Multi-modality Image-based Radiomics Analysis Applying To EBV DNA Statue And Level Of Nasopharyngeal Carcinoma

Objective:This study aims at employing multi-modality image ebased radiomics analysis within the gross tumor volume among NPC patients for the selection of candidate features,which can be correlated to EBV DNA statue and high risk level.Methods and materials:Data for NPC patients(n=230)dispositioned to definitive radiation therapy at Radiation Physics Technology Center,West China Hospital of Sichuan University between 2017/5-2018/1,were scanned with CT and MR.Contrast-enhanced CT from the same CT and Contrast-enhanced T1WI and T2WI from the same MR contours of the gross primary tumor were extracted in DICOM-RT format for patients with known EBV DNA status and Level.Plasma EBV DNA Level by PCR set as a cutoff with 100 copies for the test(less than 100 copies denotes a negative test and vice versa),and cutoffs with 4000 copies and 1 OOOcopies for the positive test(more than 4000copies denotes a high risk test,less than 1000 denotes a low risk test and the other denotes a middle risk test).A total of 1026 radiomics features was extracted from gross primary tumor by Pyradiomics.(1)DNA statue prediction:Homogeneous testing between training and validation cohorts was done by Wilcoxon-rank-sum test with a cutoff p-value<0.05.Kolmogorov-Smirnov(KS)test was done between EBV+ and EBV-subjects in the training cohort with a cutoff p-value<0.05.A biserial Spearsman correlation between EBV status and each radiomic feature was calculated with a cutoff of biserial absolute correlation>0.3.Only the top 20 features with highest marginal AUC were kept by ranking remaining features by their marginal Area Under Curve(AUC).The top 20 feature highest marginal AUC of all modality image was set as the multi-modality group.Multilayer perceptron was applied as the classifier.10 random 6-folder cross-validation and 50 times retest with random split were introduced with AUC,Confusion Matrix and Classification report.Also,the Support Vector Machine(SVM),Decision Tree(DT)and Logistic Regression(LR)were introduced as comparison.Meantime,SMOTE was used to balance the positive group and the negative group.(2)The features associated with high risk biomarker:The Kruskal-Wallis rank-sum test with a cutoff p-value<0.05 and biserial Spearsman correlation between three groups with a cutoff biserial absolute correlation>0.3 was done.We add features one by one into a Logistic Regression model according to their ranks of biserial absolute correlation from high to low,until the model AUC stopped increasing.False discovery rate(FDR)was evaluated using permutation test.Results:(1)The AUC of four Multilayer perceptron classifiers based on four modalities was 0.810(CT),0.791(T1WI),0.821(T2WI)and 0.832(multi-modality)by 10 random 6-folder cross-validation.The precision of four classifiers was 0.816(CT),0.795(T1WI),0.826(T2WI)and 0.849(multi-modality)by 50 times retest with random split.The Multilayer perceptron classifier based on multi-modality image performed well.In the comparison test,the AUC of four machine learn models based on multi-modality image 0.827(SVM),0.816(DT),0.829(LR)and 0.832(MLP).By using SMOTE to balance the positive group and the negative group,the AUC of four Multilayer perceptron classifiers based on four modalities was 0.903(CT),0.893(T1WI),0.917(T2WI)and 0.929(multi-modality)by 10 random 6-folder cross-validation.The precision of four classifiers was 0.878(CT),0.868(T1WI),0.880(T2WI)and 0.915(multi-modality)by 50 times retest with random split.The sensibility and specificity was 0.916(0.800),0.884(0.757),0.915(0.915)and 0.946(0.868).(2)In three modality group,the Major Axis feature of shape class can differentiate well the three risk group,the high risk group had a higher value.Some features associated with ellipsoid and heterogeneity of ROI can be the biomarkers associated with high risk.Conclusions:Our study show that the radiomics analysis can discriminate between EBV DNA positive and negative tumors.The multilayer perceptron classifiers based on multi-modality performed well.And radiomics features,specifically shape and heterogeneity features,can discriminate the high risk group from the middle one and the low one.