| Objectives:The purpose of our study is to establish machine learning models to predict the progression of patients with nasopharyngeal carcinoma(NPC)who received concurrent chemoradiotherapy,by using machine learning algorithms based on pretreatment magnetic resonance imaging(MRI)radiomics.Therefore,our goals are:(1)To extract radiomic features from pretreatment MRI of the patients and screen out the features related to the progression of patients with NPC;(2)To find out the radiomics features and the clinical characteristics related to the progression free survival(PFS)of patients with NPC;(3)To establish classifiers based on different machine learning algorithms for the prediction of progression within 3 years and evaluate the predictive performance of the classifiers;(4)To establish a nomogram based on the Cox proportional hazards model to predict the PFS of patients and evaluate the predictive performance of the nomogram.Materials and Methods:A retrospective analysis was conducted on first diagnosed stage II-IVA NPC patients treated in The General Hospital of the Western Theater Command from 2016 to 2018.(1)Select patients based on inclusion and exclusion criteria and divide all of the patients into training cohort and test cohort randomly according to the ratio of 7:3;(2)Primary nasopharyngeal gross tumor volume(GTVnx)was manually delineated on T1-weighted imaging(T1WI)and T2-weighted imaging(T2WI)through the open source software Itk-Snap;(3)Feature extraction was achieved using the Pyradiomics package in Python;(4)Three methods were applied to select radiomics features and six machine learning algorithms were applied to establish classifiers;(5)10-fold cross-validation was applied to evaluate the models.Area under the Receiver Operating Characteristic(ROC)curve(AUC),accuracy,precision,recall,and F1-score were used to evaluate the predictive performance of the models;(6)Radiomics features related to PFS were selected by Lasso-cox,and Rad-score(RS)of each patient was calculated;(7)The clinical characteristics related to PFS were selected by Cox proportional hazards model;(8)A nomogram based on Cox proportional hazards model was established to predict the PFS of patients;(9)The nomogram was evaluated by the concordance index and 95% confidence interval.The calibration curves of the predictive performance of 1-year PFS,2-year PFS,3-year PFS were drawn respectively.Results:A total of 176 patients with NPC were included in our study.They were randomly divided into training cohort and test cohort according to the ratio of 7:3,of which 123 were in the training cohort and 53 in the test cohort.In all classifiers for the prediction of progression within 3 years of patients with NPC,the one combined Lasso CV(Least absolute shrinkage and selection operator cross validation)and Adaboost algorithm achieve the best predictive performance.The mean AUC,accuracy,precision,recall,and F1-score are 0.73,0.71,0.79,0.67,0.77 in the training cohort,and 0.70,0.69,0.77,0.65,0.71 in the test cohort,respectively.In addition,two PFS related radiomic features were obtained through Lasso-cox.According to RS,patients were divided into high and low risk groups.In the training cohort,there was a significant difference in PFS between the two groups(P=0.016).Subsequently,the results of the Cox proportional hazards model showed that smoking and RS were independent indicators of PFS in NPC patients.The C-index and 95% confidence interval of the nomogram in the training cohort and the test cohort were 0.679(0.574-0.748),0.626(0.451-0.690),respectively.Conclusion:Predictive classifiers based on MRI radiomics can be used to predict the progression within 3 years of patients with NPC who received concurrent chemoradiotherapy.Although meaningful predictive effect was achieved,more research is needed to further improve the classification performance of machine learning models.In the Cox proportional hazards model associated with PFS of patients,smoking and RS were significantly associated with PFS.The nomogram based on multivariate Cox proportional hazards model is expected to be an effective tool for the prediction of PFS in patients with NPC.Furthermore,the prognosis of patients was affected by various factors.A large amount of data is required for the establishment of the predictive models and accurate stratification analysis is necessary.In addition,there are significant differences in the principles and parameter settings of different machine learning algorithms.The predictive performance of models established by different machine learning algorithms vary greatly.It is very important to choose applicable method.MRI-based radiomics is expected to be a noninvasive and efficient tumor biomarker and a supplementary tool in patients with NPC for the evaluation of prognosis. |
PDF Full Text Request