Title: Discriminating between p16-negative oropharyngeal and non-oropharyngeal origins by their metastatic lymph nodes using machine learning approach based on MRI radiomics

Objectives: To examine the capability of MRI radiomic analysis to differentiate oropharyngeal and non-oropharyngeal origins of cervical metastatic lymph nodes (LNs) in p16-negative patients following machine-learning approaches. Methods: : A total of 150 patients (training set: 90, validation set: 60) diagnosed with p16-negative head and neck squamous cell carcinoma with cervical metastatic LNs were retrospectively included. Radiomic features were extracted from pretreatment diagnostic T2-weighted fat-suppressed and contrast-enhanced T1-weighted images of the cervical metastatic LNs. The Wilcoxon-Mann-Whitney test (with p 0.48) were used to identify the discriminant features between the oropharyngeal and non-oropharyngeal groups. The discriminant features were used to train and test four different machine-learning classifiers. The discriminative performances of classifiers were evaluated by AUC, sensitivity, specificity, PPV, and NPV and compared. Results: : Five radiomic features were significantly different (P 0.48) between the p16-negative oropharyngeal and non-oropharyngeal groups. Among the four classifiers trained and applied for discrimination between oropharyngeal and non-oropharyngeal origins, the logistic regression and neural network both showed the highest prediction power, with AUC of 0.87 and 0.74 in the training and validation cohorts, respectively. The accuracy, sensitivity, specificity, PPV, and NPV of the two classifiers in the validation cohort were 0.72, 0.82, 0.66, 0.58, and 0.86, respectively. Conclusion: Radiomics and machine-learning diagnostic models with MRI data are potentially useful supportive tools to discriminate between oropharyngeal and non-oropharyngeal origins in p16-negative patients with cervical metastatic LNs.