligodendroglioma,4 cases of astrocytoma,2 cases of oligodendroblastoma,2 cases of glioblastoma multiforme,one case ofoligodendroglioma,one case of glioma（glial cells atypical proliferation）,one case of obese astrocytoma,one case of fibrous astrocytoma,and one case of anaplasticoligodendroglial astrocytoma.All patients underwent craniocerebral MRI before surgery,including plain T₁WI,T₂WI,and T₁WI enhanced imaging.Patient inclusion criteria:（1）the patient did not undergo any form of treatment such as chemoradiotherapy,immunotherapy,surgical resection,etc.;（2）the patient received an MRI scan before surgery,and the MRI image was clear and the lesion showed good;（3）during the operation frozen and postoperative pathological examination confirmed glioma and the clinical data were complete;（4）Informed consent of the patient.Patient exclusion criteria:（1）patients who underwent chemoradiotherapy,immunotherapy,etc.before surgery,or recurrence after surgery;（2）those who have contraindications to MRI;（3）those who have severe MRI image artifacts and poor lesions;4)Clinical data are incomplete.2.Pathological dataAll patients enrolled performed Ki-67 measurement by immunohistochemistry（Envision method）within one week after surgical resection,and calculated the percentage of positively stained malignant cells to the total number of cells,namely the Labeling Index（LI）,LI=Number of positive cells/total number of cell×100%.In this study,patients with glioma were divided into two groups according to the Ki-67 index:Ki-67 low expression group（Ki-67≤25%）and high expression group（Ki-67>25%）.There were 45 cases in the low expression group and 36 cases in the high expression group.3.Image Processing and omics analysisThe acquired magnetic resonance images were grouped and uploaded to theradiology platform in DICOM format,and the lesions were manually delineated layer by layer on the T₂WI and enhanced T₁WI images,respectively,to obtain the volume of interest（VOI）,then extract the omics feature values through omics analysis and feature calculation,select the feature values through the three-dimensional reduction methods of Variance Threshold,Select KBest,and Lasso,and finally obtain the optimaleigenvalues.Then we andomly take 80%of the VOI of all lesions as the training group and 20%of the VOI as the test group.Machine learning is performed on the extracted feature values using six classifiers.The six classifiers include:K-Nearest Neighbor（KNN）,Extreme Gradient Boosting（XGBoost）,Support Vector Machine（SVM）,Random Forest（RF）,Decision Tree（DT）,and Logistic Regression(Logistic Regression（LR）.By analyzing its area under the ROC curve（AUC）,95%confidence interval,Specificity and Sensitivity and other indicators,we evaluate the diagnostic effectiveness of the model constructed by each classifier.Results:The following information can be obtained by comparing and analyzing the AUC values of the models constructed based on the 6 classifiers of T₂WI combined enhanced T₁WI,T₂WI,and enhanced T₁WI sequence images.1.Based on the image features of the T₂WI combined enhanced T₁WI sequences,the AUC value of the prediction model constructed by the SVM and LR classifiers is better,and the AUC value of the LR classifier is the best:（1）LR classifier results:AUC of Ki-67 low expression is 0.873（95%CI:0.76-0.99）,sensitivity is 0.78,specificity is 1.00;AUC of Ki-67 high expression is 0.873（95%CI:0.76-0.99）,sensitivity is 1.00,specificity is 0.78;（2）SVM classifier results:AUC of Ki-67 low expression is 0.778（95%CI:0.60-0.95）,sensitivity is 0.56,specificity is 1.00;AUC of Ki-67 high expression is 0.778（95%CI:0.60-0.95）,sensitivity is 1.00,specificity is 0.56;2.Based on the image characteristics of the T₂WI sequence,the AUC value of the prediction model constructed by the LR classifier is the best.The AUC of Ki-67 low expression is 0.778（95%CI:0.55-1.00）,the sensitivity is 0.78,and the specificity is0.75;AUC of Ki-67 high expression is 0.778（95%CI:0.55-1.00）,sensitivity is 0.75,specificity is 0.78;3.Based on the image features of the enhanced T₁WI sequence,the AUC value of the prediction model constructed by the KNN classifier is the best.The AUC of Ki-67 low expression is 0.590（95%CI:0.35-0.83）,the sensitivity is 0.56,and the specificity is0.63;AUC of Ki-67 high expression is 0.590（95%CI:0.35-0.83）,sensitivity is 0.63,and specificity is 0.56.By comparing their AUC values,we can find that the prediction model constructed based on the LR classifier of the plain-scan T₂WI combined with the enhanced T₁WIsequence has the highest AUC value,which has a higher diagnostic efficiency,which is higher than the constructed prediction models based on plain-scan T₂WI or theenhanced T₁WI sequence.The extracted optimal feature values include:first-orderfeatures（Skewness,10Percentile,Median,Energy,Total Energy）,gray-level dependency matrices（Large Dependence High Gray Level Emphasis,Dependence Variance,Large Dependence Low Gray Level Emphasis）,and gray area size matrix（High Gray Level Zone Emphasis）.In addition,the AUC value of the prediction model constructed by the KNN classifier based on the enhanced T₁WI sequence alone is the lowest,which has the worst diagnostic performance for predicting Ki-67 expression status,and its sensitivity and specificity are not very good.Conclusion:1.The imaging histology models based on plain scan T₂WI combined with enhanced T₁WI sequence and single plain scan T₂WI sequence can predict the expression status of Ki-67 in glioma.Among them,the prediction performance of an imaging histologymodel based on plain scan T₂WI combined with enhanced T₁WI sequence is better.2.The imaging histology models based on plain scan T₂WI combined with enhanced T₁WI sequence and single plain scan T₂WI sequence had potential for non-invasive prediction of Ki-67 expression status in preoperative gliomas.