Shape-based and quantitative MRI radiomic features in the assessment of non-metastatic nasopharyngeal carcinoma

Abstract

Purpose: To investigate the association between MRI radiomic features and progression-free survival (PFS) of patients with nasopharyngeal carcinoma (NPC) after treatment. Methods and Materials: A cohort of 100 patients with histologically confirmed NPC that underwent standard treatment was retrospectively obtained. A total of 102 radiomic features were extracted from contrast-enhanced T1-weighted (CE-T1W) scans acquired at diagnosis. Univariate and multivariate analysis was conducted to assess the association between radiomic features and PFS. Mann-Whitney U test and log-rank test of Kaplan-Meier estimate was conducted for univariate analysis. Minimum redundancy and maximum relevance feature selection was used to select the most optimum subset of features for multivariate analysis. Logistic regression and support vector machine were trained with the selected features and validated using 5x8 repeated k-fold cross validation. The performance of the models was assessed by receiver operator characteristic analysis. Results: Both univariate and multivariate analysis has identified the shape feature tumour sphericity to be significantly associated with 3-year PFS (Mann-Whitney U test: p < 0.0001; log rank: p < 0.0001; logistic regression: AUC=0.765; SVM: AUC=0.799). No significant separation in PFS was seen with clinical staging. No increase in performance was found with multivariate models. Conclusion: There is a significant association between MRI radiomic features and PFS of NPC patients after standard treatment. Tumour sphericity has been identified to be most correlated to 3-year PFS. Radiomic models could potentially offer a non-invasive method of predicting treatment outcomes.