Establishing An Early Predictive Model Of Nasopharyngeal Carcinoma Based On MRI Radiomics

Objective: To explore the application value of radiomic features of pre-treatment MRI in predicting the response of nasopharyngeal carcinoma to induction chemotherapy and the residual tumor risk of primary nasopharynx after radiotherapy and chemotherapy.Methods: This study retrospectively analyzes 191 newly and pathologically diagnosed locally advanced nasopharyngeal carcinoma patients from December 2015 to December 2018 in tumor hospital of human province who underwent induction chemotherapy(IC)and concurrent chemoradiotherapy(CCRT).Each patient has MRI images before treatment within 2 weeks、after 2 cycles of induction chemotherapy and after 3 months of chemoradiotherapy and the images include T1,T2,and T1 C sequences.According to the RECIST 1.1 solid tumor efficacy evaluation criteria,chemotherapy responses were evaluated,and 191 patients are divided into response groups(CR and PR)and non-response groups(SD and PD).191 patients were divided into residual groups and non-residual groups according to MRI images after treatment by an imaging doctor and a radiation therapist.The nasopharyngeal mass in the MRI image is delineated on the MIM planning system,which is the region of interest(ROI).From the pre-treated MRI images TI,T2,and T1 C three sequences,IBEX software was used to extract six features: shapes,gray histogram,Intensity direct,Gray-Level Co-occurrence Matrix(GLCM),gray level run length matrix(GLRLM),and neighborhood intensity different matrix(NIDM).After screening the radiomic features related to chemotherapy response and tumor residuals by chi-square test,Mann-Whitney U test and Lasso logistic regression analysis,the prediction model of chemotherapy response and primary tumor residual of nasopharynx is established using SVM classifier and five-fold cross validation method.Finally,evaluate the model’s predictive ability according to AUC,accuracy,specificity and sensitivity.Results: Among the 191 patients enrolled,135 are in the chemotherapy response group and 56 are in the non-response group;58 are in the imaging residual group and 133 are in the non-residual group.IBEX extracts 1582 initial radiomic features from TI,T2,and T1 C,respectively.Eight radiomic features closely related to tumor chemotherapy response and nine radiomic features closely related to tumor residual events were selected by chi-square test,Mann-Whitney U test and Lasso logistic regression analysis,and no clinical parameters are selected in both groups.The chemotherapy response and tumor residual prediction models are established by SVM classifier and 5-fold cross-validation.The average AUC,accuracy,specificity and sensitivity of the chemotherapy response prediction model in the test set are 0.79,72%,77% and 67%;the average AUC,accuracy,specificity and sensitivity of the tumor residual prediction model in the test set are 0.73,67%,54%,77%.Conclusion: The radiomic features of the MRI image before treatment have better risk stratification ability for locally advanced nasopharyngeal carcinoma,and provide a new supplement for individualized precise treatment.