Predictive Value Of Multi-parametric MRI Radiomics Model For Local Recurrence Of Advanced Nasopharyngeal Carcinoma

Objective Based on MRI images of multiple disease progression points of advanced nasopharyngeal carcinoma,an appropriate machine learning model was established to explore the predictive value of multi-parametric MRI radiomics model for local recurrence of advanced nasopharyngeal carcinoma.Methods and materials It's a retrospective study to analysis 86 cases of advanced nasopharyngeal carcinoma with clinical data and imaging data.They were confirmed at pathology from July 2011 to December 2016 in Fujian Provincial Cancer Hospital.According to the follow-up data,35 patients of nasopharyngeal carcinoma recurrence confirmed by pathology were included in recurrent group.The rest of 51 patients without recurrent evidence at pathology and imaging data were included in non-recurrent group.All enrolled subjects underwent three routine MRI scans at Philips 3.0 T magnetic resonance.The scan time points were initial diagnosis,at the end of radiotherapy,and 3-6 months after radiotherapy respectively.Axial T2WI-STIR and T1 WI enhanced images in three scans were acquired and pre-processed.A resident delineated the region of interest(ROI)along the edge of lesions manually,then a three-dimensional volume image of the tumor was obtained.Spearman correlation analysis and clustering were used to extract and reduce the feature parameters of the image data.The feature parameters with significant correlation were filtered and kept,and their statistical results were obtained.The predictive model is constructed by classifier learning with high accuracy.Additionally,the machine learning classifiers and model were trained and verified by the 5-fold cross-validation method.The confusion matrix and the area under the receiver operating characteristic curve(AUC)were used to evaluate performance of the model.Results From all axial T2WI-STIR and T1 WI enhanced images,relative feature parameters between two groups were extracted,including the First Order Statistics,Shape-based,Gray Level Co-occurence Matrix,Gray Level Run Length Matrix,Gray Level Dependence Matrix etc.Then 13 feature parameters with significant correlation were filtered and kept.Using the 5-fold cross-validation method to train and verify 23 machine learning classifier,it is found that MLPClassifier show more accurate than other machine learning classifiers,with an AUC of more than 0.8.The artificial neural network model was constructed through MLPClassifier learning.The sensitivity of the model was 80.0%,the specificity was 88.0%,and mean AUC was 0.91.Among the 13 feature parameters,69%(9/13)was extracted at the end of radiotherapy,and T1 WI enhanced feature parameters accounted for 77.8%(7/9).In addition,the T1 WI enhanced image at the end of radiotherapy,the ShapeBased feature parameter Elongation show the best predictive performance,with a sensitivity of 80.0%,a specificity of 64.7%,AUC of 0.759,and 95% confidence interval of 0.655~0.845.Conclusion For the prediction of local recurrence of advanced nasopharyngeal carcinoma,the MLPClassifier is more accurate than other machine learning classifiers.The predictive performance of artificial neural network model is better,and T1 WI enhanced sequence combined with T2WI-STIR sequence may improve the performance of the model.Therefore,the multi-parametric MRI radiomics model shows better value in predicting local recurrence of advanced nasopharyngeal carcinoma.