A Study Of Nasopharyngeal Carcinoma Prognostic Models Based On Pretreatment MR Features And Radiomic Signatures

Purpose: To investigate the feasibility and efficiency of prognostic prediction models for nasopharyngeal carcinoma(NPC)patients without distant metastasis based on pretreatment magnetic resonance imaging(MRI)findings and radiomics signatures.Materials and methods: This was a retrospective study consisting of 529 NPC patients without distant metastasis who were first diagnosed in 2015-2016.Univariate survival analysis was used to identify possible prognostic factors among clinical characteristics and pretreatment MRI findings on primary tumor and regional lymph nodes,with Kaplan-Meier method and Log-rank test(p<0.05).Multivariate survival analysis were performed under the Cox proportional hazards models and nomogram to decide independent prognostic factors and Hazard Ratios(HR).The predictive accuracy of the nomogram models were compared with the models combining clinical characteristics with TNM classification by C-index and calibration curves.A total of 2818 radiomic signatures were extracted from T2 WI and CET1 WI sequences for each case among 203 patients by Radcloud platform 3.0.Radiomics signatures were obtained with the least absolute shrinkage and selection operator method(LASSO),followed by 4 kinds of machine learning algorithms(KNN/SVM/XGBoost/LR)to design a distant-metastasis prediction model(52:100),a local-recurrence prediction model(44:100),and a diseaseprogression prediction model(103:100)for NPC.The feasibility and efficiency of these models were then explored with receiver operating characteristic curve values(AUC),sensitivity and specificity.Results: In univariate survival analysis,age,distant metastasis were significantly correlated with overall survival(OS);pretreatment plasma lymphocyte percentage was significantly correlated with progression-free survival(PFS).Many MRI features of primary tumor extension and regional lymph nodes were significantly correlated with OS,PFS,local recurrence-free survival(LRFS)and distant metastasis-free survival(DMFS).In multivariate analysis,age>47(HR=3.203),musculus pterygoideus lateralis(MPL)invasion(HR=2.575),distant metastasis(HR=5.769)were independent negative factors for OS;levator veli palatin(LVP)or tensor veli palatin(TVP)invasion(HR=1.978),diameter of retropharyngeal lymph node>1.5cm(HR=1.532),cervical level3/4 lymph node metastasis(HR=2.068)were independent negative predictors for PFS.When combining clinical information with MR features,the nomograms performed better than those combining with TNM stage according to increased C-index.Among all the radiomics-based models,optimal machine-learning classifiers and features were identified to design 3 prediction models.For distantmetastasis prediction model,18 radiomic signatures accompany with age,CSA/PVM invasion,retropharyngeal lymph node metastasis were chosen under SVM,(AUC=0.745,95%CI:0.57-0.92).19 radiomic signatures were selected to construct the local-recurrence prediction model under KNN,with(AUC=0.878,95%CI:0.74-1.00).For disease-progression prediction model,22 radiomic signatures along with LVP/TVP invasion,diameter of retropharyngeal lymph node,cervical level3/4 lymph node metastasis were gathered to establish under LR,(AUC=0.829,95%CI:0.72-0.94).Conclusions: Compared with TNM stage,pretreatment MRI features of primary tumor extension and regional lymph nodes involvement can provide more information by nomograms,thus improving the performance of prognostic models for NPC patients.In this study,MR based-radiomic models can predict NPC prognosis in a good way,especially along with clinical characteristics and morphological MR features.But more investigations with larger samples still need to be taken for further study.