Radiomics-based prediction of head and neck cancer recurrence: a multi-centric MRI study

In this study, a classifier for head and neck cancer (HNC) recurrence was developed using MRI radiomics. In total, 51 patients (10 recurrences) from 8 different hospitals were included in this study. For each patient the following MRI-sequences were available: T1-weighted; T1-weighted with fat suppression and contrast enhancement; T2-weighted. A total of 2553 radiomic features were extracted from the images. In order to reduce the dimensionality, features selection based on features, stability and correlation was performed. Synthetic minority oversampling technique (SMOTE) was used to balance the classes. Least absolute shrinkage and selection operator (LASSO) logistic regression was used to perform the classification. A 10-fold cross-validation was performed and an unbiased estimate of the classifier sensitivity, specificity and area under the curve (AUC) were computed. Model sensitivity and specificity were 90% and 68% respectively and the AUC was 0.80. The radiomic classifier could be useful to perform a preliminary stratification, allowing to optimize the follow-up strategy for high risk patients in a more efficient and cost-effective way.