Prediction Of Short-term Outcome Of Standard Chemoradiotherapy For Residual Glioma After Surgery Using Machine Learning Based On MRI Radiomics

Background: Due to the characteristics of invasive growth,easy invasion of important brain functional areas,glioma is often difficult to completely resect.For the treatment of residual tumors,conventional radiotherapy or concurrent chemoradiotherapy is currently the standard treatment recommended internationally.However,it has been reported in the literature and found in clinical practice that some patients receiving radiotherapy and chemotherapy under the condition of suffering great physical and mental pain and possibly serious sequelae,the therapeutic results are not all effective,and some even deteriorate(lesion progression),which is also an important defect of the traditional standardized treatment of glioma.Therefore,non-invasive prediction of the effect of postoperative standardized treatment will play an important role in the choice of clinical treatment.This study attempts to use MR Radiomics based machine learning methods to predict the short-term outcome of postoperative residual lesions in patients with glioma on standard chemoradiotherapy in advance,so as to help clinicians optimize and formulate individualized treatment plans in advance,and avoid or reduce unnecessary side effects of chemoradiotherapy.Objective: Based on clinical risk factors,semantic features of MR Images,preoperative and postoperative radiomics features,to construct a model capable of predicting the efficacy of standard chemoradiotherapy for postoperative residual lesions in patients with glioma,and finally to provide a non-invasive prediction tool for the evaluation of clinical efficacy before treatment.Patients and methods:The clinical and MR Imaging data of 2108 patients with glioma who underwent surgical treatment and obtained histopathological results from five medical institutions from January 2015 to December 2021 were retrospectively collected.According to the inclusion and exclusion criteria,132 patients with residual lesions after glioma surgery were finally included in the study,including 3 institutions as the training set(95 cases)and 2 institutions as the external independent verification set(37 cases).Clinical semantic model,radiomics model and combined prediction model were constructed respectively.Three different segmentation methods were used for radiomics extraction,namely preoperative MR image delineating tumor area+ edema area(ROI-1),preoperative MR image delineating tumor area(ROI-2)and postoperative MR image delineating tumor area(ROI-3).At the same time,the MR radiomics models under each segmentation method were established based on four different machine learning algorithms,including support vector machine(SVM),decision tree(Tree),Ada algorithm(Ada),logistic regression(Log),and five different imaging sequences,including T1 WI,T2WI,T2-Flair plain scan,T1 WI contrast enhancement sequence and combined sequence.In addition,combined prediction models were established based on clinical-semantic features and radiomics features.After the establishment of different models,model calibration,prediction effectiveness comparison and clinical applicability evaluation were carried out.Evaluation indexes included: area under the receiver operating charac-teristic curve(AUC),sensitivity,specificity,accuracy,net reclassification index(NRI)and integrated discrimination improvement index(IDI).Results: The predictive efficacy of the clinical-semantic model was good in the training set and the validation set,and the AUC values were: 0.766[95%CI,0.669-0.864] and 0.650[95%CI,0.453-0.848],the sensitivity were 0.683 and 0.563,the specificity were 0.722 and 0.810,and the accuracy were 0.705 and 0.703,respectively.Under each ROI segmentation,20 radiomics prediction models were established based on different algorithms and different imaging sequences.Comparison between the radiomics models showed that the model based on preoperative T2-Flair sequence had the best performance,and its AUC value and specificity reached above 0.8 in both the training set and the validation set,but its sensitivity was poor.Comparison between the combined models showed that the model based on semantic features and preoperative T2-Flair sequence had the best performance,and the prediction efficiency was better than that of the simple radiomics model.The AUC values in the training and validation set were: 0.866[95%CI,0.790-0.942] and 0.810[95%CI,0.667-0.952],the sensitivity were 0.707 and 0.688,the specificity were 0.870 and0.810,and the accuracy were 0.800 and 0.757,respectively.The combined prediction model based on semantic features and postoperative T1 WI sequence performed the second best,and the AUC values were: 0.812[95%CI,0.726-0.898] and 0.711[95%CI,0.541-0.881],the sensitivity were 0.659 and 0.625,the specificity were 0.778 and0.714,and the accuracy were 0.726 and 0.676,respectively.Conclusion: Combined models based on MR radiomics and semantic features of multi-center patients before and after surgery were able to predict the short-term outcome of standard chemoradiotherapy before treatment.The model is stable and verified by external data,and has the prospect of clinical application,but the prediction performance of the model can be further improved.