Figure 3.

Statistical workflow. Survival algorithms were trained and evaluated according to concordance index using 100-times repeated 5-fold cross-validation and based on three types of inputs that were pre-filtered using univariable Cox regressions: (i) clinicopathological features, (ii) all radiomics-based and clinicopathological features, and (iii) uncorrelated parsimonious radiomics-based and clinicopathological features. Arrows indicate that in each fold of the 5-fold cross-validation, the models were trained on the 4 gray blocks and then applied and evaluated on the remaining light red block Abbreviations: c.p., clincopathological; LASSO, least absolute shrinkage and selection operator.