The Value Of Multimodal MRI Based On Machine Learning In Diagnosis And Post-treatment Evaluation Of Diffuse Glioma In Adults

Isocitrate dehydrogenase（IDH）is an important molecular marker in the genome for comprehensive diagnosis and dynamic classification of diffuse glioma.In the first part,we applied machine learning to predict IDH status in patients with Diffuse Glioma based on multiparameter MRI.According to the 2016 World Health Organization（WHO）classification,our study aimed to evaluate IDH status using MRI in clinically diagnosed grade II～IV glioma patients.One hundred and seventy-six patients with confirmed WHO grade II～IV glioma were retrospectively investigated as the study set,including lower-grade glioma（WHO grade II,n=64;WHO grade III,n=38）and glioblastoma（WHO grade IV,n=74）.The minimum apparent diffusion coefficient（ADCmin）in the tumor and the contralateral normal-appearing white matter（ADCn）and the r ADC（ADCmin to ADCn ratio）were defined and calculated.Intraclass correlation coefficient（ICC）analysis was carried out to evaluate interobserver and intraobserver agreement for the ADC measurements.Interobserver agreement for the morphologic categories was evaluated by Cohen’s Kappa analysis.The nonparametric Kruskal-Wallis test was used to determine whether the ADC measurements and glioma subtypes were related.By univariable analysis,if the differences in a variable were significant（P<0.05）or an image feature had high consistency（ICC>0.8;κ>0.6）,then it was chosen as a predictor variable.The performance of the area under the receiver operating characteristic curve（AUC）was evaluated using several machine learning models,including logistic regression,support vector machine,Naive Bayes and Ensemble.The optimal model was developed as the final model to predict IDH status in 40 patients with glioma as the subsequent test set.The results showed that,in the study set,six measured variables（r ADC,age,enhancement,calcification,hemorrhage,and cystic change）were selected for the machine learning model.Logistic regression had better performance than other models.Two predictive models,model 1（including all predictor variables）and model 2（excluding calcification）,correctly classified IDH status with the AUC of 0.897 and 0.890,respectively.The test set performed equally well in prediction,indicating the effectiveness of the trained classifier.The subgroup analysis revealed that the model predicted IDH status of LGG and GBM with accuracy of 84.3%and 85.1%,and the AUC of 0.873 and 0.862,respectively.Therefore,this part concluded that through the use of machine-learning algorithms,the accurate prediction of IDH-mutant versus IDH-wildtype was achieved for adult diffuse gliomas via noninvasive MR imaging characteristics,including ADC values and tumor morphologic features,which are considered widely available in most clinical workstations.At present,the best treatment for glioma is to remove the tumor tissue to the maximum extent without damaging neurological function to improve the survival rate of patients.After surgery,radiotherapy and temozolomide chemotherapy（TMZ）,combined with 6 cycles of TMZ,are required to slow down the growth of residual tumors.However,The MRI findings of radiation-induced injury are similar to those of tumor recurrence,which makes clinical diagnosis difficult.Misdiagnosis of tumor progression may lead to the opposite treatment of glioma patients.To solve this problem,in the second part,multiparameter MRI radiomics model was established and validated,including conventional MRI,diffusion-weighted imaging（DWI）and arterial spin labeling（ASL）,to differentiate radiation-induced injury from tumor recurrence in glioma patients and improve diagnostic ability.In this part,a retrospective analysis was performed on 96 pathologically confirmed gliomas with enhanced lesions after postoperative radiotherapy and chemotherapy（33 patients with WHO grade 2,17patients with WHO Grade 3,46 patients with WHO Grade 4）in the training set,and verified in an independent prospective validation set（n=30）.4199 radiomics features were extracted from conventional T₁WI,T₂WI,contrast-enhanced T₁WI,apparent diffusion coefficient（ADC）,and cerebral blood flow（CBF）.Radiomics features were selected using Mann-Whitney U Test,LASSO regression and recursive feature elimination（RFE）algorithms.Intraclass correlation coefficient（ICC）was used to test the consistency between groups.The selected features were integrated into a multi-parameter MRI radiomics model based on support vector machine（SVM）,using a 10-fold cross-validation and grid search mechanism to automatically tune the prediction model.The area under the curve（AUC）was compared with a single-parameter radiomics model to evaluate the ability of differentiating tumor recurrence from radiation-induced injury.The training set model was validated in the subsequent validation set to evaluate the diagnostic performance of the model.The results showed that a total of 8 important radiomics features（3 from conventional MRI,1 from ADC and 4 from CBF）were extracted for model construction.The multi-parameter radiomics model（AUC=0.96）was higher than the conventional MRI model（AUC=0.88）,ADC（AUC=0.91）and CBF（AUC=0.95）.It also showed better results in the validation set（AUC=0.94,P=0.001）,which was higher than any single model.Subgroup analysis showed that the prediction performance of WHO grade 4 was higher than that of WHO grade 2～3 in conventional MRI model group and ADC model group,and the AUC in the validation set was0.92 and 0.89,respectively.For CBF and multi-parameter radiomics group,the prediction performance of WHO grade 2～3 was higher than that of WHO grade 4,and the AUC in the validation set was 0.97 and 1.00,respectively.Based on the above results,the multiparameter radiomics model,especially the inclusion of DWI and ASL,showed excellent diagnostic performance in differentiating radiation-induced injury from tumor recurrence,providing a reliable noninvasive diagnostic method for clinical differential diagnosis.According to the recommendation of Response Assessment in Neuro-Oncology（RANO）,MRI examination is the standard diagnostic method for evaluating tumor grade and tumor response after treatment.Both early pseudoprogression and late radiation necrosis caused by radiotherapy are manifested in enlarged or newly developed enhanced lesions in conventional MRI,which are difficult to distinguish from tumor recurrence.Correctly distinguishing postoperative recurrence of glioma from radiation-induced brain injury is very important for the choice of clinical treatment,and advanced MR perfusion imaging technology provides the possibility to distinguish the two.Therefore,in the third part,we performed a systematic meta-analysis and regression analysis on the application of MR perfusion imaging in the prediction and differential diagnosis of glioma recurrence.We evaluated the value of dynamic susceptibility contrast enhanced（DSC）,dynamic contrast enhanced（DCE）and arterial spin labeling（ASL）in the differential diagnosis of glioma recurrence and radiation-induced brain injury.The methods were based on the Preferred Reporting Program（PRISMA）statement for systematic reviews and meta-analyses.This meta-analysis had been successfully registered on the PROSPERO（https://www.crd.york.ac.uk/PROSPERO/）,and the registration number was CRD42022304404.A comprehensive literature search was performed for clinical studies in differentiating between glioma recurrence and radiation brain injury using perfusion magnetic resonance imaging（PWI）,including DCS,DCE and ASL.The retrieval period was from the beginning of database construction to October 31,2021.Two investigators independently screened the trials that met the inclusion criteria.QUADAS-2 was used to evaluate the quality of the included studies.Cochran Q test was used to determine whether there was heterogeneity,and the heterogeneity was measured in combination with statistic I².Stata 16.0,Meta-Disc and Rev Man5.2 were used for Meta analysis and chart making.Meta-regression and subgroup analyses were applied to identify sources of heterogeneity in studies.Funnel chart was used to assess whether the included studies have publication bias.Finally,forty studies were included,including 27 English literatures and 13 Chinese literatures.There were1804 patients of glioma recurrence patients and 1206 patients of radiation injury.Meta-analysis results showed that the pooled sensitivity and specificity of DSC in differentiating glioma recurrence from radiation brain injury were 0.82（95%CI 0.78-0.86）,0.87（95%CI 0.80-0.92）,and the area under the curve（AUC）was 0.89（95%CI 0.86-0.92）,respectively.The pooled sensitivity and specificity of DCE were 0.83（95%CI 0.76-0.89）,0.83（95%CI 0.78-0.87）,and AUC was 0.88（95%CI 0.85-0.91）,respectively.The pooled sensitivity and specificity of ASL were 0.80（95%CI 0.73-0.86）,0.86（95%CI 0.76-0.92）,and AUC was 0.88（95%CI 0.85-0.91）,respectively.The results of this part demonstrated that DSC-PWI,DCE-PWI and ASL perfusion imaging techniques displayed high accuracy in distinguishing glioma recurrence from radiation brain injury.Based on current literature research evidence,DSC-PWI achieved a higher diagnostic performance than the other two.MRI perfusion imaging could be used as a feasible and quantitative examination method for postoperative follow-up after radiotherapy and chemotherapy,providing strong evidence to support the subsequent clinical treatment.